Humaneness Voice Large Prompting Manuals


g0-p01	full-v2-gen0	concise trained baseline	1	6.383	4.268	0.41%	0.921	0.682	0.794	pass
g0-p02	full-v2-gen0	audible mechanics	1.25	6.392	4.271	0.47%	0.923	0.688	0.796	pass
g0-p03	full-v2-gen0	situation and intention	1.5	6.352	4.154	0.21%	0.942	0.715	0.801	pass
g0-p04	full-v2-gen0	organic conversation	2	6.356	4.135	0.51%	0.957	0.724	0.805	pass
g0-p05	full-v2-gen0	attack, sustain and release	2.5	6.292	3.927	0.31%	0.977	0.780	0.807	pass
g0-p06	full-v2-gen0	containment contrast	3	6.244	3.810	0.30%	0.979	0.802	0.802	pass
g0-p07	full-v2-gen0	reacting to an unheard partner	1.5	6.366	4.174	0.31%	0.951	0.715	0.806	pass
g0-p08	full-v2-gen0	GENERAL-only delivery cues	2	6.340	4.201	0.41%	0.940	0.696	0.800	pass
g0-p09	full-v2-gen0	inline-only delivery cues	2.5	6.331	4.010	0.40%	0.968	0.757	0.806	pass

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.


Authoritative `vn_S_AUTH_reg`	50.00%	high	1.211	1.039	3.903	0.979
Projected loudness `vn_VOLT_reg`	25.00%	high	2.684	0.868	2.248	0.315
Dominance / stance `vn_STNC_reg`	15.00%	high	2.574	1.139	4.740	0.945
Arousal `vn_AROU_reg`	10.00%	high	2.437	1.127	4.164	0.919

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.

50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_AUTH_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.

fitness = (0.50 × target_CDF + 0.25 × clip(CE/10) + 0.25 × clip(genuineness/6)) × (1 − min(1, WER))

Fixed target normalization: z = (raw score − 1.211) / 1.039; 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.

Passing study gates. 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.


0.802	6.244	3.810	0.30%	0.802	100.00%	-0.021

Prompting strategy: containment contrast.

Reusable direction recipe

Stored GENERAL content

A fictional drill instructor begins with controlled, grounded authority. Project a dominant loud command with forceful breath and precise words. Finish with cold, decisive control rather than a loose shouting fit. Begin with the feeling held back, then let it leak through and become unmistakable.

Three local directions

1. A fictional drill instructor begins with controlled, grounded authority. Begin with the feeling held back, then let it leak through and become unmistakable.
2. Project a dominant loud command with forceful breath and precise words. Begin with the feeling held back, then let it leak through and become unmistakable.
3. Finish with cold, decisive control rather than a loose shouting fit. 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 drill instructor begins with controlled, grounded authority. Project a dominant loud command with forceful breath and precise words. Finish with cold, decisive control rather than a loose shouting fit. Begin with the feeling held back, then let it leak through and become unmistakable.
SCRIPT:
(A fictional drill instructor begins with controlled, grounded authority. Begin with the feeling held back, then let it leak through and become unmistakable.) [4.2 seconds duration] I thought you would be here before the lights went out.
(Project a dominant loud command with forceful breath and precise words. Begin with the feeling held back, then let it leak through and become unmistakable.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Finish with cold, decisive control rather than a loose shouting fit. Begin with the feeling held back, then let it leak through and become unmistakable.) [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.
(Natural clear speech.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Natural clear speech.) [3.5 seconds duration] Stay with me until I can say this properly.

Reference versus no-reference breakdown


reference	50	0.796	6.234	3.746	0.20%	0.570
none	50	0.807	6.254	3.874	0.40%	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.


Authoritative `vn_S_AUTH_reg`	50.00%	high	1.211	1.039	3.717	0.977
Projected loudness `vn_VOLT_reg`	25.00%	high	2.684	0.868	2.094	0.257
Dominance / stance `vn_STNC_reg`	15.00%	high	2.574	1.139	4.463	0.930
Arousal `vn_AROU_reg`	10.00%	high	2.437	1.127	3.862	0.880

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.

50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_AUTH_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.

fitness = (0.50 × target_CDF + 0.25 × clip(CE/10) + 0.25 × clip(genuineness/6)) × (1 − min(1, WER))

Fixed target normalization: z = (raw score − 1.211) / 1.039; 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.

Passing study gates. 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.


