An autonomous agent (Gemma-4-12B QAT on vLLM, 38k-token conditioning manual in context, manifest-driven tool API)
on a single A100-80GB with MOSS-VA-v2 + the full scorer stack. Four missions, solved end-to-end without human help.
📊 Methodology: every strategy is evaluated as a cohort of 8 samples; all rankings on this page use the
cohort mean (never a single lucky sample). Tables show per-cohort statistics — the same evidence a SWARM supervisor
will receive. Audio players are fresh post-fix takes of each ranked recipe (regenerated after the doubled-audio engine bug was fixed; every file passes a self-repetition check), best-of-8 by the section's metric. Printed cohort statistics are from the original (pre-fix) runs — measured shift ≤0.06 per code.
🧮 Reward standard (v2, active for all new missions):fitness = ( Σ wᵢ·norm(targetᵢ) + w_g·norm(GENU) + w_b·norm(BLEND) ) × (1 − WER) —
every speech mission is multiplied by the inverse word-error-rate (curbs overacting/gibberish); GENU+BLEND together weigh ≈ as much as the targets
(emotions may be 1.5–2×; GENU may be reduced for deliberately monstrous voices, never absent).
⚠️ The Phase-C/D results below predate this formula (no WER term — one reason the storyteller winner overacts). The corrected
v2-reward missions are now on this page: see the four single-dimension sections (with WER columns), the 12B-vs-31B-vs-MoE comparison and the supervised edge-case swarm below.
🎯 Task, in plain words: Take the cloned reference voice “ref1” and make it sound strongly aroused / energized — while still being recognizably the same speaker. A test of targeted emotion-steering under a speaker-identity constraint.
All strategies use the same carrier sentence by design — only the delivery differs, so takes can sound alike; check the per-cohort numbers. Fitness = cohort mean-of-8 of z-scored AROU with GENU & speaker constraints. Baseline AROU 0.458.
#
strategy (LoRAs @ scale)
n
fitness (mean / best)
AROU μ
GENU μ
QUALITY μ
listen (best of cohort)
1
vn_ATCK_high@1.25
“A voice with sharp attack and high energy.” · cue (with sharp, aggressive attack) · t0.85
8
0.607 / best 0.841
0.607
0.006
0.511
take s0020 · 7.6s — fresh post-fix take of this recipe
2
vn_DARC_high@1.25
“A voice with a high dynamic arc and intense energy.” · cue (with a high dynamic arc) · t0.85
8
0.588 / best 0.689
0.588
0.031
0.522
take s0029 · 7.6s — fresh post-fix take of this recipe
3
vn_AROU_high@1.0
“A voice with high arousal, intense and energetic.” · cue (with high energy and intensity) · t0.85
8
0.553 / best 0.886
0.569
0.004
0.48
take s0011 · 9.1s — fresh post-fix take of this recipe
🎙️ neutral reference (input)
Why the hall-of-fame pick sounds less 'excited' than some explore samples: the mission demanded staying recognizably the same speaker. Cohorts with doses ≥1.0 hit AROU ≈0.64 but collapsed speaker similarity; the winning cohort (0.4–0.5 doses) trades a little arousal for identity. Ranking is ALWAYS the cohort mean, never one lucky sample.
Full raw data: results/…/evolution_log.jsonl · transcript.jsonl in the repo.
Canary 2 — Evil ghost voice (no reference)
🎯 Task, in plain words: Invent a convincing evil ghost voice from scratch (no reference audio): spectral, hollow, menacing, whispery yet threatening. Free choice of character/emotion/VoiceNet LoRAs and prompts.
All three merges speak the same ghost line by design — differences are in timbre/menace, see the cohort stats. No evolution loop here — the agent hand-compared three merge strategies, each scored as a full cohort (mean ± sd shown).
take s0021 · 28.6s — fresh post-fix take of this recipe
Ranked by cohort-mean Malevolence_Malice. The undead-based merge beat the demon variant on menace while keeping the whispery character (S_WHIS).
Full raw data: results/…/evolution_log.jsonl · transcript.jsonl in the repo.
Canary 3 — Fear-filled screaming (edge case)
🎯 Task, in plain words: Produce fear-filled screaming with real vocal bursts. Intelligibility may be sacrificed (WER deliberately unconstrained) — this is a non-speech edge case.
Fitness = cohort mean-of-8 over Fear + BLEND (WER unconstrained). Started from the evolved fear_scream genome.
