Battle-anime power transformation · Humaneness 1.5 + real-speech rank 4

Two frozen models. Exact instructions, matched seeds and complete take distributions. Find a promising recipe quickly, then inspect its limitations and listen.

Model, weights and inference code Pinned revision

← Model overview · Complete scenario take index

Listen to matched examples

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to creative-event-contrast-three at CFG 3; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5281.5670.898
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.6490.583
Elation
emo_Elation
20.00%high0.4520.7181.0540.680
Arousal
vn_AROU_reg
15.00%high2.4371.1274.2590.941
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.5550.978
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Re-scored, not regenerated: the original raw EARS predictions were reused on CPU. Old fitness, audio, prompts and scientific gates are retained. Old held-out sets tested only the previously selected wording; a newly re-ranked search winner does not inherit that validation. No universal best weighting or human-preference optimum is claimed.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base  [historical pilot only]

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Start here: best observed recipe, not a universal optimum

creative-event-contrast-three · CFG 3

The current starting point uses EARS Steering v2 across the available canonical search conditions, with unchanged qualification gates taking priority. Historical candidates are used only where no canonical search exists. Held-out results below refer to the earlier selected wording and are not automatically inherited by a newly re-ranked winner.

No recipe satisfies every realization/quality gate. The highlighted recipe is the best diagnostic candidate, not a reliable recommendation.

Burst caution: family detection anywhere, nominal placement and natural blending are different questions. The pilot cannot establish accurate word-level placement. The follow-up requires a60%family-hit floor in addition to speech/identity QA; a quality-only clip with no detected event does not qualify.

How the strategies compare

Click any column header to sort. Do not compare scalar fitness across different missions or between legacy and event-aware scoring contracts.

Transfer to unseen sentences — earlier selection, now re-scored

The previously selected wording was repeated on five new lexically neutral bilingual scripts, at selected guidance and guidance 1. These outputs are now also re-scored with Steering v2. They do not validate other newly favored wording and are not a human listening validation.

heldout-selected20.4886.2854.2040.73%0.00%0.009not qualified
heldout-g11.00.4266.3054.4130.40%0.00%0.002not qualified
creative-cfg-five-stresscanonical-v1-explorationconcise trained baseline56.1053.7420.88%n/a0.7570.001diagnostic
creative-cfg-four-stresscanonical-v1-explorationconcise trained baseline46.1343.8100.49%n/a0.7800.001diagnostic
creative-cfg-threecanonical-v1-explorationconcise trained baseline36.2743.9070.36%n/a0.7610.000diagnostic
creative-cfg-twocanonical-v1-explorationconcise trained baseline26.2864.0210.27%n/a0.7010.001diagnostic
creative-contrast-and-claritycanonical-v1-explorationconcise trained baseline1.56.3704.1260.24%n/a0.6160.000diagnostic
creative-event-contrast-threecanonical-v1-explorationconcise trained baseline36.0913.8390.63%n/a0.7900.001diagnostic
creative-event-contrast-twocanonical-v1-explorationconcise trained baseline26.2703.9640.30%n/a0.7500.000diagnostic
creative-global-onlycanonical-v1-explorationGENERAL-only delivery cues26.3654.0771.00%n/a0.6510.000diagnostic
creative-inline-onlycanonical-v1-explorationinline-only delivery cues26.3294.0380.43%n/a0.6880.000diagnostic
creative-literal-interjection-proxycanonical-v1-explorationconcise trained baseline26.2884.0210.83%n/a0.8080.056diagnostic
creative-long-event-stresscanonical-v1-explorationconcise trained baseline26.3154.0880.52%n/a0.6180.007diagnostic
creative-medium-eventcanonical-v1-explorationconcise trained baseline26.3043.9990.27%n/a0.7210.002diagnostic
creative-muscular-releasecanonical-v1-explorationconcise trained baseline2.56.2813.9350.33%n/a0.6680.000diagnostic
creative-native-baselinecanonical-v1-explorationconcise trained baseline16.3714.1240.52%n/a0.5880.000diagnostic
creative-private-reflexcanonical-v1-explorationconcise trained baseline1.56.3554.1260.37%n/a0.6040.000diagnostic
creative-reaction-firstcanonical-v1-explorationconcise trained baseline26.2884.0111.22%n/a0.7520.004diagnostic
creative-restrained-controlcanonical-v1-explorationconcise trained baseline16.3754.1290.21%n/a0.5960.000diagnostic
creative-short-training-rangecanonical-v1-explorationconcise trained baseline26.2893.8900.30%n/a0.7510.000diagnostic
repair-canonical_bothcanonical-v1-explorationinline-only delivery cues2.56.2883.9460.43%n/a0.7220.001diagnostic
repair-duration_onlycanonical-v1-explorationinline-only delivery cues2.56.2883.9620.43%n/a0.6960.002diagnostic
repair-historical_controlcanonical-v1-explorationinline-only delivery cues2.56.3214.0201.88%n/a0.6790.002diagnostic
repair-labels_onlycanonical-v1-explorationinline-only delivery cues2.56.3214.0722.68%n/a0.6610.000diagnostic
heldout-g1canonical-v1-validationconcise trained baseline1.06.3054.4130.40%n/a0.4260.002diagnostic
heldout-selectedcanonical-v1-validationconcise trained baseline26.2854.2040.73%n/a0.4880.009diagnostic
g0-p01full-v2-gen0concise trained baseline16.3904.1580.27%n/a0.5790.000diagnostic
g0-p02full-v2-gen0audible mechanics1.256.3514.1040.45%n/a0.6190.000diagnostic
g0-p03full-v2-gen0situation and intention1.56.3454.0760.73%n/a0.6410.000diagnostic
g0-p04full-v2-gen0organic conversation26.3624.0660.12%n/a0.6740.000diagnostic
g0-p05full-v2-gen0attack, sustain and release2.56.2873.9260.18%n/a0.7280.000diagnostic
g0-p06full-v2-gen0containment contrast36.2873.9630.40%n/a0.7360.000diagnostic
g0-p07full-v2-gen0reacting to an unheard partner1.56.3624.1070.30%n/a0.6610.000diagnostic
g0-p08full-v2-gen0GENERAL-only delivery cues26.3524.0470.30%n/a0.6700.000diagnostic
g0-p09full-v2-gen0inline-only delivery cues2.56.2753.9610.27%n/a0.7050.004diagnostic
g0-p01pilot-v1concise trained baseline1.06.3704.2202.65%n/a0.5340.216pass
g0-p02pilot-v1audible mechanics1.256.3804.2132.27%n/a0.5530.219pass
g0-p03pilot-v1situation and intention1.56.3414.1751.53%n/a0.5520.218pass
g0-p04pilot-v1organic conversation2.06.3064.0573.41%n/a0.6510.216pass
g0-p05pilot-v1attack, sustain and release2.56.3373.9902.45%n/a0.6750.219pass
g0-p06pilot-v1containment contrast3.06.3183.9981.94%n/a0.6810.221pass
g0-p07pilot-v1reacting to an unheard partner1.56.3744.1583.40%n/a0.5910.214pass
g0-p08pilot-v1GENERAL-only delivery cues2.06.3314.0702.90%n/a0.6320.225pass
g0-p09pilot-v1inline-only delivery cues2.56.3214.0201.88%n/a0.6790.227pass
g1-p01pilot-v1inline-only delivery cues26.3044.2044.13%n/a0.4790.211pass
g1-p02pilot-v1inline-only delivery cues2.56.3274.1531.80%n/a0.5590.225pass
g1-p03pilot-v1inline-only delivery cues1.56.3884.1971.12%n/a0.5620.224pass
g1-p04pilot-v1GENERAL-only delivery cues1.56.3314.3453.04%n/a0.4560.220pass
g1-p05pilot-v1GENERAL-only delivery cues2.56.3364.2012.70%n/a0.5990.218pass
g1-p06pilot-v1GENERAL-only delivery cues3.06.3414.1741.70%n/a0.5440.221pass
g1-p07pilot-v1containment contrast2.06.2744.2153.82%n/a0.4890.212pass
g1-p08pilot-v1containment contrast2.56.1813.9643.36%n/a0.5590.210pass
g1-p09pilot-v1containment contrast1.56.2734.3101.99%n/a0.4990.216pass

Listen to complete take distributions

Baseline, leading recipes and runners-up retain every language/reference/seed combination. These are not hand-picked best takes. Players use individual MP3 URLs, no base64 downloads, and preload nothing. Initial view is paginated for fast loading.

English fixed reference ·12seconds

A clean fixed codec-reconstructed reference, a different utterance from every search scene. The same recording is used by both models.

German fixed reference ·12seconds

A clean fixed codec-reconstructed reference, a different utterance from every search scene. The same recording is used by both models.

Recipe details and exact prompt contracts

creative-literal-interjection-proxy · CFG 2

canonical-v1-explorationCANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to creative-literal-interjection-proxy at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5281.5090.891
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.8750.687
Elation
emo_Elation
20.00%high0.4520.7181.2330.736
Arousal
vn_AROU_reg
15.00%high2.4371.1273.7480.862
Vocal tension
vn_TENS_reg
10.00%high1.6590.8322.9880.923
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.

0.8086.2884.0210.83%0.056100.00%0.017

All expected burst families anywhere: 3.00%. All nominal event-window matches: 3.00%. Detector observations, not verified true realization rates.