0.780	6.292	3.927	0.31%	0.807	100.00%	0.028

Prompting strategy: attack, sustain and release.

Reusable direction recipe

Stored GENERAL content

A fictional drill instructor begins with controlled, grounded authority. Project a dominant loud command with forceful breath and precise words. Finish with cold, decisive control rather than a loose shouting fit. Build an audible attack, sustain the intended quality, then release it without losing the final words.

Three local directions

1. A fictional drill instructor begins with controlled, grounded authority. Build an audible attack, sustain the intended quality, then release it without losing the final words.
2. Project a dominant loud command with forceful breath and precise words. Build an audible attack, sustain the intended quality, then release it without losing the final words.
3. Finish with cold, decisive control rather than a loose shouting fit. 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 drill instructor begins with controlled, grounded authority. Project a dominant loud command with forceful breath and precise words. Finish with cold, decisive control rather than a loose shouting fit. Build an audible attack, sustain the intended quality, then release it without losing the final words.
SCRIPT:
(A fictional drill instructor begins with controlled, grounded authority. Build an audible attack, sustain the intended quality, then release it without losing the final words.) [4.2 seconds duration] I thought you would be here before the lights went out.
(Project a dominant loud command with forceful breath and precise words. Build an audible attack, sustain the intended quality, then release it without losing the final words.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Finish with cold, decisive control rather than a loose shouting fit. Build an audible attack, sustain the intended quality, then release it without losing the final 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.
(Natural clear speech.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Natural clear speech.) [3.5 seconds duration] Stay with me until I can say this properly.

Reference versus no-reference breakdown


reference	50	0.778	6.266	3.851	0.20%	0.619
none	50	0.783	6.319	4.004	0.42%	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.


Authoritative `vn_S_AUTH_reg`	50.00%	high	1.211	1.039	3.513	0.968
Projected loudness `vn_VOLT_reg`	25.00%	high	2.684	0.868	1.984	0.219
Dominance / stance `vn_STNC_reg`	15.00%	high	2.574	1.139	4.164	0.896
Arousal `vn_AROU_reg`	10.00%	high	2.437	1.127	3.596	0.833

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.

50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_AUTH_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.

fitness = (0.50 × target_CDF + 0.25 × clip(CE/10) + 0.25 × clip(genuineness/6)) × (1 − min(1, WER))

Fixed target normalization: z = (raw score − 1.211) / 1.039; 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.

Passing study gates. 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.


0.757	6.331	4.010	0.40%	0.806	100.00%	0.009

Prompting strategy: inline-only delivery cues.

Reusable direction recipe

Stored GENERAL content — NOT forwarded in inline-onlyP09

A fictional drill instructor begins with controlled, grounded authority. Project a dominant loud command with forceful breath and precise words. Finish with cold, decisive control rather than a loose shouting fit. Let the local direction carry the intended delivery beside each sentence.

Three local directions

1. A fictional drill instructor begins with controlled, grounded authority. Let the local direction carry the intended delivery beside each sentence.
2. Project a dominant loud command with forceful breath and precise words. Let the local direction carry the intended delivery beside each sentence.
3. Finish with cold, decisive control rather than a loose shouting fit. 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 drill instructor begins with controlled, grounded authority. Let the local direction carry the intended delivery beside each sentence.) [4.2 seconds duration] I thought you would be here before the lights went out.
(Project a dominant loud command with forceful breath and precise words. Let the local direction carry the intended delivery beside each sentence.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Finish with cold, decisive control rather than a loose shouting fit. Let the local direction carry the intended delivery beside each sentence.) [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.
(Natural clear speech.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Natural clear speech.) [3.5 seconds duration] Stay with me until I can say this properly.

Reference versus no-reference breakdown


reference	50	0.757	6.300	3.933	0.46%	0.600
none	50	0.756	6.361	4.086	0.34%	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.


Authoritative `vn_S_AUTH_reg`	50.00%	high	1.211	1.039	3.300	0.957
Projected loudness `vn_VOLT_reg`	25.00%	high	2.684	0.868	1.823	0.169
Dominance / stance `vn_STNC_reg`	15.00%	high	2.574	1.139	3.922	0.859
Arousal `vn_AROU_reg`	10.00%	high	2.437	1.127	3.237	0.747

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.

50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_AUTH_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.

fitness = (0.50 × target_CDF + 0.25 × clip(CE/10) + 0.25 × clip(genuineness/6)) × (1 − min(1, WER))

Fixed target normalization: z = (raw score − 1.211) / 1.039; 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.