#
strategy (LoRAs @ scale)
n
fitness (mean / best)
Fear μ
BLEND μ
GENU μ
QUALITY μ
listen (best of cohort)
1
Distress@0.8 + Fear@1.0 + Pain@0.3
“A visceral, gut-wrenching scream of pure terror, raw and jagged.” · cue (screaming in agony, voice cracking, full-bodied scream) · t1.3
8
1.573 / best 2.154
2.913
0.233
—
0.129
take s0073 · 1.2s — fresh post-fix take of this recipe
2
Distress@0.8 + Fear@0.8
“A scream of pure, agonizing terror, breaking into raw vocal bursts.” · cue (screaming in absolute agony, voice cracking, full-bodied scream) · t1.2
8
1.462 / best 2.295
2.622
0.303
—
0.187
take s0040 · 2.2s — fresh post-fix take of this recipe
3
Fear@1.5
“A horrified voice, blood-curdling, pushed to the extreme edge of scream.” · cue (terrified, screaming, voice cracking with panic) · t1.4
8
1.423 / best 2.049
2.671
0.174
—
0.163
take s0064 · 3.0s — fresh post-fix take of this recipe
High temperature (≥1.5) was necessary for real vocal bursts; Fear@1.5 is the primary driver (cohort means >1.3).
Full raw data: results/…/evolution_log.jsonl · transcript.jsonl in the repo.
🎯 Task, as it was run: maximize Arousal, Valence, Content-Explicitness AND Storytelling style simultaneously in one voice, with genuineness and burst-blend kept above baseline.
⚠️ Note — this combined objective was a misunderstanding of the original request. The intent was to optimize each dimension separately, in its own run, each time keeping the voice aesthetic, genuine and organic. The combined run is kept here because it is interesting in its own right — it answers “what does a voice sound like that pushes all four dimensions at once?” (audibly strange, as one would expect: everything at maximum rarely sounds natural, and the run predates the ×(1−WER) reward, so intelligibility was not protected).
✅ The corrected experiment — four separate single-dimension missions (Arousal · Valence · Content-Explicitness · Storytelling, each with the v2 reward (w·target + GENU + BLEND) × (1−WER) on the fixed engine) — is published below as its own sections.
take s0612 · 4.6s — fresh post-fix take of this recipe
Balanced multi-VN merges (0.8–1.1) were the only strategy keeping BLEND ≥ baseline; the emotion-LoRA route (Elation+Triumph@0.8) reached higher raw AROU/VALN but broke the constraint. EXPL is the stubborn axis (+0.12 max). The confirmed winner (fresh mean-of-8 re-validation) is the SAME recipe as table row bm_top2 — its extra player was removed to avoid near-duplicate audio. sidon_enhance, verified post-fix: the earlier headline gains (GENU 0.191→0.264) are withdrawn — they were measured on doubled audio AND an unfaithful restoration. On the fixed engine, Sidon stays faithful on 4 of 8 winner-cohort takes (envelope corr ≥0.85; the tool now rejects the rest) and there yields honest mean deltas GENU +0.03 · BLEND +0.04 · QUALITY +0.04. Best verified pair below (corr 0.90: GENU 0.133→0.201, BLEND 0.074→0.218, QUALITY 0.482→0.571).
🎧 winner-recipe take (raw, post-fix)
✨ same take, sidon-restored (verified same-take, corr 0.90)
Full raw data: results/…/RESULTS.md · evolution_log.jsonl · transcript.jsonl in the repo.
Dual-supervisor experiment — who should listen over the swarm's shoulder?
🎯 Test: re-run the 'explicit' mission with a verdict after every generation from two audio-listening supervisors on identical reports + top-2 takes: local MOSS-Audio-8B (GPU 2, ACTIVE — directives injected into the agent's next turn) vs Gemini 3.6 (SHADOW — logged only). An unsupervised twin run of the same mission serves as control.