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Use a raw open vowel before recovering the intended words.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Use a raw open vowel before recovering the intended words.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Use a raw open vowel before recovering the intended words.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Use a raw open vowel before recovering the intended words.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Use a raw open vowel before recovering the intended words.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Use a raw open vowel before recovering the intended words.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Use a raw open vowel before recovering the intended words.) [4.2 seconds duration] Ah! Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Use a raw open vowel before recovering the intended words.) [3.8 seconds duration] Ah! Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Ah! Every part of me is shaking... this power is finally mine. Ah! Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] Ah! Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Ah! Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.8126.3134.0611.25%0.624
none500.8046.2633.9800.41%n/a

creative-event-contrast-three · CFG 3

canonical-v1-explorationCANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to creative-event-contrast-three at CFG 3; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5281.5670.898
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.6490.583
Elation
emo_Elation
20.00%high0.4520.7181.0540.680
Arousal
vn_AROU_reg
15.00%high2.4371.1274.2590.941
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.5550.978
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.

0.7906.0913.8390.63%0.001100.00%0.008

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. A separate raw vocal reaction interrupts the sentence rhythm.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. A separate raw vocal reaction interrupts the sentence rhythm.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. A separate raw vocal reaction interrupts the sentence rhythm.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. A separate raw vocal reaction interrupts the sentence rhythm.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. A separate raw vocal reaction interrupts the sentence rhythm.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. A separate raw vocal reaction interrupts the sentence rhythm.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. A separate raw vocal reaction interrupts the sentence rhythm.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. A separate raw vocal reaction interrupts the sentence rhythm.) [3.8 seconds duration] Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
[0.48 seconds pause]
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
[0.80 seconds pause]
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.7876.1173.8240.79%0.615
none500.7946.0653.8540.48%n/a

creative-cfg-four-stress · CFG 4

canonical-v1-explorationCANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to creative-cfg-four-stress at CFG 4; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5281.6680.898
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.5440.548
Elation
emo_Elation
20.00%high0.4520.7180.9700.658
Arousal
vn_AROU_reg
15.00%high2.4371.1274.3840.952
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.8130.989
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.

0.7806.1343.8100.49%0.001100.00%0.006

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Make the reaction physically immediate and unmistakable.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Make the reaction physically immediate and unmistakable.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Make the reaction physically immediate and unmistakable.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Make the reaction physically immediate and unmistakable.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Make the reaction physically immediate and unmistakable.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Make the reaction physically immediate and unmistakable.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Make the reaction physically immediate and unmistakable.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Make the reaction physically immediate and unmistakable.) [3.8 seconds duration] Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.7956.1553.8250.12%0.612
none500.7656.1123.7950.85%n/a

creative-cfg-three · CFG 3

canonical-v1-explorationCANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to creative-cfg-three at CFG 3; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5281.4070.858
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.5920.570
Elation
emo_Elation
20.00%high0.4520.7180.9010.647
Arousal
vn_AROU_reg
15.00%high2.4371.1273.9950.900
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.4620.969
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.

0.7616.2743.9070.36%0.000100.00%0.039

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the reaction escape before you can control it. Do not merely describe a reaction.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Let the reaction escape before you can control it. Do not merely describe a reaction.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Let the reaction escape before you can control it. Do not merely describe a reaction.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the reaction escape before you can control it. Do not merely describe a reaction.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the reaction escape before you can control it. Do not merely describe a reaction.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Let the reaction escape before you can control it. Do not merely describe a reaction.) [3.8 seconds duration] I spent my whole life waiting for this impossible moment.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Let the reaction escape before you can control it. Do not merely describe a reaction.) [4.2 seconds duration] It is tearing through me... and I am not afraid anymore.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the reaction escape before you can control it. Do not merely describe a reaction.) [4.6 seconds duration] This is what freedom feels like when nothing can hold you down.

Literal words in the separate text field:

I spent my whole life waiting for this impossible moment. It is tearing through me... and I am not afraid anymore. This is what freedom feels like when nothing can hold you down.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I spent my whole life waiting for this impossible moment.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] It is tearing through me... and I am not afraid anymore.
(scream, 0.80 seconds)
(Natural clear speech.) [4.6 seconds duration] This is what freedom feels like when nothing can hold you down.
Reference versus no-reference breakdown
reference500.7726.2843.9020.18%0.645
none500.7506.2653.9130.55%n/a

creative-cfg-five-stress · CFG 5

canonical-v1-explorationCANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to creative-cfg-five-stress at CFG 5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5281.7410.908
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.4550.494
Elation
emo_Elation
20.00%high0.4520.7180.7750.594
Arousal
vn_AROU_reg
15.00%high2.4371.1274.4240.956
Vocal tension
vn_TENS_reg
10.00%high1.6590.8324.0100.993
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.

0.7576.1053.7420.88%0.001100.00%-0.033

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. The involuntary sound breaks through the spoken performance. Preserve all words.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. The involuntary sound breaks through the spoken performance. Preserve all words.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. The involuntary sound breaks through the spoken performance. Preserve all words.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. The involuntary sound breaks through the spoken performance. Preserve all words.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. The involuntary sound breaks through the spoken performance. Preserve all words.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. The involuntary sound breaks through the spoken performance. Preserve all words.) [4.2 seconds duration] For one more second I thought this strength would break me.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. The involuntary sound breaks through the spoken performance. Preserve all words.) [3.5 seconds duration] Then something changed... and every doubt turned into fire.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. The involuntary sound breaks through the spoken performance. Preserve all words.) [4.6 seconds duration] I am still here, and I am more than I ever imagined.

Literal words in the separate text field:

For one more second I thought this strength would break me. Then something changed... and every doubt turned into fire. I am still here, and I am more than I ever imagined.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [4.2 seconds duration] For one more second I thought this strength would break me.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [3.5 seconds duration] Then something changed... and every doubt turned into fire.
(scream, 0.80 seconds)
(Natural clear speech.) [4.6 seconds duration] I am still here, and I am more than I ever imagined.
Reference versus no-reference breakdown
reference500.7676.1243.7590.54%0.574
none500.7486.0863.7241.22%n/a

creative-reaction-first · CFG 2

canonical-v1-explorationCANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to creative-reaction-first at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5281.2910.849
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.6010.559
Elation
emo_Elation
20.00%high0.4520.7180.9540.650
Arousal
vn_AROU_reg
15.00%high2.4371.1273.8570.879
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.2640.954
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.

0.7526.2884.0111.22%0.004100.00%0.018

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. React before the first words, as if the physical event has already happened.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. React before the first words, as if the physical event has already happened.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. React before the first words, as if the physical event has already happened.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. React before the first words, as if the physical event has already happened.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. React before the first words, as if the physical event has already happened.
SCRIPT:
(exhausted groan, 0.48 seconds)
(scream, 0.80 seconds)
(A fictional adult fighter under enormous strain, trying to contain rising energy. React before the first words, as if the physical event has already happened.) [3.8 seconds duration] I spent my whole life waiting for this impossible moment.
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. React before the first words, as if the physical event has already happened.) [4.2 seconds duration] It is tearing through me... and I am not afraid anymore.
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. React before the first words, as if the physical event has already happened.) [4.6 seconds duration] This is what freedom feels like when nothing can hold you down.

Literal words in the separate text field:

I spent my whole life waiting for this impossible moment. It is tearing through me... and I am not afraid anymore. This is what freedom feels like when nothing can hold you down.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(exhausted groan, 0.48 seconds)
(scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] I spent my whole life waiting for this impossible moment.
(Natural clear speech.) [4.2 seconds duration] It is tearing through me... and I am not afraid anymore.
(Natural clear speech.) [4.6 seconds duration] This is what freedom feels like when nothing can hold you down.
Reference versus no-reference breakdown
reference500.7706.2963.9951.22%0.625
none500.7346.2804.0261.22%n/a

creative-short-training-range · CFG 2

canonical-v1-explorationCANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to creative-short-training-range at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5281.4520.894
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.5720.543
Elation
emo_Elation
20.00%high0.4520.7180.8800.626
Arousal
vn_AROU_reg
15.00%high2.4371.1273.7120.858
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.0320.932
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.

0.7516.2893.8900.30%0.000100.00%0.025

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Use a brief, crisp involuntary reaction, not an extended dramatic sound.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Use a brief, crisp involuntary reaction, not an extended dramatic sound.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Use a brief, crisp involuntary reaction, not an extended dramatic sound.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Use a brief, crisp involuntary reaction, not an extended dramatic sound.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Use a brief, crisp involuntary reaction, not an extended dramatic sound.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Use a brief, crisp involuntary reaction, not an extended dramatic sound.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.30 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Use a brief, crisp involuntary reaction, not an extended dramatic sound.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.30 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Use a brief, crisp involuntary reaction, not an extended dramatic sound.) [3.8 seconds duration] Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.30 seconds)
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.30 seconds)
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.7586.3333.9150.36%0.632
none500.7446.2453.8640.24%n/a

creative-event-contrast-two · CFG 2

canonical-v1-explorationCANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to creative-event-contrast-two at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5281.1990.835
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.6610.590
Elation
emo_Elation
20.00%high0.4520.7180.9410.636
Arousal
vn_AROU_reg
15.00%high2.4371.1273.7700.870
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.1200.941
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.

0.7506.2703.9640.30%0.000100.00%0.048

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. The nonverbal event must actually be heard, then return to the words.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. The nonverbal event must actually be heard, then return to the words.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. The nonverbal event must actually be heard, then return to the words.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. The nonverbal event must actually be heard, then return to the words.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. The nonverbal event must actually be heard, then return to the words.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. The nonverbal event must actually be heard, then return to the words.) [4.2 seconds duration] For one more second I thought this strength would break me.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. The nonverbal event must actually be heard, then return to the words.) [3.5 seconds duration] Then something changed... and every doubt turned into fire.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. The nonverbal event must actually be heard, then return to the words.) [4.6 seconds duration] I am still here, and I am more than I ever imagined.

