← runs

# Best candidates tracker

Live record of the highest-scoring concepts and full episodes the system has produced. Sorted by `combined = 0.6·text + 0.4·audio` when audio is measured, else by text alone. Updated on every milestone.

---

## Current peaks (2026-05-01)

### Full episodes with audio (combined ≥ 8.5)

| Rank | Combined | Text | Audio | Concept | Run path |
|---|---|---|---|---|---|
| 1 | **9.08** | 9.39 | 8.60 | "Drop Match" | `2026-04-28/2026-04-28T17-09-33__90b81484/01` |
| 2 | **8.81** | 8.34 | **9.50** | "Night Board" (11.5 min) † | `2026-05-01/2026-05-01T11-31-13__db63046c/01` |
| 3 | 8.74 | 8.90 | 8.50 | "The Drop Match" (regen) | `2026-04-30/2026-04-30T11-00-45__fb49479e/01` |
| 4 | 8.74 | 9.17 | 8.10 | "Equal Temperament" | `2026-04-29/2026-04-29T17-07-57__c6092e4f/02` |
| 5 | 8.71 | 9.14 | 8.05 | "Holding Pattern" | `2026-04-28/2026-04-28T16-00-39__fc988792/01` |
| 6 | 8.64 | 9.00 | 8.10 | "Measure Twice" | `2026-04-28/732670d724fa/02` |
| 7 | 8.63 | 9.48 | 7.35 | "Plumb Line" | `2026-04-29/2026-04-29T22-56-55__b3240200/01` |
| 8 | 8.59 | 9.26 | 7.60 | "Paste the Wall" | `2026-04-29/2026-04-29T19-35-37__5d8baa65/01` |
| 9 | 8.51 | 9.42 | 7.15 | "The Zugzwang Protocol" | `2026-04-29/2026-04-29T15-56-53__829d1105/02` |
| 10 | 8.51 | 8.98 | 7.80 | "Green Ink" (vicar/bell-ringer, spec-decode) | `2026-05-01/2026-05-01T19-28-49__ad61f6f6/01` |

### Short-form runs (`--target_duration_min=8`)

| Formula | Text | Audio | vc.score | Gate (post-penalty) | Concept | Run path |
|---|---|---|---|---|---|---|
| **8.47** | 8.61 | 8.25 | **1.00** | **8.47** | "Primary Contacts" (v6d episode, render-pick n=5 attempt 2 — first clean vc) ★★★ | `2026-05-03/render_pick__63e273d0/01` |
| 8.16 | 8.73 | 7.30 | 0.92 | 7.54 | "The Phantom Exchange" (v9, telephone operators) | `2026-05-03/2026-05-03T23-09-41__6c6def04/01` |
| 8.67 | 8.61 | 8.75 | 0.72 | 6.24 | "Primary Contacts" (v6d episode, render-pick n=3 attempt 3) | `2026-05-03/render_pick__9ca8e087/attempt_03` |
| 8.97 | 8.61 | 9.50 | 0.15 | 0.56 | "Primary Contacts" (v6d original — vc-fail, 49 Hz drift) ‡ | `2026-05-03/2026-05-03T10-07-57__9aa1a759/01` |
| 8.66 | 8.70 | 8.60 | ? | 8.66† | "The Final Arrangement" (undertakers, v4) | `2026-05-02/2026-05-02T12-15-35__38c4bd49/01` |
| 8.32 | 8.60 | 7.90 | ? | 8.32† | "The Primary Tribute" (undertakers, v2) | `2026-05-02/2026-05-02T11-19-53__f2e392a0/01` |
| 7.76 | 8.33 | 6.90 | ? | 7.76† | "The Platinum Package" (undertakers, v1 baseline) | `2026-05-01/2026-05-01T19-47-32__f401f9ae/01` |

★★★ **First clean voice-consistency render (gate 8.47).** Same v6d
episode JSON, fifth re-render in render-pick n=5. **vc.score 1.0** —
no voice drift, vc_passed=True, **no penalty applied**. Audio judge
axes uniformly strong (8/8/8/8/9). WPM 159 — first render under the
160 wpm threshold. WER is 0.14 (higher than other attempts at 0.04;
some text mangled at the STT round-trip layer; the audio judge confirms
listening quality is fine). Validates the render-pick approach: across
8 total re-renders of the same episode JSON (3+5), vc.score landed
0.4–1.0; persistence pays off. Audio at
`data/generations/2026-05-03/render_pick__63e273d0/01/audio/final.mp3`.

★★ **Earlier best gate (7.54).** v9 — fresh meta run on the
telephone operators seed with `--audio_judge_n_repeats=2` and the
recalibrated pacing preamble. **30/31 rules pass (best yet).** vc.score
0.92 (vs v6d 0.15, render-pick 0.72). One vc violation: Mrs. Mavis
Corby drifted only 28.8 Hz — just 4 Hz over the 25 Hz threshold —
making the multiplicative penalty very mild (0.92×). WPM 164 is still
above the 145 target, but down from 179 in earlier renders. Audio at
`data/generations/2026-05-03/2026-05-03T23-09-41__6c6def04/01/audio/final.mp3`.