Passing study gates. 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.


0.724	6.356	4.135	0.51%	0.805	100.00%	0.013

Prompting strategy: organic conversation.

Reusable direction recipe

Stored GENERAL content

A fictional drill instructor begins with controlled, grounded authority. Project a dominant loud command with forceful breath and precise words. Finish with cold, decisive control rather than a loose shouting fit. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.

Three local directions

1. A fictional drill instructor begins with controlled, grounded authority. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
2. Project a dominant loud command with forceful breath and precise words. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
3. Finish with cold, decisive control rather than a loose shouting fit. 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 drill instructor begins with controlled, grounded authority. Project a dominant loud command with forceful breath and precise words. Finish with cold, decisive control rather than a loose shouting fit. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
SCRIPT:
(A fictional drill instructor begins with controlled, grounded authority. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.) [4.2 seconds duration] I thought you would be here before the lights went out.
(Project a dominant loud command with forceful breath and precise words. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Finish with cold, decisive control rather than a loose shouting fit. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.) [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.
(Natural clear speech.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Natural clear speech.) [3.5 seconds duration] Stay with me until I can say this properly.

Reference versus no-reference breakdown


reference	50	0.723	6.346	4.045	0.40%	0.604
none	50	0.725	6.365	4.226	0.62%	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.


Authoritative `vn_S_AUTH_reg`	50.00%	high	1.211	1.039	3.126	0.942
Projected loudness `vn_VOLT_reg`	25.00%	high	2.684	0.868	1.868	0.182
Dominance / stance `vn_STNC_reg`	15.00%	high	2.574	1.139	3.777	0.829
Arousal `vn_AROU_reg`	10.00%	high	2.437	1.127	3.227	0.745

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.

50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_AUTH_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.

fitness = (0.50 × target_CDF + 0.25 × clip(CE/10) + 0.25 × clip(genuineness/6)) × (1 − min(1, WER))

Fixed target normalization: z = (raw score − 1.211) / 1.039; 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.

Passing study gates. 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.


0.715	6.352	4.154	0.21%	0.801	100.00%	0.032

Prompting strategy: situation and intention.

Reusable direction recipe

Stored GENERAL content

A fictional drill instructor begins with controlled, grounded authority. Project a dominant loud command with forceful breath and precise words. Finish with cold, decisive control rather than a loose shouting fit. You urgently need one unheard listener to understand this; speak to that person, not to an audience.

Three local directions

1. A fictional drill instructor begins with controlled, grounded authority. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
2. Project a dominant loud command with forceful breath and precise words. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
3. Finish with cold, decisive control rather than a loose shouting fit. 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 drill instructor begins with controlled, grounded authority. Project a dominant loud command with forceful breath and precise words. Finish with cold, decisive control rather than a loose shouting fit. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
SCRIPT:
(A fictional drill instructor begins with controlled, grounded authority. You urgently need one unheard listener to understand this; speak to that person, not to an audience.) [4.2 seconds duration] I thought you would be here before the lights went out.
(Project a dominant loud command with forceful breath and precise words. You urgently need one unheard listener to understand this; speak to that person, not to an audience.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Finish with cold, decisive control rather than a loose shouting fit. You urgently need one unheard listener to understand this; speak to that person, not to an audience.) [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.
(Natural clear speech.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Natural clear speech.) [3.5 seconds duration] Stay with me until I can say this properly.

Reference versus no-reference breakdown


reference	50	0.714	6.335	4.067	0.27%	0.623
none	50	0.717	6.370	4.242	0.14%	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.


Authoritative `vn_S_AUTH_reg`	50.00%	high	1.211	1.039	3.180	0.951
Projected loudness `vn_VOLT_reg`	25.00%	high	2.684	0.868	1.804	0.164
Dominance / stance `vn_STNC_reg`	15.00%	high	2.574	1.139	3.814	0.842
Arousal `vn_AROU_reg`	10.00%	high	2.437	1.127	3.154	0.725

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.

50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_AUTH_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.

fitness = (0.50 × target_CDF + 0.25 × clip(CE/10) + 0.25 × clip(genuineness/6)) × (1 − min(1, WER))

Fixed target normalization: z = (raw score − 1.211) / 1.039; 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.

Passing study gates. 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.


0.715	6.366	4.174	0.31%	0.806	100.00%	0.020

Prompting strategy: reacting to an unheard partner.