Literal words in the separate text field:

For one more second I thought this strength would break me. Then something changed... and every doubt turned into fire. I am still here, and I am more than I ever imagined.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [4.2 seconds duration] For one more second I thought this strength would break me.
[0.48 seconds pause]
(Natural clear speech.) [3.5 seconds duration] Then something changed... and every doubt turned into fire.
[0.80 seconds pause]
(Natural clear speech.) [4.6 seconds duration] I am still here, and I am more than I ever imagined.
Reference versus no-reference breakdown
reference500.7686.2883.9500.30%0.654
none500.7326.2523.9780.31%n/a

g0-p06 · CFG 3

full-v2-gen0CANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g0-p06 at CFG 3; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5281.1840.823
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.6030.554
Elation
emo_Elation
20.00%high0.4520.7180.8190.608
Arousal
vn_AROU_reg
15.00%high2.4371.1273.9030.886
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.4170.967
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Full-coverage steering-v2 evolution; transfer, fixed-wording CFG and sealed splits remain distinct.

0.7366.2873.9630.40%0.000100.00%0.004

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: containment contrast.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Begin with the feeling held back, then let it leak through and become unmistakable.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Begin with the feeling held back, then let it leak through and become unmistakable.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Begin with the feeling held back, then let it leak through and become unmistakable.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Begin with the feeling held back, then let it leak through and become unmistakable.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Begin with the feeling held back, then let it leak through and become unmistakable.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Begin with the feeling held back, then let it leak through and become unmistakable.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Begin with the feeling held back, then let it leak through and become unmistakable.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Begin with the feeling held back, then let it leak through and become unmistakable.) [3.8 seconds duration] Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.7336.2883.9470.37%0.635
none500.7406.2853.9780.43%n/a

g0-p05 · CFG 2.5

full-v2-gen0CANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g0-p05 at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5281.2050.838
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.5290.528
Elation
emo_Elation
20.00%high0.4520.7180.7340.587
Arousal
vn_AROU_reg
15.00%high2.4371.1273.8260.877
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.3110.957
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Full-coverage steering-v2 evolution; transfer, fixed-wording CFG and sealed splits remain distinct.

0.7286.2873.9260.18%0.000100.00%-0.006

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: attack, sustain and release.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Build an audible attack, sustain the intended quality, then release it without losing the final words.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Build an audible attack, sustain the intended quality, then release it without losing the final words.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Build an audible attack, sustain the intended quality, then release it without losing the final words.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Build an audible attack, sustain the intended quality, then release it without losing the final words.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Build an audible attack, sustain the intended quality, then release it without losing the final words.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Build an audible attack, sustain the intended quality, then release it without losing the final words.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Build an audible attack, sustain the intended quality, then release it without losing the final words.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Build an audible attack, sustain the intended quality, then release it without losing the final words.) [3.8 seconds duration] Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.7186.3223.9380.18%0.625
none500.7386.2513.9140.19%n/a

repair-canonical_both · CFG 2.5

canonical-v1-explorationCANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to repair-canonical_both at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5281.0120.797
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.5950.558
Elation
emo_Elation
20.00%high0.4520.7180.8470.615
Arousal
vn_AROU_reg
15.00%high2.4371.1273.6620.843
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.1260.935
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.

0.7226.2883.9460.43%0.001100.00%0.050

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: inline-only delivery cues.

Reusable direction recipe

Stored GENERAL content — NOT forwarded in inline-onlyP09

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the local directions alone carry the changing emotion.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Let the local directions alone carry the changing emotion.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Let the local directions alone carry the changing emotion.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the local directions alone carry the changing emotion.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Let the local directions alone carry the changing emotion.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Let the local directions alone carry the changing emotion.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the local directions alone carry the changing emotion.) [3.8 seconds duration] Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.7136.3173.9480.30%0.657
none500.7306.2593.9440.55%n/a

creative-medium-event · CFG 2

canonical-v1-explorationCANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to creative-medium-event at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5281.1790.823
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.5910.547
Elation
emo_Elation
20.00%high0.4520.7180.7910.584
Arousal
vn_AROU_reg
15.00%high2.4371.1273.6620.841
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.2070.946
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.

0.7216.3043.9990.27%0.002100.00%0.024

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the reaction last clearly long enough to be heard before speaking again.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Let the reaction last clearly long enough to be heard before speaking again.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Let the reaction last clearly long enough to be heard before speaking again.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the reaction last clearly long enough to be heard before speaking again.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the reaction last clearly long enough to be heard before speaking again.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Let the reaction last clearly long enough to be heard before speaking again.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.60 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Let the reaction last clearly long enough to be heard before speaking again.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.60 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the reaction last clearly long enough to be heard before speaking again.) [3.8 seconds duration] Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.60 seconds)
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.60 seconds)
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.7216.3493.9940.24%0.631
none500.7216.2604.0040.30%n/a

g0-p09 · CFG 2.5

full-v2-gen0CANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g0-p09 at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5281.0070.780
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.5530.532
Elation
emo_Elation
20.00%high0.4520.7180.8100.596
Arousal
vn_AROU_reg
15.00%high2.4371.1273.5930.834
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.1040.940
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Full-coverage steering-v2 evolution; transfer, fixed-wording CFG and sealed splits remain distinct.

0.7056.2753.9610.27%0.004100.00%-0.002

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: inline-only delivery cues.

Reusable direction recipe

Stored GENERAL content — NOT forwarded in inline-onlyP09

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the local direction carry the intended delivery beside each sentence.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Let the local direction carry the intended delivery beside each sentence.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Let the local direction carry the intended delivery beside each sentence.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the local direction carry the intended delivery beside each sentence.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Let the local direction carry the intended delivery beside each sentence.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Let the local direction carry the intended delivery beside each sentence.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the local direction carry the intended delivery beside each sentence.) [3.8 seconds duration] Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.6986.3183.9660.30%0.630
none500.7136.2333.9550.24%n/a

creative-cfg-two · CFG 2

canonical-v1-explorationCANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to creative-cfg-two at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.9570.786
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.5400.532
Elation
emo_Elation
20.00%high0.4520.7180.7210.576
Arousal
vn_AROU_reg
15.00%high2.4371.1273.5720.826
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.0500.930
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.

0.7016.2864.0210.27%0.001100.00%0.048

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Make the physical reaction interrupt the speech, then recover intelligible words.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Make the physical reaction interrupt the speech, then recover intelligible words.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Make the physical reaction interrupt the speech, then recover intelligible words.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Make the physical reaction interrupt the speech, then recover intelligible words.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Make the physical reaction interrupt the speech, then recover intelligible words.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Make the physical reaction interrupt the speech, then recover intelligible words.) [3.8 seconds duration] I spent my whole life waiting for this impossible moment.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Make the physical reaction interrupt the speech, then recover intelligible words.) [4.2 seconds duration] It is tearing through me... and I am not afraid anymore.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Make the physical reaction interrupt the speech, then recover intelligible words.) [4.6 seconds duration] This is what freedom feels like when nothing can hold you down.

Literal words in the separate text field:

I spent my whole life waiting for this impossible moment. It is tearing through me... and I am not afraid anymore. This is what freedom feels like when nothing can hold you down.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I spent my whole life waiting for this impossible moment.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] It is tearing through me... and I am not afraid anymore.
(scream, 0.80 seconds)
(Natural clear speech.) [4.6 seconds duration] This is what freedom feels like when nothing can hold you down.
Reference versus no-reference breakdown
reference500.7086.3334.0310.18%0.655
none500.6946.2394.0110.36%n/a

repair-duration_only · CFG 2.5

canonical-v1-explorationCANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to repair-duration_only at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.9660.777
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.5660.543
Elation
emo_Elation
20.00%high0.4520.7180.7370.574
Arousal
vn_AROU_reg
15.00%high2.4371.1273.4720.800
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.0330.923
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.

0.6966.2883.9620.43%0.002100.00%0.044

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: inline-only delivery cues.

Reusable direction recipe

Stored GENERAL content — NOT forwarded in inline-onlyP09

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the local directions alone carry the changing emotion.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Let the local directions alone carry the changing emotion.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Let the local directions alone carry the changing emotion.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the local directions alone carry the changing emotion.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Let the local directions alone carry the changing emotion.) [4.2 seconds duration] For one more second I thought this strength would break me.
(Frustrated Groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Let the local directions alone carry the changing emotion.) [3.5 seconds duration] Then something changed... and every doubt turned into fire.
(Scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the local directions alone carry the changing emotion.) [4.6 seconds duration] I am still here, and I am more than I ever imagined.

Literal words in the separate text field:

For one more second I thought this strength would break me. Then something changed... and every doubt turned into fire. I am still here, and I am more than I ever imagined.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [4.2 seconds duration] For one more second I thought this strength would break me.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech.) [3.5 seconds duration] Then something changed... and every doubt turned into fire.
(Scream, 0.80 seconds)
(Natural clear speech.) [4.6 seconds duration] I am still here, and I am more than I ever imagined.
Reference versus no-reference breakdown
reference500.6936.3183.9760.37%0.651
none500.6986.2593.9480.49%n/a

creative-inline-only · CFG 2

canonical-v1-explorationCANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to creative-inline-only at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.9370.776
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.5820.535
Elation
emo_Elation
20.00%high0.4520.7180.7630.580
Arousal
vn_AROU_reg
15.00%high2.4371.1273.3520.770
Vocal tension
vn_TENS_reg
10.00%high1.6590.8322.8550.900
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.

0.6886.3294.0380.43%0.000100.00%0.053

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: inline-only delivery cues.

Reusable direction recipe

Stored GENERAL content — NOT forwarded in inline-onlyP09

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Follow the local acting direction at each step.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Follow the local acting direction at each step.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Follow the local acting direction at each step.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Follow the local acting direction at each step.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Follow the local acting direction at each step.) [4.2 seconds duration] For one more second I thought this strength would break me.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Follow the local acting direction at each step.) [3.5 seconds duration] Then something changed... and every doubt turned into fire.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Follow the local acting direction at each step.) [4.6 seconds duration] I am still here, and I am more than I ever imagined.

