Chapter 8 — Performance
Status: DRAFT 1 (2026-09-21), DESIGN v6 §4. New in v6, on the director review's first finding. Sources: production/phase3/DIRECTOR-REJECTIONS.md; ledger rows by line. Note for the reader: our shot cards carried a DIRECTION block with goal, obstacle, task and acting for every card (HALFWAY-SHOTLIST.md, e.g. lines 238–250). What we did not have was any acceptance step that listened to a recorded line against that block; the voice audition tested identity and the transcript tested words. This chapter adds the per-line acceptance we should have run, built from the moment we found out.
8.0 Objective
Given a card, you can write its intention line (want, obstacle), name the listening beat and the reaction, and judge a delivered line for intention by ear. Given two takes of a face, you can say which one has the beat in the eyes.
8.1 What you are deciding
Chapter 7 chose a voice. This chapter decides what each line is doing, so that you can hear when a delivery is not doing it. Three things per line:
- The want. What this character wants from the other one, right now, in this line. A thing, not a meaning.
- The obstacle. What is in the way, including what it costs to ask.
- The listening beat. What the other face does while the line is said. In a generated film this is the shot you can most reliably get, and the one nobody plans.
And one thing per scene: the reaction take, a silent single of the face at the turn, judged for whether the beat lands in the eyes.
8.2 How we did it
We did not, and then we paid for it. After the third assembly the Director wrote the review that changed the rest of the production (DIRECTOR-REJECTIONS.md:6):
"Dialogue overall sounds like book reading: flat, without scene-appropriate energy, intention or emotional response. Approval of Naomi/Holden as voice identities is not approval of each delivery. Word accuracy and timing do not establish acting quality."
Every line had passed its transcript. Every voice had been approved by ear on one audition line. None of that was performance. The recovery gate the Director set (DIRECTOR-REJECTIONS.md:14): "review actual line deliveries by ear for intention, naturalness and emotional timing"; "test the hardest short performance before extending or assembling"; casting, performance, posture, continuity and sync "must each pass independently."
What we then did, line by line, was put the intention into the tone cue and re-record:
| Line | First cue and result | Director's note | Second cue and result |
|---|---|---|---|
| 2A 是巧克力 | flat take; measured "a −44 dB hole at 1.40 s between syllables and hard digital-silence padding" (row 561) | "2A 是巧克力 audio is choppy" (row 560) | [casual] (row 564, job 61f4e035): 1.68 s, speech 0.20–0.85 s, transcript exact; approved (row 569) |
| 2B 你是一朵鲜花… | approved as voice, rejected as delivery | "should be witty/teasing" (row 560) | [playfully teasing] (row 565, job e6f03a9d): three phrases 0.23–1.34 / 1.70–2.47 / 2.63–4.16 s; approved (row 569) |
| 4C opening | "A better but opening still announcer-like" (row 467) | "melodic mischievous opening, not flat" (row 468) | [playfully teasing] opening then [casual] question |
| 1D | v1/v2 "delivery wrong, not like daily talk" (row 296) | "must sound like a normal person speaking" | reworded cue |
The benchmark the Director gave for the whole cast (row 446): "例如你吗 非常好; 我耳环掉了 and 走吧 也不错; use these as performance benchmark." Three lines that had a want in them: a tease turned back, a small confession, a decision.
Each of these retakes cost 0.15 credits. The assemblies they replaced had cost the video takes under them, generated with the flat audio attached as reference, and one of them (2B) had been "lip-driven by the v1 recording", so the fence warned "a new rhythm may drift from the mouth" (row 562); the Director approved the voices on playback (row 569) without a separate sync ruling, so whether it drifted is not established. Performance before picture would have cost nothing.
8.3 Case 8-A — approved as a voice, rejected as a delivery
- What I saw: twenty lines transcribed word for word, two voices approved by ear on one audition line each, three assemblies built on them, and then the Director's review: "book reading". The specific lines above and their retakes.
- Why I think so: the audition tested identity (does this sound like her) with one line. Nothing tested intention (does this line want something) for the other nineteen. A transcript checks words; it has no column for want.
- What I changed: an intention line per card (below), written before the recording, and acceptance by ear against it, separately from the voice approval and the sync check. In v6 this is a rule (DESIGN §6, "Performance acceptance").
- Would this happen on your tool? Yes. Every TTS model reads. Whether it acts is decided by what you ask it for, line by line, and by whether anyone listens for it.
8.4 Do it with your tools
| Need | We used | Any tool | If yours lacks it |
|---|---|---|---|
| A cue the model honours | bracketed tags before the text ([casual], [playfully teasing]) |
style, emotion or "instruction" fields; rewording; punctuation | split the line in two files with different cues and cut them |
| A listening ear | the Director, then the second reader | same | play the line to someone who has not read the script and ask what the speaker wants |
| A reaction take | silent singles of the listening face (we planned them as "listening" cards, e.g. 4B) | any image-to-video of a still face with one written affect | a held still with a single eye movement is enough for the animatic |
8.5 Your turn
Time cap: one evening. Spend: from your voice row; a retake is one line's price.
- Write an intention line for every line in the script (template: Appendix C.13): want, obstacle, listening beat, "accept by ear when". Do it before you look at the Chapter 7 recordings again.
- Listen to every recording against its intention line, sound only, with the second reader. Mark each: has the want / does not. Keep a director's-notes line per marked line (C.8).
- Re-record the marked ones with the cue changed to name the want, one variable at a time. Fence, price, one click, ledger. Stop at two takes per line; a line that will not come is rewritten (rung 5 belongs to you).
- Reaction takes on paper. For each scene's turn, write the reaction card: whose face, what it does, in one sentence with no adjective ("her shoulders drop, then the smile"). These are generated in Chapter 9's stress test.
- Blind pairs. For three of the lines you re-recorded in step 3, keep the first take and the retake as a pair. Write down, before anyone listens, which file of each pair has the want. Play each pair to the second reader in an order they do not know, sound only, and ask one question: "which one wants something, and what?" Record their pick and their answer beside yours (C.13, last two columns).
- Two-face drill (ungated, no spend): any two short face clips, from Chapter 9's stress strips when you have them or two phone clips of a friend reacting to two different sentences. Watch silent. Write which one has the beat in the eyes and what the eyes did.
8.6 Gate
Each line needs valid evidence per DESIGN §5; outcomes are pass / fail / unverified, and unverified blocks.
- Every line has an intention line with a want that is a thing, not a meaning (evidence: the cards; the judgment "a thing" is yours, by name).
- Every recording was listened to against its intention line by you and the second reader, sound only (evidence: the marks; self-attested listening; ★ second reader initials).
- Every marked line was re-recorded or rewritten, with its ledger rows (evidence: ledger).
- Performance task: the three blind pairs exist as six files with your pick written before the listening, and the second reader's blind pick matched yours on at least two of the three pairs, with their answer to "what does it want" naming the want on your intention line for each match (evidence: the six files, your dated sheet, ★ the reader's picks and answers in their hand; a pair where the reader could not tell is a miss).
- Reaction cards exist for every scene's turn (evidence: the cards).
Pass example from HALFWAY: 2B, [playfully teasing], three phrases measured, approved by the Director by ear (row 569). Fail example: the first twenty lines, word-perfect and flat, three assemblies deep before anyone listened for a want.