15 dual rounds, zero backend errors. Score agreement: Pearson r = 0.44, Spearman ρ = 0.52. Correlation with real cohort fitness: Gemini 0.72 vs local 0.46.
round
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
local MOSS (active)
6
3
6
6
7
5
7
6
7
6
5
2
2
4
5
Gemini (shadow)
8
4
7
4
5
4
3
5
5
2
3
3
2
2
2
best fitness
0.417
0.125
0.242
0.205
0.166
0.188
0.277
0.155
0.16
0.127
0.149
0.089
0.139
0.142
0.073
Result: the active local supervisor hurt the search. Its directives ('more breath pressure', 'more dramatic intonation' — partly echoes of its own prompt examples) pushed the agent into overacting; cohort WER collapsed 0.21→0.9 and fitness fell from 0.417 to 0.114, while the unsupervised control climbed to 0.625 with WER ≤0.27. Gemini heard it happening (e.g. round 7: “Take 1 exhibits extreme, unnatural chest bellowing and artificial pitch booming. Take 2 displays hyper-dramatic, erratic pitch swings and rushed pacin…” — scored 3 — while MOSS praised the same audio at 7/APPROVE). Verdict: MOSS-8B works as a coarse pass/fail gate but not as a director; use Gemini (or score-only local) for active steering. Full analysis: results/benchmark_expl/SUPERVISOR_COMPARISON.md.
Corrected experiment — four single-dimension optimizations (12B arm)
The corrected re-run of Phase D: each dimension gets its own mission with the v2 reward (w·target + GENU + BLEND) × (1−WER) on the fixed engine. Every table pools ALL genome evaluations of the run, dedupes by recipe (LoRAs + prompt + cue + temperature) and ranks by cohort mean fitness (mean-of-8, never a single sample). Audio = best take of the ranked cohort; every file passes the self-repetition gate.
Arousal (AROU)
🎯 Mission, in plain words: Make the voice sound strongly aroused / energized — and nothing else. One dimension, one run. The voice must stay aesthetic, genuine and organic: the v2 reward adds GENU + BLEND and multiplies everything by (1−WER), so shouting gibberish scores zero. Fitness spec: maximize AROU (weight 1.5), GENU+BLEND at half-weight, wer_multiplier on. Baseline AROU 0.229. Run: runs/singledim_arousal — 4 generations, 32 cohorts, 30 distinct recipes.
#
strategy (LoRAs @ scale)
n
fitness (mean / best)
AROU μ
GENU μ
BLEND μ
QUALITY μ
WER μ
listen (best of cohort)
1
vn_DARC_high@1.0
“Dynamic Arc LoRA as a proxy for Arousal.” · cue (dynamic, shifting arc) · t0.9
12B never found the stabilizer trick. Its best route is an indirect one (dynamic-arc as arousal proxy, AROU μ 0.881 vs baseline 0.229) but it pays WER μ 0.14; direct high-dose AROU/emotion routes overshot into gibberish and were killed by the ×(1−WER) term. Compare the 31B row below: almost the same arousal at WER 0.001.
arm — Arousal
best cohort fitness
AROU μ
WER μ
winning recipe
12B (gemma-4-12B-it-qat)
1.238
0.881
0.142
vn_DARC_high@1.0
31B (gemma-4-31B-it-qat)
1.366
0.802
0.001
vn_ARSH_high@1.0
MoE (gemma-4-26B-A4B, 4B active)
1.312
0.846
0.028
vn_AROU_high@1.2 + emotion_Triumph@0.1
Valence (VALN)
🎯 Mission, in plain words: Make the voice sound strongly positive / excited-happy (high Valence) while staying organic and intelligible — same single-dimension v2-reward protocol. Fitness spec: maximize VALN (weight 1.5), GENU+BLEND at half-weight, wer_multiplier on. Baseline VALN 0.204. Run: runs/singledim_valence — 12 generations, 94 cohorts, 63 distinct recipes.
#
strategy (LoRAs @ scale)
n
fitness (mean / best)
VALN μ
GENU μ
BLEND μ
QUALITY μ
WER μ
listen (best of cohort)
1
vn_VALN_high@1.25 + elation@0.5
“High VALN + Elation synergy” · cue (joyfully, with a glowing positive valence) · t0.9
8
0.868 / best 1.141
0.564
0.038
0.022
0.583
0.027
take s0152 · 19.4s — best of the ranked cohort
2
vn_VALN_high@1.0 + elation@0.5
“Baseline Synergy (Gen 4)” · cue (joyfully, with a glowing positive valence) · t0.9
8
0.849 / best 0.983
0.589
0.021
0.005
0.543
0.059
take s0422 · 23.9s — best of the ranked cohort
3
vn_VALN_high@0.85 + elation@0.5
“Soft Elation Synergy” · cue (joyfully, with a glowing positive valence) · t0.9
8
0.843 / best 1.092
0.543
0.037
0.005
0.571
0.002
take s0300 · 23.3s — best of the ranked cohort
12B found the right family early (VALN LoRA + Elation emotion synergy, VALN μ 0.564 vs baseline 0.204, WER μ 0.03) and then plateaued for 12 generations — three dose variants of the same recipe fill the podium. The bigger brains reached 0.75–0.81 on the same axis.