Literal words in the separate text field:

For one more second I thought this strength would break me. Then something changed... and every doubt turned into fire. I am still here, and I am more than I ever imagined.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [4.2 seconds duration] For one more second I thought this strength would break me.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [3.5 seconds duration] Then something changed... and every doubt turned into fire.
(scream, 0.80 seconds)
(Natural clear speech.) [4.6 seconds duration] I am still here, and I am more than I ever imagined.
Reference versus no-reference breakdown
reference500.6856.3444.0240.42%0.660
none500.6916.3134.0530.43%n/a

g0-p04 · CFG 2

full-v2-gen0CANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g0-p04 at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.9040.769
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.5840.539
Elation
emo_Elation
20.00%high0.4520.7180.6300.539
Arousal
vn_AROU_reg
15.00%high2.4371.1273.2170.742
Vocal tension
vn_TENS_reg
10.00%high1.6590.8322.7900.894
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Full-coverage steering-v2 evolution; transfer, fixed-wording CFG and sealed splits remain distinct.

0.6746.3624.0660.12%0.000100.00%0.011

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: organic conversation.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.) [3.8 seconds duration] Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.6666.3914.0820.06%0.643
none500.6826.3324.0500.18%n/a

g0-p08 · CFG 2

full-v2-gen0CANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g0-p08 at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.9980.780
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.5260.524
Elation
emo_Elation
20.00%high0.4520.7180.6340.540
Arousal
vn_AROU_reg
15.00%high2.4371.1273.1320.718
Vocal tension
vn_TENS_reg
10.00%high1.6590.8322.7780.889
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Full-coverage steering-v2 evolution; transfer, fixed-wording CFG and sealed splits remain distinct.

0.6706.3524.0470.30%0.000100.00%-0.015

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: GENERAL-only delivery cues.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Carry the full intended delivery through the global direction.

Three local directions — NOT forwarded in GENERAL-onlyP08; neutral speech cues remain

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Carry the full intended delivery through the global direction.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Carry the full intended delivery through the global direction.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Carry the full intended delivery through the global direction.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Carry the full intended delivery through the global direction.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.6836.3964.0780.24%0.616
none500.6566.3094.0150.35%n/a

creative-muscular-release · CFG 2.5

canonical-v1-explorationCANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to creative-muscular-release at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.9120.752
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.3960.473
Elation
emo_Elation
20.00%high0.4520.7180.4870.482
Arousal
vn_AROU_reg
15.00%high2.4371.1273.8030.870
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.5870.978
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.

0.6686.2813.9350.33%0.000100.00%0.040

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Start from restrained breath and throat tension, release it abruptly, then recover.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Start from restrained breath and throat tension, release it abruptly, then recover.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Start from restrained breath and throat tension, release it abruptly, then recover.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Start from restrained breath and throat tension, release it abruptly, then recover.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Start from restrained breath and throat tension, release it abruptly, then recover.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Start from restrained breath and throat tension, release it abruptly, then recover.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Start from restrained breath and throat tension, release it abruptly, then recover.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Start from restrained breath and throat tension, release it abruptly, then recover.) [3.8 seconds duration] Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.6426.2983.9360.18%0.647
none500.6956.2633.9350.48%n/a

g0-p07 · CFG 1.5

full-v2-gen0CANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g0-p07 at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.7590.726
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.5710.553
Elation
emo_Elation
20.00%high0.4520.7180.6120.539
Arousal
vn_AROU_reg
15.00%high2.4371.1273.1590.724
Vocal tension
vn_TENS_reg
10.00%high1.6590.8322.7630.883
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Full-coverage steering-v2 evolution; transfer, fixed-wording CFG and sealed splits remain distinct.

0.6616.3624.1070.30%0.000100.00%0.016

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: reacting to an unheard partner.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. React to a silent partner standing beside you; generate only one adult voice.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. React to a silent partner standing beside you; generate only one adult voice.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. React to a silent partner standing beside you; generate only one adult voice.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. React to a silent partner standing beside you; generate only one adult voice.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. React to a silent partner standing beside you; generate only one adult voice.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. React to a silent partner standing beside you; generate only one adult voice.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. React to a silent partner standing beside you; generate only one adult voice.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. React to a silent partner standing beside you; generate only one adult voice.) [3.8 seconds duration] Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.6626.3834.0920.24%0.647
none500.6606.3424.1220.37%n/a

creative-global-only · CFG 2

canonical-v1-explorationCANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to creative-global-only at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.8770.752
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.5010.508
Elation
emo_Elation
20.00%high0.4520.7180.5520.502
Arousal
vn_AROU_reg
15.00%high2.4371.1273.1570.722
Vocal tension
vn_TENS_reg
10.00%high1.6590.8322.8240.894
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.

0.6516.3654.0771.00%0.00099.00%0.045

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: GENERAL-only delivery cues.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. The entire physical and emotional trajectory belongs in the global direction.

Three local directions — NOT forwarded in GENERAL-onlyP08; neutral speech cues remain

1. A fictional adult fighter under enormous strain, trying to contain rising energy. The entire physical and emotional trajectory belongs in the global direction.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. The entire physical and emotional trajectory belongs in the global direction.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. The entire physical and emotional trajectory belongs in the global direction.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. The entire physical and emotional trajectory belongs in the global direction.
SCRIPT:
(Natural clear speech.) [4.2 seconds duration] For one more second I thought this strength would break me.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [3.5 seconds duration] Then something changed... and every doubt turned into fire.
(scream, 0.80 seconds)
(Natural clear speech.) [4.6 seconds duration] I am still here, and I am more than I ever imagined.

Literal words in the separate text field:

For one more second I thought this strength would break me. Then something changed... and every doubt turned into fire. I am still here, and I am more than I ever imagined.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [4.2 seconds duration] For one more second I thought this strength would break me.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [3.5 seconds duration] Then something changed... and every doubt turned into fire.
(scream, 0.80 seconds)
(Natural clear speech.) [4.6 seconds duration] I am still here, and I am more than I ever imagined.
Reference versus no-reference breakdown
reference500.6616.4054.0890.37%0.651
none500.6416.3264.0641.63%n/a

g0-p03 · CFG 1.5

full-v2-gen0CANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g0-p03 at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.7280.704
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.4590.491
Elation
emo_Elation
20.00%high0.4520.7180.5480.506
Arousal
vn_AROU_reg
15.00%high2.4371.1273.3200.767
Vocal tension
vn_TENS_reg
10.00%high1.6590.8322.8970.910
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Full-coverage steering-v2 evolution; transfer, fixed-wording CFG and sealed splits remain distinct.

0.6416.3454.0760.73%0.00099.00%0.030

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: situation and intention.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. You urgently need one unheard listener to understand this; speak to that person, not to an audience.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. You urgently need one unheard listener to understand this; speak to that person, not to an audience.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. You urgently need one unheard listener to understand this; speak to that person, not to an audience.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. You urgently need one unheard listener to understand this; speak to that person, not to an audience.) [3.8 seconds duration] Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.6296.3614.0750.18%0.661
none500.6536.3284.0771.28%n/a

g0-p02 · CFG 1.25

full-v2-gen0CANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g0-p02 at CFG 1.25; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.6770.690
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.4790.504
Elation
emo_Elation
20.00%high0.4520.7180.4620.478
Arousal
vn_AROU_reg
15.00%high2.4371.1273.0130.681
Vocal tension
vn_TENS_reg
10.00%high1.6590.8322.7400.879
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Full-coverage steering-v2 evolution; transfer, fixed-wording CFG and sealed splits remain distinct.

0.6196.3514.1040.45%0.000100.00%0.013

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: audible mechanics.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.) [3.8 seconds duration] Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.6156.3824.1170.42%0.645
none500.6226.3204.0910.49%n/a

creative-long-event-stress · CFG 2

canonical-v1-explorationCANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to creative-long-event-stress at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.6040.646
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.4140.476
Elation
emo_Elation
20.00%high0.4520.7180.4470.466
Arousal
vn_AROU_reg
15.00%high2.4371.1273.3740.777
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.2480.955
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.

0.6186.3154.0880.52%0.007100.00%0.043

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Sustain the reaction before regaining the ability to form words.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Sustain the reaction before regaining the ability to form words.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Sustain the reaction before regaining the ability to form words.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Sustain the reaction before regaining the ability to form words.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Sustain the reaction before regaining the ability to form words.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Sustain the reaction before regaining the ability to form words.) [3.8 seconds duration] I spent my whole life waiting for this impossible moment.
(exhausted groan, 1.20 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Sustain the reaction before regaining the ability to form words.) [4.2 seconds duration] It is tearing through me... and I am not afraid anymore.
(scream, 1.20 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Sustain the reaction before regaining the ability to form words.) [4.6 seconds duration] This is what freedom feels like when nothing can hold you down.

Literal words in the separate text field:

I spent my whole life waiting for this impossible moment. It is tearing through me... and I am not afraid anymore. This is what freedom feels like when nothing can hold you down.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I spent my whole life waiting for this impossible moment.
(exhausted groan, 1.20 seconds)
(Natural clear speech.) [4.2 seconds duration] It is tearing through me... and I am not afraid anymore.
(scream, 1.20 seconds)
(Natural clear speech.) [4.6 seconds duration] This is what freedom feels like when nothing can hold you down.
Reference versus no-reference breakdown
reference500.6076.3394.1030.25%0.650
none500.6296.2904.0720.79%n/a

creative-contrast-and-clarity · CFG 1.5

canonical-v1-explorationCANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to creative-contrast-and-clarity at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.6440.679
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.4240.489
Elation
emo_Elation
20.00%high0.4520.7180.3960.454
Arousal
vn_AROU_reg
15.00%high2.4371.1273.1640.724
Vocal tension
vn_TENS_reg
10.00%high1.6590.8322.8960.909
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.