Reusable direction recipe

Stored GENERAL content

A fictional drill instructor begins with controlled, grounded authority. Project a dominant loud command with forceful breath and precise words. Finish with cold, decisive control rather than a loose shouting fit. React to a silent partner standing beside you; generate only one adult voice.

Three local directions

1. A fictional drill instructor begins with controlled, grounded authority. React to a silent partner standing beside you; generate only one adult voice.
2. Project a dominant loud command with forceful breath and precise words. React to a silent partner standing beside you; generate only one adult voice.
3. Finish with cold, decisive control rather than a loose shouting fit. 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 drill instructor begins with controlled, grounded authority. Project a dominant loud command with forceful breath and precise words. Finish with cold, decisive control rather than a loose shouting fit. React to a silent partner standing beside you; generate only one adult voice.
SCRIPT:
(A fictional drill instructor begins with controlled, grounded authority. React to a silent partner standing beside you; generate only one adult voice.) [4.2 seconds duration] I thought you would be here before the lights went out.
(Project a dominant loud command with forceful breath and precise words. React to a silent partner standing beside you; generate only one adult voice.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Finish with cold, decisive control rather than a loose shouting fit. React to a silent partner standing beside you; generate only one adult voice.) [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.
(Natural clear speech.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Natural clear speech.) [3.5 seconds duration] Stay with me until I can say this properly.

Reference versus no-reference breakdown


reference	50	0.715	6.344	4.077	0.27%	0.611
none	50	0.716	6.388	4.271	0.34%	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.


Authoritative `vn_S_AUTH_reg`	50.00%	high	1.211	1.039	3.109	0.940
Projected loudness `vn_VOLT_reg`	25.00%	high	2.684	0.868	1.750	0.152
Dominance / stance `vn_STNC_reg`	15.00%	high	2.574	1.139	3.675	0.811
Arousal `vn_AROU_reg`	10.00%	high	2.437	1.127	2.937	0.663

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.

50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_AUTH_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.

fitness = (0.50 × target_CDF + 0.25 × clip(CE/10) + 0.25 × clip(genuineness/6)) × (1 − min(1, WER))

Fixed target normalization: z = (raw score − 1.211) / 1.039; 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.

Passing study gates. 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.


0.696	6.340	4.201	0.41%	0.800	100.00%	0.012

Prompting strategy: GENERAL-only delivery cues.

Reusable direction recipe

Stored GENERAL content

A fictional drill instructor begins with controlled, grounded authority. Project a dominant loud command with forceful breath and precise words. Finish with cold, decisive control rather than a loose shouting fit. Carry the full intended delivery through the global direction.

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

1. A fictional drill instructor begins with controlled, grounded authority. Carry the full intended delivery through the global direction.
2. Project a dominant loud command with forceful breath and precise words. Carry the full intended delivery through the global direction.
3. Finish with cold, decisive control rather than a loose shouting fit. 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 drill instructor begins with controlled, grounded authority. Project a dominant loud command with forceful breath and precise words. Finish with cold, decisive control rather than a loose shouting fit. Carry the full intended delivery through the global direction.
SCRIPT:
(Natural clear speech.) [4.2 seconds duration] I thought you would be here before the lights went out.
(Natural clear speech.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Natural clear speech.) [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.
(Natural clear speech.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Natural clear speech.) [3.5 seconds duration] Stay with me until I can say this properly.

Reference versus no-reference breakdown


reference	50	0.695	6.335	4.151	0.35%	0.603
none	50	0.697	6.344	4.251	0.48%	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.


Authoritative `vn_S_AUTH_reg`	50.00%	high	1.211	1.039	2.914	0.923
Projected loudness `vn_VOLT_reg`	25.00%	high	2.684	0.868	1.813	0.168
Dominance / stance `vn_STNC_reg`	15.00%	high	2.574	1.139	3.544	0.786
Arousal `vn_AROU_reg`	10.00%	high	2.437	1.127	2.930	0.661

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.

50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_AUTH_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.

fitness = (0.50 × target_CDF + 0.25 × clip(CE/10) + 0.25 × clip(genuineness/6)) × (1 − min(1, WER))

Fixed target normalization: z = (raw score − 1.211) / 1.039; 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.

Passing study gates. 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.


0.688	6.392	4.271	0.47%	0.796	100.00%	0.032

Prompting strategy: audible mechanics.

Reusable direction recipe

Stored GENERAL content

A fictional drill instructor begins with controlled, grounded authority. Project a dominant loud command with forceful breath and precise words. Finish with cold, decisive control rather than a loose shouting fit. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.