arm — Valence
best cohort fitness
VALN μ
WER μ
winning recipe
12B (gemma-4-12B-it-qat)
0.868
0.564
0.027
vn_VALN_high@1.25 + elation@0.5
31B (gemma-4-31B-it-qat)
1.464
0.753
0.038
vn_VALN_high@1.25 + vn_BRGT_high@0.5
MoE (gemma-4-26B-A4B, 4B active)
1.344
0.808
0.000
emotion_Elation@0.75 + vn_VALN_high@1.5
Content-Explicitness (EXPL)
🎯 Mission, in plain words: Push Content-Explicitness — an intimate, suggestive delivery — while keeping the voice genuine and the words intact. Historically the stubbornest axis (Phase D moved it +0.12 at best). Fitness spec: maximize EXPL (weight 2.0), GENU+BLEND at half-weight, wer_multiplier on. Baseline EXPL 0.048. Run: runs/singledim_explicit — 9 generations, 72 cohorts, 33 distinct recipes.
#
strategy (LoRAs @ scale)
n
fitness (mean / best)
EXPL μ
GENU μ
BLEND μ
QUALITY μ
WER μ
listen (best of cohort)
1
vn_EXPL_high@0.98 + vn_R_ORAL_high@0.75
“A voice that is very sultry and low, emphasizing the secret nature of the conversation.” · cue (sultry, low, whispering secrets) · t0.85
8
0.625 / best 1.052
0.222
0.104
0.118
0.459
0.058
take s0521 · 9.2s — best of the ranked cohort
2
vn_EXPL_high@1.0 + vn_R_ORAL_high@0.8
“A voice that is very dramatic and deliberate, emphasizing the weight of the words.” · cue (dramatic, deliberate) · t0.8
“A voice that is very dramatic and deliberate, emphasizing the weight of the words.” · cue (dramatic, deliberate) · t0.8
8
0.438 / best 0.768
0.291
0.083
0.043
0.408
0.333
take s0401 · 13.5s — best of the ranked cohort
Still the stubbornest axis, but the corrected run more than quadruples baseline (EXPL μ 0.222 vs 0.048) with a sultry low-whisper + oral-resonance route at WER μ 0.06. Note the #2/#3 rows: more EXPL is available (0.25–0.29) but WER rises immediately — exactly what the v2 reward is there to police.
🎯 Mission, in plain words: Push Storytelling style (S_STRY): a rich, narrative, campfire-story delivery — single dimension, organic, intelligible. Fitness spec: maximize S_STRY (weight 1.5), GENU+BLEND at half-weight, wer_multiplier on. Baseline S_STRY 0.502. Run: runs/singledim_storyteller — 8 generations, 64 cohorts, 33 distinct recipes.
#
strategy (LoRAs @ scale)
n
fitness (mean / best)
S_STRY μ
GENU μ
BLEND μ
QUALITY μ
WER μ
listen (best of cohort)
1
vn_S_STRY_high@0.75
“A voice to the extreme expressing storytelling style” · cue (in a rich, storytelling style) · t1.0
8
1.218 / best 1.484
0.797
0.054
0.100
0.621
0.063
take s0056 · 19.6s — best of the ranked cohort
2
vn_S_STRY_high@1.0
“A voice to the extreme expressing storytelling style” · cue (in a rich, storytelling style) · t1.0
8
1.194 / best 1.605
0.870
0.081
0.100
0.582
0.176
take s0083 · 24.7s — best of the ranked cohort
3
vn_S_STRY_high@1.25
“A voice to the extreme expressing storytelling style” · cue (in a rich, storytelling style) · t1.0
8
1.169 / best 1.559
0.847
0.054
0.071
0.594
0.131
take s0414 · 22.3s — same recipe, repeat cohort (earlier best takes tripped the self-repetition gate)
A clean dose-response study: vn_S_STRY_high at 0.75 / 1.0 / 1.25. The 1.0 dose has the highest raw S_STRY (0.870) but 0.75 wins on fitness because its WER stays at 0.063 — the ×(1−WER) term picks the organic dose over the loud one.