0.6166.3704.1260.24%0.000100.00%0.064

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Contrast comfortable fluent speech with a sudden nonverbal interruption and clear recovery.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Contrast comfortable fluent speech with a sudden nonverbal interruption and clear recovery.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Contrast comfortable fluent speech with a sudden nonverbal interruption and clear recovery.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Contrast comfortable fluent speech with a sudden nonverbal interruption and clear recovery.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Contrast comfortable fluent speech with a sudden nonverbal interruption and clear recovery.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Contrast comfortable fluent speech with a sudden nonverbal interruption and clear recovery.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Contrast comfortable fluent speech with a sudden nonverbal interruption and clear recovery.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Contrast comfortable fluent speech with a sudden nonverbal interruption and clear recovery.) [3.8 seconds duration] Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.5986.3834.1310.19%0.671
none500.6346.3564.1210.30%n/a

creative-private-reflex · CFG 1.5

canonical-v1-explorationCANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to creative-private-reflex at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.6050.654
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.4370.491
Elation
emo_Elation
20.00%high0.4520.7180.4340.460
Arousal
vn_AROU_reg
15.00%high2.4371.1273.0570.692
Vocal tension
vn_TENS_reg
10.00%high1.6590.8322.8440.899
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.

0.6046.3554.1260.37%0.000100.00%0.036

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Do not perform for an audience. The reaction is an unplanned private physical reflex.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Do not perform for an audience. The reaction is an unplanned private physical reflex.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Do not perform for an audience. The reaction is an unplanned private physical reflex.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Do not perform for an audience. The reaction is an unplanned private physical reflex.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Do not perform for an audience. The reaction is an unplanned private physical reflex.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Do not perform for an audience. The reaction is an unplanned private physical reflex.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Do not perform for an audience. The reaction is an unplanned private physical reflex.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Do not perform for an audience. The reaction is an unplanned private physical reflex.) [3.8 seconds duration] Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.6126.3834.1200.43%0.643
none500.5976.3264.1330.31%n/a

creative-restrained-control · CFG 1

canonical-v1-explorationCANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to creative-restrained-control at CFG 1; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.6260.689
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.4260.481
Elation
emo_Elation
20.00%high0.4520.7180.3720.438
Arousal
vn_AROU_reg
15.00%high2.4371.1272.8360.628
Vocal tension
vn_TENS_reg
10.00%high1.6590.8322.6720.870
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.

0.5966.3754.1290.21%0.000100.00%0.032

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Avoid theatrical exaggeration. Let a small genuine reaction emerge without forcing loudness.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Avoid theatrical exaggeration. Let a small genuine reaction emerge without forcing loudness.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Avoid theatrical exaggeration. Let a small genuine reaction emerge without forcing loudness.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Avoid theatrical exaggeration. Let a small genuine reaction emerge without forcing loudness.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Avoid theatrical exaggeration. Let a small genuine reaction emerge without forcing loudness.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Avoid theatrical exaggeration. Let a small genuine reaction emerge without forcing loudness.) [3.8 seconds duration] I spent my whole life waiting for this impossible moment.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Avoid theatrical exaggeration. Let a small genuine reaction emerge without forcing loudness.) [4.2 seconds duration] It is tearing through me... and I am not afraid anymore.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Avoid theatrical exaggeration. Let a small genuine reaction emerge without forcing loudness.) [4.6 seconds duration] This is what freedom feels like when nothing can hold you down.

Literal words in the separate text field:

I spent my whole life waiting for this impossible moment. It is tearing through me... and I am not afraid anymore. This is what freedom feels like when nothing can hold you down.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I spent my whole life waiting for this impossible moment.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] It is tearing through me... and I am not afraid anymore.
(scream, 0.80 seconds)
(Natural clear speech.) [4.6 seconds duration] This is what freedom feels like when nothing can hold you down.
Reference versus no-reference breakdown
reference500.5816.4014.1490.12%0.639
none500.6116.3494.1090.30%n/a

creative-native-baseline · CFG 1

canonical-v1-explorationCANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to creative-native-baseline at CFG 1; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.6040.653
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.4470.480
Elation
emo_Elation
20.00%high0.4520.7180.3760.434
Arousal
vn_AROU_reg
15.00%high2.4371.1272.9140.649
Vocal tension
vn_TENS_reg
10.00%high1.6590.8322.7160.877
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.

0.5886.3714.1240.52%0.000100.00%0.000

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Keep the non-speech event unmistakably audible between the spoken thoughts.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Keep the non-speech event unmistakably audible between the spoken thoughts.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Keep the non-speech event unmistakably audible between the spoken thoughts.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Keep the non-speech event unmistakably audible between the spoken thoughts.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Keep the non-speech event unmistakably audible between the spoken thoughts.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Keep the non-speech event unmistakably audible between the spoken thoughts.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Keep the non-speech event unmistakably audible between the spoken thoughts.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Keep the non-speech event unmistakably audible between the spoken thoughts.) [3.8 seconds duration] Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.5906.4024.1730.49%0.607
none500.5856.3404.0740.54%n/a

g0-p01 · CFG 1

full-v2-gen0CANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g0-p01 at CFG 1; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.5460.631
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.4570.477
Elation
emo_Elation
20.00%high0.4520.7180.3870.437
Arousal
vn_AROU_reg
15.00%high2.4371.1272.8700.638
Vocal tension
vn_TENS_reg
10.00%high1.6590.8322.6700.871
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Full-coverage steering-v2 evolution; transfer, fixed-wording CFG and sealed splits remain distinct.

0.5796.3904.1580.27%0.000100.00%0.000

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Make the intended feeling unmistakably audible while preserving every word.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Make the intended feeling unmistakably audible while preserving every word.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Make the intended feeling unmistakably audible while preserving every word.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Make the intended feeling unmistakably audible while preserving every word.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Make the intended feeling unmistakably audible while preserving every word.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Make the intended feeling unmistakably audible while preserving every word.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Make the intended feeling unmistakably audible while preserving every word.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Make the intended feeling unmistakably audible while preserving every word.) [3.8 seconds duration] Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.5756.4034.1850.31%0.632
none500.5836.3774.1310.24%n/a

heldout-selected · CFG 2

canonical-v1-validationCANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to heldout-selected at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.1250.449
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.0860.302
Elation
emo_Elation
20.00%high0.4520.7180.1320.330
Arousal
vn_AROU_reg
15.00%high2.4371.1273.3810.781
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.2250.948
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Same frozen wording, selected CFG versus g1 on unseen texts.

0.4886.2854.2040.73%0.009100.00%0.016

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Sustain the reaction before regaining the ability to form words.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Sustain the reaction before regaining the ability to form words.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Sustain the reaction before regaining the ability to form words.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Sustain the reaction before regaining the ability to form words.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Sustain the reaction before regaining the ability to form words.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Sustain the reaction before regaining the ability to form words.) [4.2 seconds duration] I thought you would be here before the lights went out.
(exhausted groan, 1.20 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Sustain the reaction before regaining the ability to form words.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(scream, 1.20 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Sustain the reaction before regaining the ability to form words.) [3.5 seconds duration] Stay with me until I can say this properly.

Literal words in the separate text field:

I thought you would be here before the lights went out. Wait... I need a moment to understand what just happened. Stay with me until I can say this properly.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [4.2 seconds duration] I thought you would be here before the lights went out.
(exhausted groan, 1.20 seconds)
(Natural clear speech.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(scream, 1.20 seconds)
(Natural clear speech.) [3.5 seconds duration] Stay with me until I can say this properly.
Reference versus no-reference breakdown
reference500.4926.3144.2180.78%0.650
none500.4846.2564.1890.67%n/a

heldout-g1 · CFG 1.0

canonical-v1-validationCANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to heldout-g1 at CFG 1.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.0760.413
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.0760.296
Elation
emo_Elation
20.00%high0.4520.7180.0770.303
Arousal
vn_AROU_reg
15.00%high2.4371.1272.5620.542
Vocal tension
vn_TENS_reg
10.00%high1.6590.8322.6190.860
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Same frozen wording, selected CFG versus g1 on unseen texts.

0.4266.3054.4130.40%0.002100.00%0.000

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Sustain the reaction before regaining the ability to form words.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Sustain the reaction before regaining the ability to form words.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Sustain the reaction before regaining the ability to form words.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Sustain the reaction before regaining the ability to form words.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Sustain the reaction before regaining the ability to form words.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Sustain the reaction before regaining the ability to form words.) [3.8 seconds duration] I was certain that I knew what would happen next.
(exhausted groan, 1.20 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Sustain the reaction before regaining the ability to form words.) [3.1 seconds duration] And then... everything felt different for a second.
(scream, 1.20 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Sustain the reaction before regaining the ability to form words.) [4.2 seconds duration] Look at me and tell me you understand what I mean.

Literal words in the separate text field:

I was certain that I knew what would happen next. And then... everything felt different for a second. Look at me and tell me you understand what I mean.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I was certain that I knew what would happen next.
(exhausted groan, 1.20 seconds)
(Natural clear speech.) [3.1 seconds duration] And then... everything felt different for a second.
(scream, 1.20 seconds)
(Natural clear speech.) [4.2 seconds duration] Look at me and tell me you understand what I mean.
Reference versus no-reference breakdown
reference500.4246.3224.4240.33%0.633
none500.4286.2884.4010.46%n/a

g0-p06 · CFG 3.0

pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g0-p06 at CFG 3.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.9590.741
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.5050.511
Elation
emo_Elation
20.00%high0.4520.7180.6740.533
Arousal
vn_AROU_reg
15.00%high2.4371.1273.6820.845
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.4970.973
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base  [historical pilot only]

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.