★ **Earlier best (6.24).** Same episode JSON as v6d, re-rendered.
Across 3 re-renders the same script produced vc.score 0.50/0.50/0.72;
attempt 3 was the cleanest. Audio judge 8.75 (single call). Pre-penalty
formula 8.67. The post-penalty gate (×0.72 voice consistency multiplier)
is 6.24 — **roughly 11× v6d's 0.56**. v6d's recorded "8.97" was the
formula score; the actual gate after penalties was 0.56 because Chloe
Sparks drifted 49 Hz mid-episode. Audio at
`data/generations/2026-05-03/render_pick__9ca8e087/attempt_03/render_03.mp3`.

† **Older entries lack a recorded vc.score** — the voice_consistency layer
landed after these runs. Their "combined" was the formula value with no
multiplicative penalties; not directly comparable to the post-penalty gate.

‡ **v6d's headline 8.97 was misleading.** The number was the pre-penalty
formula `0.6×8.61 + 0.4×9.50`. After voice_consistency ×0.146 + (now-removed)
diarization penalty ×0.425, the manifest gate score was 0.556. Documented in
`docs/eval_battery_findings.md` section 5b. Diarization penalty has since
been removed (Exp 10); only voice_consistency multiplicatively penalises now.

‡ **New peak (combined 8.97).** Same seed as v4. Audio +0.90 over v4. Diagnostic
v6d intentionally ran with `--audio_adversarial` OFF; a parallel v6c run with the
flag ON regressed audio. Conclusion: the audio adversarial critic was patching
delivery_notes that introduced same-speaker dead air. Run with audio_adversarial
disabled by default. WAV at `data/generations/2026-05-03/2026-05-03T10-07-57__9aa1a759/01/audio/final.mp3`.

**v4 vs baseline:** +0.90 combined / +1.70 audio across 4 iterations on the same
seed. Path: 6 fixes (e8aabab) → A/B confirmed → revert TTS pacing cue +
strengthen delivery_note cap (2bb589c) → v4 lands.

Notable v4 wins over v2:
- character_distinctness 8 → **10** (max)
- WPM 151 → 135 (corpus pace finally)
- delivery_note coverage 100% → 35% (cap finally took)
- avg words/line 13.8 → 12.2 (target ≤12)
- genre_register_collision (text axis) 8 → 10

A/B against same seed, **+0.56 combined / +1.00 audio** after the
6 fixes from `e8aabab`:
- consistency_across_episode 7.0 → 9.0 (voice anchor working)
- STT WER 0.089 → 0.020 (clearer enunciation)
- Same-speaker gap 0.00s → 0.40s (breath beat landing)
- delivery_note pacing 7.0 → 8.0; prosody 6.0 → 7.0
- SFX restraint 15 → 8 events; music 0 → 3 (into target band)

See `2026-05-02/.../ab_comparison.md` for the full diff.

† **Highest audio judge to date (9.50).** First episode under the new TTS prompting stack (per-speaker style cue + conversation framing + sticky voice routing). Audio committed at `data/generations/2026-05-01/2026-05-01T11-31-13__db63046c/01/audio/final_mixed__2026-05-_iter01_20260501T114131.mp3`.

### Text-only peaks (audio not yet rendered)

| Rank | Text | Concept | Run path |
|---|---|---|---|
| 1 | **9.74** | "Air in the System" | `2026-04-29/2026-04-29T18-56-52__9b344b58/02` |
| 2 | 9.49 | "Vulnerable" | `2026-04-28/2026-04-28T20-15-22__633053a9/02` |
| 3 | 9.44 | "The French Drop" | `2026-04-28/2026-04-28T19-30-42__64bb978d/01` |
| 4 | 9.30 | "Drop Match" (concept-only) | `2026-04-29/2026-04-29T22-35-32__a788835d/01` |
| 5 | 9.21 | "The Drop Match" iter 2 | `2026-04-30/2026-04-30T11-00-45__fb49479e/02` |

## Reference points

- **Corpus mean composite:** ~8.45 (computed across the 12 corpus episodes via 8-axis judge + 25 lint rules).
- **Corpus best episode:** "The Wroxton Box" at ~9.34.
- **Calibration separation (corpus vs. contrast set):** 4.65+ points.

The current text peak (9.74) beats the best human-corpus episode by +0.40, and the current full-audio peak (8.60 audio) is mid-corpus. Audio is the bottleneck — text saturated.

## Audio bottleneck — open questions (from expert listener feedback, 2026-04-30)

1. **No emotional arc within a character** — a character on edge stays on edge for 30 minutes. Need rising/falling intensity within scenes.
2. **Voice mid-episode flip** — audio listener flagged a voice change at ~6:30. Voice consistency check needs to be a hard test, not a warning.
3. **Trailing dead air** — generated 30 min, audio stops at ~16. Need `trailing_silence_seconds` deterministic check.
4. **One-note humor** — same gag template repeated 7×. Need joke-diversity check.
5. **Plot empty except final reveal** — needs multiple perception shifts à la *English for Pony Lovers*.

See `docs/nice-to-haves.md` for the operationalization plan.