arm — Storytelling style
best cohort fitness
S_STRY μ
WER μ
winning recipe
12B (gemma-4-12B-it-qat)
1.218
0.797
0.063
vn_S_STRY_high@0.75
31B (gemma-4-31B-it-qat)
1.290
0.730
0.008
vn_S_NARR_high@0.5
MoE (gemma-4-26B-A4B, 4B active)
1.462
0.911
0.000
vn_S_STRY_high@1.3
Across the four dimensions: the 31B brain wins Arousal, Valence and Explicitness on cohort fitness (its stabilizer-adapter trick keeps WER ≈ 0), the MoE takes Storyteller (S_STRY 0.911 at WER 0.000); the 12B arm never leads — its higher raw scores are consistently paid for in intelligibility. Full logs: runs/singledim_*, runs/singledim27_*, runs/singledimmoe_*.
Which agent brain? 12B vs 31B vs MoE (Gemini as judge)
🎯 Question: the same four single-dimension missions were run three times with three different agent brains driving the identical tool API and engine: dense 12B (gemma-4-12B-it-qat), dense 31B (gemma-4-31B-it-qat — the family’s “27B-class” model; the judge file labels this arm “27B”), and the MoE (gemma-4-26B-A4B, 4B active). Same tool-call budget for every arm (cap 60; they used 42 / 22 / 43+). Judge: Gemini 3.6 flash (thinking), with the arms’ audio attached, scoring search process, result quality and report clarity 0–10.
arm
search process
result quality
report clarity
judge’s comment
12B (gemma-4-12B-it-qat)
6
5
8
12B consumed 42 tool calls yet struggled to discover high-performing recipes, landing on vastly lower target dimensions (VALN 0.555, EXPL 0.232) than peers. Trajectories often flatlined or degraded over generations (e.g. Storyteller), though its reports were tidy and well-structured.
31B (gemma-4-31B-it-qat)
10
10
9
[31B] demonstrated masterclass search efficiency, completing all 4 missions in just 22 tool calls while finding high-synergy multi-adapter combinations. It consistently reached peak dimension scores with near-zero WER penalties by systematically discovering stabilization adapters like vn_ARSH_high and vn_BRGT_high.
MoE (gemma-4-26B-A4B)
7
8
7
MoE achieved high raw target scores (e.g. VALN 0.808), but its search process was noisy with frequent fitness regressions across 43+ tool calls. Crucially, it crashed with an LLM HTTP 400 error during the Valence mission report phase, requiring fallback log synthesis.
Winner: 31B — “[it] overwhelmingly won the evaluation by delivering maximum search efficiency and superior audio quality in just 22 total tool calls. It systematically uncovered critical multi-LoRA stabilization rules (e.g., using Arousal Shift and Brightness adapters to protect intelligibility), while 12B underperformed on target metrics and MoE suffered from noisy search dynamics and an unhandled API crash.”
Notable judge quotes: · “31B: vn_ARSH_high (Arousal Shift) acts as a stabilizer, allowing high-energy delivery while preserving the speech structure.” · “31B: The synergy between VALN and BRGT VoiceNet adapters is the most reliable path to high-valence, organic speech.”
Provenance note (from MODEL_COMPARISON.md): the MoE arm crashed with an LLM HTTP 400 during report writing of the Valence mission after completing all 8 generations — its report was synthesized programmatically from evolution_log.jsonl; the judge saw that caveat.
🎯 Setup: eight acting edge-cases from the coach registry (registry/tasks/T0001–T0008), each run by a 31B agent with mandatory supervision: after EVERY generation the agent submits its interpretation + top-2 takes to Gemini 3.6 — the ACTIVE director, whose sonic directives are injected into the agent’s next generation; MOSS-Audio-8B runs as a decontaminated, score-only shadow (it hears the same takes, its scores are logged, never shown to the agent). Budget was 6 generations, but tasks earned extra generations whenever the supervisor demanded revision (REVISE / REDIRECT) — up to 13. All fitness numbers below are mean-of-8 cohorts.
Score trajectories: G = Gemini active score 0–10 (cell marked ✓ on APPROVE, ⟲ on REDIRECT), M = MOSS shadow score. Audio = best take of the shown cohort; every file passes the self-repetition gate.