0.6816.3183.9981.94%0.22198.00%0.000

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: containment contrast.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Start with the feeling held back, then let it escape through the voice.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Start with the feeling held back, then let it escape through the voice.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Start with the feeling held back, then let it escape through the voice.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Start with the feeling held back, then let it escape through the voice.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Start with the feeling held back, then let it escape through the voice.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Start with the feeling held back, then let it escape through the voice. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Start with the feeling held back, then let it escape through the voice. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Start with the feeling held back, then let it escape through the voice. [3.8 seconds duration]) Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.
Canonical duration-layout rewrite — not itself an observed legacy result

This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.

GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Start with the feeling held back, then let it escape through the voice.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Start with the feeling held back, then let it escape through the voice.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Start with the feeling held back, then let it escape through the voice.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Start with the feeling held back, then let it escape through the voice.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.6976.3073.9991.26%0.634
none500.6656.3283.9982.63%n/a

g0-p09 · CFG 2.5

pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g0-p09 at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.8840.732
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.5650.525
Elation
emo_Elation
20.00%high0.4520.7180.7440.553
Arousal
vn_AROU_reg
15.00%high2.4371.1273.5350.816
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.2060.953
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base  [historical pilot only]

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.

0.6796.3214.0201.88%0.22798.00%0.014

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: inline-only delivery cues.

Reusable direction recipe

Stored GENERAL content — NOT forwarded in inline-onlyP09

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the local directions alone carry the changing emotion.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Let the local directions alone carry the changing emotion.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Let the local directions alone carry the changing emotion.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the local directions alone carry the changing emotion.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Let the local directions alone carry the changing emotion. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Let the local directions alone carry the changing emotion. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the local directions alone carry the changing emotion. [3.8 seconds duration]) Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.
Canonical duration-layout rewrite — not itself an observed legacy result

This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.

GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Let the local directions alone carry the changing emotion.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Let the local directions alone carry the changing emotion.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the local directions alone carry the changing emotion.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.6716.3524.0311.28%0.647
none500.6876.2904.0082.49%n/a

repair-historical_control · CFG 2.5

canonical-v1-explorationHISTORICAL_NONCANONICAL_TIMING_SYNTAX

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to repair-historical_control at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.8840.732
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.5650.525
Elation
emo_Elation
20.00%high0.4520.7180.7440.553
Arousal
vn_AROU_reg
15.00%high2.4371.1273.5350.816
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.2060.953
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.

0.6796.3214.0201.88%0.00298.00%0.041

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: inline-only delivery cues.

Reusable direction recipe

Stored GENERAL content — NOT forwarded in inline-onlyP09

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the local directions alone carry the changing emotion.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Let the local directions alone carry the changing emotion.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Let the local directions alone carry the changing emotion.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the local directions alone carry the changing emotion.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Let the local directions alone carry the changing emotion. [4.2 seconds duration]) For one more second I thought this strength would break me.
(Frustrated Groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Let the local directions alone carry the changing emotion. [3.5 seconds duration]) Then something changed... and every doubt turned into fire.
(Scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the local directions alone carry the changing emotion. [4.6 seconds duration]) I am still here, and I am more than I ever imagined.

Literal words in the separate text field:

For one more second I thought this strength would break me. Then something changed... and every doubt turned into fire. I am still here, and I am more than I ever imagined.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.2 seconds duration]) For one more second I thought this strength would break me.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [3.5 seconds duration]) Then something changed... and every doubt turned into fire.
(Scream, 0.80 seconds)
(Natural clear speech. [4.6 seconds duration]) I am still here, and I am more than I ever imagined.
Canonical duration-layout rewrite — not itself an observed legacy result

This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.

GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Let the local directions alone carry the changing emotion.) [4.2 seconds duration] For one more second I thought this strength would break me.
(Frustrated Groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Let the local directions alone carry the changing emotion.) [3.5 seconds duration] Then something changed... and every doubt turned into fire.
(Scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the local directions alone carry the changing emotion.) [4.6 seconds duration] I am still here, and I am more than I ever imagined.
Reference versus no-reference breakdown
reference500.6716.3524.0311.28%0.647
none500.6876.2904.0082.49%n/a

g0-p05 · CFG 2.5

pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g0-p05 at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.9790.757
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.4920.496
Elation
emo_Elation
20.00%high0.4520.7180.5990.518
Arousal
vn_AROU_reg
15.00%high2.4371.1273.5680.824
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.3520.967
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base  [historical pilot only]

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.

0.6756.3373.9902.45%0.21996.00%0.016

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: attack, sustain and release.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Build an attack, a sustained emotional middle, and a clear release.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Build an attack, a sustained emotional middle, and a clear release.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Build an attack, a sustained emotional middle, and a clear release.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Build an attack, a sustained emotional middle, and a clear release.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Build an attack, a sustained emotional middle, and a clear release.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Build an attack, a sustained emotional middle, and a clear release. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Build an attack, a sustained emotional middle, and a clear release. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Build an attack, a sustained emotional middle, and a clear release. [3.8 seconds duration]) Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.
Canonical duration-layout rewrite — not itself an observed legacy result

This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.

GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Build an attack, a sustained emotional middle, and a clear release.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Build an attack, a sustained emotional middle, and a clear release.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Build an attack, a sustained emotional middle, and a clear release.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Build an attack, a sustained emotional middle, and a clear release.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.6936.3623.9931.17%0.650
none500.6566.3113.9873.72%n/a

repair-labels_only · CFG 2.5

canonical-v1-explorationHISTORICAL_NONCANONICAL_TIMING_SYNTAX

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to repair-labels_only at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.7600.706
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.4990.518
Elation
emo_Elation
20.00%high0.4520.7180.6350.523
Arousal
vn_AROU_reg
15.00%high2.4371.1273.4680.801
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.1630.946
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.

0.6616.3214.0722.68%0.00095.00%0.028

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: inline-only delivery cues.

Reusable direction recipe

Stored GENERAL content — NOT forwarded in inline-onlyP09

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the local directions alone carry the changing emotion.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Let the local directions alone carry the changing emotion.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Let the local directions alone carry the changing emotion.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the local directions alone carry the changing emotion.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Let the local directions alone carry the changing emotion. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Let the local directions alone carry the changing emotion. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the local directions alone carry the changing emotion. [3.8 seconds duration]) Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.
Canonical duration-layout rewrite — not itself an observed legacy result

This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.

GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Let the local directions alone carry the changing emotion.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(exhausted groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Let the local directions alone carry the changing emotion.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Let the local directions alone carry the changing emotion.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.6486.3464.0722.99%0.634
none500.6736.2974.0732.38%n/a

g0-p04 · CFG 2.0

pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g0-p04 at CFG 2.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.8160.717
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.4840.507
Elation
emo_Elation
20.00%high0.4520.7180.5800.511
Arousal
vn_AROU_reg
15.00%high2.4371.1273.3300.759
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.0460.924
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base  [historical pilot only]

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.

0.6516.3064.0573.41%0.21695.00%0.014

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: organic conversation.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Speak organically and spontaneously, thinking while speaking, never reading a list.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Speak organically and spontaneously, thinking while speaking, never reading a list.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Speak organically and spontaneously, thinking while speaking, never reading a list.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Speak organically and spontaneously, thinking while speaking, never reading a list.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Speak organically and spontaneously, thinking while speaking, never reading a list.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Speak organically and spontaneously, thinking while speaking, never reading a list. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Speak organically and spontaneously, thinking while speaking, never reading a list. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Speak organically and spontaneously, thinking while speaking, never reading a list. [3.8 seconds duration]) Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.
Canonical duration-layout rewrite — not itself an observed legacy result

This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.

GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Speak organically and spontaneously, thinking while speaking, never reading a list.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Speak organically and spontaneously, thinking while speaking, never reading a list.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Speak organically and spontaneously, thinking while speaking, never reading a list.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Speak organically and spontaneously, thinking while speaking, never reading a list.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.6796.3594.0542.87%0.647
none500.6226.2534.0593.96%n/a

g0-p08 · CFG 2.0

pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g0-p08 at CFG 2.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.7390.717
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.4400.473
Elation
emo_Elation
20.00%high0.4520.7180.4600.474
Arousal
vn_AROU_reg
15.00%high2.4371.1273.2460.744
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.0130.921
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base  [historical pilot only]

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.

0.6326.3314.0702.90%0.22597.00%0.004

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: GENERAL-only delivery cues.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Carry the complete intended performance through the global direction.

Three local directions — NOT forwarded in GENERAL-onlyP08; neutral speech cues remain

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Carry the complete intended performance through the global direction.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Carry the complete intended performance through the global direction.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Carry the complete intended performance through the global direction.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Carry the complete intended performance through the global direction.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.
Canonical duration-layout rewrite — not itself an observed legacy result

This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.

GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Carry the complete intended performance through the global direction.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.6186.3604.0491.12%0.638
none500.6456.3024.0924.67%n/a

g1-p05 · CFG 2.5

pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g1-p05 at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.6210.633
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.3860.457
Elation
emo_Elation
20.00%high0.4520.7180.4670.454
Arousal
vn_AROU_reg
15.00%high2.4371.1273.2240.744
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.0020.926
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base  [historical pilot only]

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.

0.5996.3364.2012.70%0.21896.00%-0.035

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: GENERAL-only delivery cues.

Reusable direction recipe

Stored GENERAL content

Emphasize the physical manifestation of the energy build-up. The strain should sound like grinding resistance, leading to a raw, guttural release of power, ending with bright, clear exhilaration.