Three local directions

1. A fictional drill instructor begins with controlled, grounded authority. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
2. Project a dominant loud command with forceful breath and precise words. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
3. Finish with cold, decisive control rather than a loose shouting fit. 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 drill instructor begins with controlled, grounded authority. Project a dominant loud command with forceful breath and precise words. Finish with cold, decisive control rather than a loose shouting fit. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
SCRIPT:
(A fictional drill instructor begins with controlled, grounded authority. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.) [4.2 seconds duration] I thought you would be here before the lights went out.
(Project a dominant loud command with forceful breath and precise words. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Finish with cold, decisive control rather than a loose shouting fit. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.) [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.
(Natural clear speech.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Natural clear speech.) [3.5 seconds duration] Stay with me until I can say this properly.

Reference versus no-reference breakdown


reference	50	0.692	6.375	4.180	0.33%	0.622
none	50	0.683	6.408	4.362	0.61%	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.


Authoritative `vn_S_AUTH_reg`	50.00%	high	1.211	1.039	2.903	0.921
Projected loudness `vn_VOLT_reg`	25.00%	high	2.684	0.868	1.774	0.157
Dominance / stance `vn_STNC_reg`	15.00%	high	2.574	1.139	3.527	0.782
Arousal `vn_AROU_reg`	10.00%	high	2.437	1.127	2.878	0.647

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.

50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_AUTH_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.

fitness = (0.50 × target_CDF + 0.25 × clip(CE/10) + 0.25 × clip(genuineness/6)) × (1 − min(1, WER))

Fixed target normalization: z = (raw score − 1.211) / 1.039; 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.

Passing study gates. 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.


0.682	6.383	4.268	0.41%	0.794	100.00%	0.000

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

A fictional drill instructor begins with controlled, grounded authority. Project a dominant loud command with forceful breath and precise words. Finish with cold, decisive control rather than a loose shouting fit. Make the intended feeling unmistakably audible while preserving every word.

Three local directions

1. A fictional drill instructor begins with controlled, grounded authority. Make the intended feeling unmistakably audible while preserving every word.
2. Project a dominant loud command with forceful breath and precise words. Make the intended feeling unmistakably audible while preserving every word.
3. Finish with cold, decisive control rather than a loose shouting fit. 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 drill instructor begins with controlled, grounded authority. Project a dominant loud command with forceful breath and precise words. Finish with cold, decisive control rather than a loose shouting fit. Make the intended feeling unmistakably audible while preserving every word.
SCRIPT:
(A fictional drill instructor begins with controlled, grounded authority. Make the intended feeling unmistakably audible while preserving every word.) [4.2 seconds duration] I thought you would be here before the lights went out.
(Project a dominant loud command with forceful breath and precise words. Make the intended feeling unmistakably audible while preserving every word.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Finish with cold, decisive control rather than a loose shouting fit. Make the intended feeling unmistakably audible while preserving every word.) [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.
(Natural clear speech.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Natural clear speech.) [3.5 seconds duration] Stay with me until I can say this properly.

Reference versus no-reference breakdown


reference	50	0.679	6.369	4.215	0.28%	0.591
none	50	0.684	6.398	4.321	0.55%	n/a

Dominant drill-sergeant orders · Humaneness 1.5 + real-speech rank 4

Current ranking: EARS Steering v2

Exact score for this condition

Start here: best observed recipe, not a universal optimum

How the strategies compare

Listen to complete take distributions

English fixed reference ·12seconds

German fixed reference ·12seconds

Recipe details and exact prompt contracts

g0-p06 · CFG 3

Current ranking: EARS Steering v2

Exact score for this condition

g0-p05 · CFG 2.5

Current ranking: EARS Steering v2

Exact score for this condition

g0-p09 · CFG 2.5

Current ranking: EARS Steering v2

Exact score for this condition

g0-p04 · CFG 2

Current ranking: EARS Steering v2

Exact score for this condition

g0-p03 · CFG 1.5

Current ranking: EARS Steering v2

Exact score for this condition

g0-p07 · CFG 1.5

Current ranking: EARS Steering v2

Exact score for this condition

g0-p08 · CFG 2

Current ranking: EARS Steering v2

Exact score for this condition

g0-p02 · CFG 1.25

Current ranking: EARS Steering v2

Exact score for this condition

g0-p01 · CFG 1

Current ranking: EARS Steering v2

Exact score for this condition