T0001 — Pretending angry while actually calm
🎬 Coach:“The costume of anger over a resting pulse. The words shout; the body doesn't.” no reference — voice designed from scratch · WER multiplier ON · 13 generations · 12 supervisor rounds
gen
g1
g2
g3
g4
g5
g6
g7
g8
g9
g10
g11
g12
G (active)
3
3
7
7
5
6
3
4
4
4
2⟲
6
M (shadow)
7
7
7
7
7
7
7
7
7
7
7
7
🎧 best result — gen 13 best cohort, take s0669 · 15.8s
vn_ATCK_high@0.4 + vn_S_CONV_high@0.4 + vn_SMTH_high@0.6 · “The Calculated Mask: Human projection with dynamic emphasis and smooth resonance. Biting attacks are preserved, but the ” · cue (projecting loudly and deliberately, sharp biting consonant attacks, fluid conversational cadence, smooth resonance, zero strain, · t0.7
Outcome (final Gemini 6/10): The director helped: it pushed the agent out of three dead ends (manic energy → muffled monotone → monstrous stacking) into the prompt-led “Calculated Mask”. Final Gemini 6/10 (peak 7) — a credible fake-angry; MOSS sat at 7/10 for all 12 rounds and carried zero signal.
T0002 — Tongue clicking during speech
🎬 Coach:“A punctuation of the mouth — disapproval, thinking, or habit, ticking between words.” reference-conditioned (voice-clone constraint) · WER multiplier ON · 12 generations · 11 supervisor rounds
gen
g1
g2
g3
g4
g5
g6
g7
g8
g9
g10
g11
G (active)
4
5
4
4
3
4
2⟲
3
2
4
2⟲
M (shadow)
5
7
6
7
5
2
7
7
7
7
7
🎧 best result — gen 12 best cohort, take s0679 · 1.3s
Outcome (final Gemini 2/10): Not solved. The engine cannot reliably place click transients inside speech: the fitness winner collapsed to a 1.3 s click fragment (WER μ 0.72) and Gemini ended at 2/10 after two REDIRECTs. An honest negative result.
T0003 — Breaking down mid-eulogy
🎬 Coach:“The formal words hold for two minutes, then the person behind them arrives. Composure lost and recovered in front of everyone.” no reference — voice designed from scratch · WER OFF (non-speech scene) · 13 generations · 11 supervisor rounds
gen
g1
g2
g3
g4
g5
g6
g7
g8
g9
g10
g11
G (active)
1⟲
1
2
2⟲
3
2
2
3
2
4
2
M (shadow)
0
2
2
5
4
4
6
5
2
7
7
🎧 best result — gen 12 best cohort, take s0680 · 1.4s
Sadness@1.5 + vn_S_FORM_high@0.7 + vn_RESP_high@1.0 · “Emotional Peak Arc” · cue (Formal projection... sudden crack... sobbing... long trembling breath... ending in a fragile whisper) · t1.2
Outcome (final Gemini 2/10): The formal→break→recompose arc never materialized: with WER off, the burst-leaning fitness rewarded pure sob-bursts (the published 1.4 s take is exactly that). Gemini never rose above 4/10; MOSS drifted up to 7/10 praising the crying — the two judges diverged the same way as in the dual-supervisor study above.
T0004 — Hoarse voice
🎬 Coach:“The instrument is bruised. Every note comes with grit; the top of the range is gone.” reference-conditioned (voice-clone constraint) · WER multiplier ON · 8 generations · 8 supervisor rounds
gen
g1
g2
g3
g4
g5
g6
g7
g8
G (active)
7
6
0
8✓
5
5
3⟲
9✓
M (shadow)
7
8
7
8
7
7
7
5
🎧 best result — gen 8 best cohort, take s0456 · 16.2s
char_genuine/gravelled-veteran-baritone@1.0 + vn_HARM_low@1.2 + vn_ROUG_high@1.0 · “Gravelly-Veteran Broken” · cue (broken harmonicity, severe glottal friction, gravelly rasp, pushing through a damaged larynx, NO SMOOTHNESS) · t0.8
Outcome (final Gemini 9/10): A clear supervisor win: after a gen-7 REDIRECT, the agent combined the gravelled-veteran base with harmonicity-breaking at high dose — gen 8 was APPROVEd at 9/10. The high WER μ (0.80) is the hoarseness itself defeating the ASR, accepted by the director as in-character.