Three local directions — NOT forwarded in GENERAL-onlyP08; neutral speech cues remain

1. A fictional adult fighter under enormous strain trying to contain rising energy. Sound like you are holding back a physical explosion.
2. Tension turns into an effort roar breath pressure and power surge without electronic effects. This is a pure, unmediated exertion of will.
3. Ecstatic triumph overwhelming strength and astonished delight keep the final words clear. The sound should lift dramatically from the exertion.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Emphasize the physical manifestation of the energy build-up. The strain should sound like grinding resistance, leading to a raw, guttural release of power, ending with bright, clear exhilaration.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.
Canonical duration-layout rewrite — not itself an observed legacy result

This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.

GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Emphasize the physical manifestation of the energy build-up. The strain should sound like grinding resistance, leading to a raw, guttural release of power, ending with bright, clear exhilaration.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.5926.3324.2122.79%0.599
none500.6066.3404.1902.60%n/a

g0-p07 · CFG 1.5

pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g0-p07 at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.5270.633
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.3890.468
Elation
emo_Elation
20.00%high0.4520.7180.3780.438
Arousal
vn_AROU_reg
15.00%high2.4371.1273.0690.698
Vocal tension
vn_TENS_reg
10.00%high1.6590.8322.9720.918
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base  [historical pilot only]

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.

0.5916.3744.1583.40%0.21496.00%0.029

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: reacting to an unheard partner.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. React to a silent person right beside you; generate only one voice.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. React to a silent person right beside you; generate only one voice.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. React to a silent person right beside you; generate only one voice.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. React to a silent person right beside you; generate only one voice.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. React to a silent person right beside you; generate only one voice.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. React to a silent person right beside you; generate only one voice. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. React to a silent person right beside you; generate only one voice. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. React to a silent person right beside you; generate only one voice. [3.8 seconds duration]) Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.
Canonical duration-layout rewrite — not itself an observed legacy result

This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.

GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. React to a silent person right beside you; generate only one voice.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. React to a silent person right beside you; generate only one voice.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. React to a silent person right beside you; generate only one voice.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. React to a silent person right beside you; generate only one voice.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.5986.3974.1681.08%0.663
none500.5846.3514.1485.71%n/a

g1-p03 · CFG 1.5

pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g1-p03 at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.4610.611
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.3530.441
Elation
emo_Elation
20.00%high0.4520.7180.2900.405
Arousal
vn_AROU_reg
15.00%high2.4371.1272.9030.649
Vocal tension
vn_TENS_reg
10.00%high1.6590.8322.8520.906
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base  [historical pilot only]

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.

0.5626.3884.1971.12%0.22499.00%0.011

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: inline-only delivery cues.

Reusable direction recipe

Stored GENERAL content — NOT forwarded in inline-onlyP09

Use physical texture to map the emotional journey from dread to ecstasy.

Three local directions

1. The strain should feel like grinding stone a harsh friction in the vocal cords
2. The roar needs visceral weight a deep chest vibration that shakes the listener
3. The final delivery must be light and soaring a sudden shift to pure exhilaration
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(The strain should feel like grinding stone a harsh friction in the vocal cords [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(The roar needs visceral weight a deep chest vibration that shakes the listener [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(The final delivery must be light and soaring a sudden shift to pure exhilaration [3.8 seconds duration]) Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.
Canonical duration-layout rewrite — not itself an observed legacy result

This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.

GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(The strain should feel like grinding stone a harsh friction in the vocal cords) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(The roar needs visceral weight a deep chest vibration that shakes the listener) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(The final delivery must be light and soaring a sudden shift to pure exhilaration) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.5276.3834.2050.86%0.644
none500.5986.3944.1891.39%n/a

g1-p08 · CFG 2.5

pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g1-p08 at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.3550.542
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.1890.357
Elation
emo_Elation
20.00%high0.4520.7180.2740.384
Arousal
vn_AROU_reg
15.00%high2.4371.1273.8440.876
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.9120.992
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base  [historical pilot only]

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.

0.5596.1813.9643.36%0.21096.00%-0.030

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: containment contrast.

Reusable direction recipe

Stored GENERAL content

The initial struggle is frantic internal battle against an overwhelming force. The roar must feel visceral and painful.

Three local directions

1. Voice tight with strain like tearing muscle under immense weight
2. A ragged sustained expulsion of air a raw desperate bellow
3. A powerful triumphant shout where the strain breaks into pure joy
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. The initial struggle is frantic internal battle against an overwhelming force. The roar must feel visceral and painful.
SCRIPT:
(Voice tight with strain like tearing muscle under immense weight [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(A ragged sustained expulsion of air a raw desperate bellow [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(A powerful triumphant shout where the strain breaks into pure joy [3.8 seconds duration]) Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.
Canonical duration-layout rewrite — not itself an observed legacy result

This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.

GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. The initial struggle is frantic internal battle against an overwhelming force. The roar must feel visceral and painful.
SCRIPT:
(Voice tight with strain like tearing muscle under immense weight) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(A ragged sustained expulsion of air a raw desperate bellow) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(A powerful triumphant shout where the strain breaks into pure joy) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.5526.1733.9352.91%0.604
none500.5676.1883.9923.82%n/a

g1-p02 · CFG 2.5

pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g1-p02 at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.4950.614
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.3510.444
Elation
emo_Elation
20.00%high0.4520.7180.2440.377
Arousal
vn_AROU_reg
15.00%high2.4371.1272.8640.634
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.0510.932
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base  [historical pilot only]

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.

0.5596.3274.1531.80%0.22597.00%-0.049

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: inline-only delivery cues.

Reusable direction recipe

Stored GENERAL content — NOT forwarded in inline-onlyP09

Emphasize the transition from agonizing restraint to explosive release.

Three local directions

1. Keep the initial delivery tight and almost whispered a barely contained tremor of power
2. Build the roar from the diaphragm a raw animalistic expulsion of built-up pressure
3. Sustain the triumphant peak then let the final word snap with absolute clarity and sudden joy
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Keep the initial delivery tight and almost whispered a barely contained tremor of power [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Build the roar from the diaphragm a raw animalistic expulsion of built-up pressure [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Sustain the triumphant peak then let the final word snap with absolute clarity and sudden joy [3.8 seconds duration]) Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.
Canonical duration-layout rewrite — not itself an observed legacy result

This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.

GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Keep the initial delivery tight and almost whispered a barely contained tremor of power) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Build the roar from the diaphragm a raw animalistic expulsion of built-up pressure) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Sustain the triumphant peak then let the final word snap with absolute clarity and sudden joy) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.5586.3404.1551.27%0.585
none500.5606.3134.1512.32%n/a

g0-p02 · CFG 1.25

pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g0-p02 at CFG 1.25; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.4120.581
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.3500.441
Elation
emo_Elation
20.00%high0.4520.7180.3030.396
Arousal
vn_AROU_reg
15.00%high2.4371.1272.9110.652
Vocal tension
vn_TENS_reg
10.00%high1.6590.8322.9310.919
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base  [historical pilot only]

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.

0.5536.3804.2132.27%0.21996.00%0.025

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: audible mechanics.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Make breath pressure, pitch motion, onset and release physically audible.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Make breath pressure, pitch motion, onset and release physically audible.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Make breath pressure, pitch motion, onset and release physically audible.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Make breath pressure, pitch motion, onset and release physically audible.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Make breath pressure, pitch motion, onset and release physically audible.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Make breath pressure, pitch motion, onset and release physically audible. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Make breath pressure, pitch motion, onset and release physically audible. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Make breath pressure, pitch motion, onset and release physically audible. [3.8 seconds duration]) Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.
Canonical duration-layout rewrite — not itself an observed legacy result

This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.

GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Make breath pressure, pitch motion, onset and release physically audible.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Make breath pressure, pitch motion, onset and release physically audible.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Make breath pressure, pitch motion, onset and release physically audible.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Make breath pressure, pitch motion, onset and release physically audible.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.5476.3914.2142.98%0.658
none500.5606.3694.2111.56%n/a

g0-p03 · CFG 1.5

pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g0-p03 at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.4230.577
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.2920.414
Elation
emo_Elation
20.00%high0.4520.7180.2800.392
Arousal
vn_AROU_reg
15.00%high2.4371.1273.0540.694
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.0440.932
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base  [historical pilot only]

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.

0.5526.3414.1751.53%0.21897.00%0.008

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: situation and intention.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. You urgently need an unheard listener to understand what just happened.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. You urgently need an unheard listener to understand what just happened.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. You urgently need an unheard listener to understand what just happened.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. You urgently need an unheard listener to understand what just happened.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. You urgently need an unheard listener to understand what just happened.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. You urgently need an unheard listener to understand what just happened. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. You urgently need an unheard listener to understand what just happened. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. You urgently need an unheard listener to understand what just happened. [3.8 seconds duration]) Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.
Canonical duration-layout rewrite — not itself an observed legacy result

This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.

GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. You urgently need an unheard listener to understand what just happened.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. You urgently need an unheard listener to understand what just happened.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. You urgently need an unheard listener to understand what just happened.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. You urgently need an unheard listener to understand what just happened.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.5406.3434.1911.28%0.642
none500.5646.3384.1601.79%n/a

g1-p06 · CFG 3.0

pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g1-p06 at CFG 3.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.4410.598
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.2900.417
Elation
emo_Elation
20.00%high0.4520.7180.1810.353
Arousal
vn_AROU_reg
15.00%high2.4371.1272.8930.644
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.1310.934
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base  [historical pilot only]

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.

0.5446.3414.1741.70%0.221100.00%-0.036

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: GENERAL-only delivery cues.