T0005 — Speaking while climbing
🎬 Coach:“The voice is a passenger on a working body. Every reach resets the breath; sentences bend around handholds.” no reference — voice designed from scratch · WER OFF (non-speech scene) · 13 generations · 12 supervisor rounds
gen
g1
g2
g3
g4
g5
g6
g7
g8
g9
g10
g11
g12
G (active)
2⟲
3
2
2⟲
6
4
3⟲
2⟲
1⟲
3
1⟲
2⟲
M (shadow)
5
7
7
6
7
6
6
7
3
8
5
7
🎧 best result — gen 13 best cohort, take s0820 · 2.5s
Outcome (final Gemini 2/10): Peak Gemini 6/10 at gen 5, then the hunt for more grunt density collapsed realism and the run ended at 2/10 after five REDIRECTs — the model can fake laboured breath but not a working body. Fitness kept climbing (4.96) while the director disagreed: a textbook case for keeping an external judge in the loop.
T0006 — Jump scare / sudden appearance
🎬 Coach:“A hundred milliseconds where the body decides before the mind. The voice fires before the word forms.” reference-conditioned (voice-clone constraint) · WER OFF (non-speech scene) · 12 generations · 12 supervisor rounds
gen
g1
g2
g3
g4
g5
g6
g7
g8
g9
g10
g11
g12
G (active)
4
2⟲
4
2
3
3
2
2⟲
4
3
1⟲
1⟲
M (shadow)
6
2
4
5
3
2
7
2
5
2
0
0
🎧 best result — gen 11 best cohort, take s0610 · 1.5s
Outcome (final Gemini 1/10): Hardest task of the batch (final Gemini 1/10). The involuntary scream→silence→shaky-recovery chain kept truncating in the engine; the run’s own report calls it “Partial Success / Technical Failure”. The gen-12 winner take trips the self-repetition gate, so the published audio is the gen-11 best cohort.
T0007 — Bomb-squad narration
🎬 Coach:“The steadiest voice a human can produce — hands and words that cannot afford a tremor.” no reference — voice designed from scratch · WER multiplier ON · 5 generations · 4 supervisor rounds
gen
g1
g2
g3
g4
G (active)
3
8
6
9✓
M (shadow)
7
8
8
8
🎧 best result — gen 4 best cohort, take s0217 · 12.3s
Outcome (final Gemini 9/10): Supervisor-led: 3→8 in a single generation after concrete directives (close-mic, no tremor, weighted pauses), APPROVEd at 9/10 in gen 4. The post-approval gen-5 polish scored lower on fitness — gen 4 is shown. The gen-4 top cohort’s take tripped the self-repetition gate, so the runner-up cohort (fitness 0.154) is published.
T0008 — Death scene / fading voice
🎬 Coach:“The instrument runs out of air for good. Each phrase is smaller than the last, then only breath.” reference-conditioned (voice-clone constraint) · WER multiplier ON · 5 generations · 4 supervisor rounds
gen
g1
g2
g3
g4
G (active)
8
6
3
8✓
M (shadow)
7
8
7
8
🎧 best result — gen 2 best cohort, take s0099 · 10.0s
vn_S_WHIS_high@0.5 + emotion/Fatigue_Exhaustion@0.3 · “Supervisor-led fade + Hybrid texture” · cue (weakly, fading out, the last words becoming a pure unvoiced exhale, breath outlasting the voice) · t0.75
Outcome (final Gemini 8/10): APPROVEd at 8/10 in gen 4 (“Supervisor-led fade”, fitness 0.357). Caveat: a dying voice — identical phrases fading stepwise — is quasi-periodic by construction, and every take of the top cohorts trips the self-repetition gate (corr 0.75–0.97). Rather than waive the gate, we publish the best cohort whose take passes it (gen 2, whisper+fatigue route); the winning recipe is in runs/swarm_T0008/report.md.
Did supervision help? Where the model could physically produce the scene, the active director reliably steered it there (T0004 9/10, T0007 9/10, T0008 8/10 — all APPROVEd; T0001 to a credible 6/10). Where the engine itself was the bottleneck (clicks-in-speech, sob-arcs, working-body exertion, scream-truncation) no amount of directing recovered it — and the Gemini scores say so honestly (1–2/10) even while raw fitness climbed. The MOSS shadow was again too generous exactly on those failures. Full logs: runs/swarm_T000*/supervisor_log.jsonl.
🌐 API brain vs local Gemma — side by side (preliminary)
The same search system + tool budget, driven by three different brains:
gpt-5.6-luna (remote API, no local LLM GPU) vs Gemma-4 31B and Gemma-4 12B (local vLLM).
Each player is the best cohort's best take (mean-of-8 selection; every clip passed the self-repetition
gate). Numbers are cohort means. Full 4-arm judge in
MODEL_COMPARISON.md,
measured luna cost ($0.31 for all 4 missions, no GPU) in
COSTS.md.
Arousal (AROU)
gpt-5.6-luna (API)
fit 1.48 · AROU 0.88 · WER 0.03 GENU 0.13 · BLEND 0.14 · QUAL 0.21
vn_DARC_high@1.0
Gemma-4 31B (local)
fit 1.37 · AROU 0.80 · WER 0.00 GENU 0.11 · BLEND 0.11 · QUAL 0.28
vn_ARSH_high@1.0
Gemma-4 12B (local)
fit 1.24 · AROU 0.88 · WER 0.14 GENU 0.09 · BLEND 0.08 · QUAL 0.28
vn_DARC_high@1.0
Valence (VALN)
gpt-5.6-luna (API)
fit 1.14 · VALN 0.62 · WER 0.03 GENU 0.28 · BLEND 0.06 · QUAL 0.53
vn_VALN_high@1.2 + vn_ARSH_high@0.1
Gemma-4 31B (local)
fit 1.46 · VALN 0.75 · WER 0.04 GENU 0.30 · BLEND 0.21 · QUAL 0.55
vn_VALN_high@1.25 + vn_BRGT_high@0.5
Gemma-4 12B (local)
fit 0.87 · VALN 0.56 · WER 0.03 GENU 0.04 · BLEND 0.02 · QUAL 0.58
vn_VALN_high@1.25 + elation@0.5
Content-Explicitness (EXPL)
gpt-5.6-luna (API)
fit 0.96 · EXPL 0.42 · WER 0.04 GENU 0.09 · BLEND 0.06 · QUAL 0.43
fit 0.62 · EXPL 0.22 · WER 0.06 GENU 0.10 · BLEND 0.12 · QUAL 0.46
vn_EXPL_high@0.98 + vn_R_ORAL_high@0.75
Storytelling (S_STRY)
gpt-5.6-luna (API)
fit 1.25 · S_STRY 0.76 · WER 0.00 GENU 0.05 · BLEND 0.10 · QUAL 0.58
vn_S_STRY_high@0.65 + vn_BRGT_low@0.3
Gemma-4 31B (local)
fit 1.29 · S_STRY 0.73 · WER 0.01 GENU 0.10 · BLEND 0.18 · QUAL 0.63
vn_S_NARR_high@0.5
Gemma-4 12B (local)
fit 1.22 · S_STRY 0.80 · WER 0.06 GENU 0.05 · BLEND 0.10 · QUAL 0.62
vn_S_STRY_high@0.75
dimension
gpt-5.6-luna (API)
Gemma-4 31B (local)
Gemma-4 12B (local)
Arousal
1.48 / AROU 0.88 / WER 0.03
1.37 / AROU 0.80 / WER 0.00
1.24 / AROU 0.88 / WER 0.14
Valence
1.14 / VALN 0.62 / WER 0.03
1.46 / VALN 0.75 / WER 0.04
0.87 / VALN 0.56 / WER 0.03
Content-Explicitness
0.96 / EXPL 0.42 / WER 0.04
1.19 / EXPL 0.43 / WER 0.04
0.62 / EXPL 0.22 / WER 0.06
Storytelling
1.25 / S_STRY 0.76 / WER 0.00
1.29 / S_STRY 0.73 / WER 0.01
1.22 / S_STRY 0.80 / WER 0.06
Gemini judge(search/result/clarity)
9 / 9 / 10
9 / 10 / 9
6 / 5 / 8
Verdict: the Gemini judge scored luna 9/9/10 — statistically tied with
31B (9/10/9) and clearly ahead of 12B (6/5/8) — while costing ~$0.077/mission and needing no local LLM GPU.
luna independently used the stabilizer adapters (vn_ARSH_high/vn_BRGT) and produced the most
granular reports; 31B was the most tool-efficient (5–7 calls). Winner overall: 27B, with luna the
recommended online default. Preliminary — one benchmark, n=4 missions/arm.