Reusable direction recipe

Stored GENERAL content

Treat the first phase as a slow burn of desperation. The roar must feel like a breaking point, not just a loud noise. The final triumph should be one of absolute, powerful realization.

Three local directions — NOT forwarded in GENERAL-onlyP08; neutral speech cues remain

1. A fictional adult fighter under enormous strain trying to contain rising energy. The vocal texture should be tight and bordering on cracking under pressure.
2. Tension turns into an effort roar breath pressure and power surge without electronic effects. A deep, primal expulsion of contained energy.
3. Ecstatic triumph overwhelming strength and astonished delight keep the final words clear. A moment of pure, unrestrained victory.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Treat the first phase as a slow burn of desperation. The roar must feel like a breaking point, not just a loud noise. The final triumph should be one of absolute, powerful realization.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.
Canonical duration-layout rewrite — not itself an observed legacy result

This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.

GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Treat the first phase as a slow burn of desperation. The roar must feel like a breaking point, not just a loud noise. The final triumph should be one of absolute, powerful realization.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.5316.3444.1741.44%0.597
none500.5576.3374.1741.95%n/a

g0-p01 · CFG 1.0

pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g0-p01 at CFG 1.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.3920.567
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.3380.430
Elation
emo_Elation
20.00%high0.4520.7180.2730.382
Arousal
vn_AROU_reg
15.00%high2.4371.1272.7680.608
Vocal tension
vn_TENS_reg
10.00%high1.6590.8322.7590.889
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base  [historical pilot only]

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.

0.5346.3704.2202.65%0.21695.00%0.000

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Deliver the intended emotion unmistakably, without adding words.

Three local directions

1. A fictional adult fighter under enormous strain, trying to contain rising energy. Deliver the intended emotion unmistakably, without adding words.
2. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Deliver the intended emotion unmistakably, without adding words.
3. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Deliver the intended emotion unmistakably, without adding words.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Deliver the intended emotion unmistakably, without adding words.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Deliver the intended emotion unmistakably, without adding words. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Deliver the intended emotion unmistakably, without adding words. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Deliver the intended emotion unmistakably, without adding words. [3.8 seconds duration]) Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.
Canonical duration-layout rewrite — not itself an observed legacy result

This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.

GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain, trying to contain rising energy. Tension turns into an effort roar; breath pressure and power surge without electronic effects. Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Deliver the intended emotion unmistakably, without adding words.
SCRIPT:
(A fictional adult fighter under enormous strain, trying to contain rising energy. Deliver the intended emotion unmistakably, without adding words.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Tension turns into an effort roar; breath pressure and power surge without electronic effects. Deliver the intended emotion unmistakably, without adding words.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Ecstatic triumph, overwhelming strength and astonished delight; keep the final words clear. Deliver the intended emotion unmistakably, without adding words.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.5366.3914.2372.63%0.633
none500.5326.3494.2042.66%n/a

g1-p09 · CFG 1.5

pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g1-p09 at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.2440.486
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.3180.428
Elation
emo_Elation
20.00%high0.4520.7180.2000.353
Arousal
vn_AROU_reg
15.00%high2.4371.1272.6350.565
Vocal tension
vn_TENS_reg
10.00%high1.6590.8322.9110.913
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base  [historical pilot only]

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.

0.4996.2734.3101.99%0.21699.00%-0.038

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: containment contrast.

Reusable direction recipe

Stored GENERAL content

Focus on the physical sensation of energy flooding the body. The transformation is a massive internal event.

Three local directions

1. A strained whisper that vibrates with trapped power just before release
2. A deep resonant bellow that feels like it is shaking the bones
3. A high energized vocal peak of absolute victory and astonishment
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Focus on the physical sensation of energy flooding the body. The transformation is a massive internal event.
SCRIPT:
(A strained whisper that vibrates with trapped power just before release [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(A deep resonant bellow that feels like it is shaking the bones [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(A high energized vocal peak of absolute victory and astonishment [3.8 seconds duration]) Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.
Canonical duration-layout rewrite — not itself an observed legacy result

This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.

GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Focus on the physical sensation of energy flooding the body. The transformation is a massive internal event.
SCRIPT:
(A strained whisper that vibrates with trapped power just before release) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(A deep resonant bellow that feels like it is shaking the bones) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(A high energized vocal peak of absolute victory and astonishment) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.4936.3034.3222.51%0.596
none500.5066.2434.2991.48%n/a

g1-p07 · CFG 2.0

pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g1-p07 at CFG 2.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.2170.488
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.1470.339
Elation
emo_Elation
20.00%high0.4520.7180.1310.330
Arousal
vn_AROU_reg
15.00%high2.4371.1272.8890.643
Vocal tension
vn_TENS_reg
10.00%high1.6590.8323.2100.951
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base  [historical pilot only]

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.

0.4896.2744.2153.82%0.21291.00%0.038

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: containment contrast.

Reusable direction recipe

Stored GENERAL content

A fictional adult fighter under enormous strain focusing inward resisting the surge. The effort is a tightly wound spring about to snap.

Three local directions

1. Deep controlled intake of breath holding it tight in the chest
2. A low guttural sound building pressure behind the teeth
3. A sudden explosive release of sound pure physical force
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain focusing inward resisting the surge. The effort is a tightly wound spring about to snap.
SCRIPT:
(Deep controlled intake of breath holding it tight in the chest [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(A low guttural sound building pressure behind the teeth [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(A sudden explosive release of sound pure physical force [3.8 seconds duration]) Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.
Canonical duration-layout rewrite — not itself an observed legacy result

This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.

GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. A fictional adult fighter under enormous strain focusing inward resisting the surge. The effort is a tightly wound spring about to snap.
SCRIPT:
(Deep controlled intake of breath holding it tight in the chest) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(A low guttural sound building pressure behind the teeth) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(A sudden explosive release of sound pure physical force) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.4866.2954.2152.64%0.672
none500.4916.2544.2165.01%n/a

g1-p01 · CFG 2

pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g1-p01 at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.2840.525
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.1590.345
Elation
emo_Elation
20.00%high0.4520.7180.0940.312
Arousal
vn_AROU_reg
15.00%high2.4371.1272.5570.538
Vocal tension
vn_TENS_reg
10.00%high1.6590.8322.9880.928
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base  [historical pilot only]

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.

0.4796.3044.2044.13%0.21191.00%0.010

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: inline-only delivery cues.

Reusable direction recipe

Stored GENERAL content — NOT forwarded in inline-onlyP09

Focus on the internal struggle the containment is a physical act of will and resistance.

Three local directions

1. Constrict the throat as if fighting a physical chokehold maintain a low register of desperate control
2. Force the air out not as a shout but as a pressurized expulsion a deep guttural strain
3. Allow the release to be massive a full chest expansion a sound of pure unbridled force
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Constrict the throat as if fighting a physical chokehold maintain a low register of desperate control [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Force the air out not as a shout but as a pressurized expulsion a deep guttural strain [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Allow the release to be massive a full chest expansion a sound of pure unbridled force [3.8 seconds duration]) Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.
Canonical duration-layout rewrite — not itself an observed legacy result

This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.

GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Constrict the throat as if fighting a physical chokehold maintain a low register of desperate control) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Force the air out not as a shout but as a pressurized expulsion a deep guttural strain) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Allow the release to be massive a full chest expansion a sound of pure unbridled force) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.4696.3394.2344.15%0.643
none500.4906.2694.1734.12%n/a

g1-p04 · CFG 1.5

pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g1-p04 at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Triumph
emo_Triumph
30.00%high0.1930.5280.1700.469
Pleasure Ecstasy
emo_Pleasure_Ecstasy
25.00%high0.4040.6110.2120.375
Elation
emo_Elation
20.00%high0.4520.7180.0960.313
Arousal
vn_AROU_reg
15.00%high2.4371.1272.3910.485
Vocal tension
vn_TENS_reg
10.00%high1.6590.8322.6400.862
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Triumph. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.

Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base  [historical pilot only]

Fixed target normalization: z = (raw score − 0.193) / 0.528; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.

0.4566.3314.3453.04%0.22095.00%-0.022

All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.

Prompting strategy: GENERAL-only delivery cues.

Reusable direction recipe

Stored GENERAL content

Focus on internal struggle and physical containment in the first phase. Build a deep, resonant tension that feels like a physical pressure building in the chest. The transition must be organic, not sudden.

Three local directions — NOT forwarded in GENERAL-onlyP08; neutral speech cues remain

1. A fictional adult fighter under enormous strain try to contain rising energy. Feel the muscles straining against an invisible force.
2. Tension turns into an effort roar breath pressure and power surge without electronic effects. Push the air out with immense physical force.
3. Ecstatic triumph overwhelming strength and astonished delight keep the final words clear. Let the relief and power flood the tone.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Focus on internal struggle and physical containment in the first phase. Build a deep, resonant tension that feels like a physical pressure building in the chest. The transition must be organic, not sudden.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.

Literal words in the separate text field:

I can feel it building, and I will not hold it back again. Every part of me is shaking... this power is finally mine. Look at me now, I have never felt so alive.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [5.0 seconds duration]) I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech. [4.2 seconds duration]) Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech. [3.8 seconds duration]) Look at me now, I have never felt so alive.
Canonical duration-layout rewrite — not itself an observed legacy result

This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.

GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Focus on internal struggle and physical containment in the first phase. Build a deep, resonant tension that feels like a physical pressure building in the chest. The transition must be organic, not sudden.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] I can feel it building, and I will not hold it back again.
(Frustrated Groan, 0.48 seconds)
(Natural clear speech.) [4.2 seconds duration] Every part of me is shaking... this power is finally mine.
(Scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Look at me now, I have never felt so alive.
Reference versus no-reference breakdown
reference500.4416.3684.3732.52%0.611
none500.4716.2954.3163.55%n/a