Repository documentation

Directed onboarding cut QA

Status: directed pilot visual inspection passed by the user; see the separate visual decision. The user validated the voices and rejected the first cut's animation. The original review decision is preserved; the new MP4 has a different identity in machine-readable observations. No creative approval or publication is inferred from technical verification.

New cut

• The 38.437989-second H.264/AAC film has eight captions, visible captions in the picture, retained SRT/WebVTT and an offline browser companion. Independent litai video verify --manifest accepts the canonical output.
• All eight approved WAVs and their original generation records remain unchanged. Sidecar caption bytes also match the previous audition. There was no new speech synthesis, cast selection or voice substitution.
• The renderer realizes the performance score through 46 finite beats and 23 recording-bound phrase cues. Two-joint working arms, a source folder, gaze, brow, expression, listener reactions, deliberate holds and pending-gate emphasis replace the old arm/nod cycle. The composition remains a stable two-shot.
• Local Whisper base.en provided estimated word times. The cue map retains raw recognition and curated word spans, model and recording identities. Recognition errors are not promoted into dialogue. These are not verified phoneme timings.
• Nine Python guards and lint/format pass. Ten score-admission refusals cover changed recordings, unknown targets/poses, bad timing, omitted turns and off-story cues. Ten real builder/capture refusals cover drift in scene, combined/individual audio, story, dependency lock, narrative, score, cues, compiled score and HTML.
• Nine browser checks pass. Five requested times repeat in both pixels and DOM after different seek histories. Captions fit; keyboard controls, offline local-file playback, reduced motion on load/change, static rigs and no-JavaScript transcript are checked. The browser performs no external requests.
• Directed checks observe an intentional hand hold, a hand stopping short of Apply, unchanged folder contents and spoken-order emphasis on three still-pending gates. Every adjacent hand position across the full 24 fps timeline is checked; the maximum step is 29.190382 pixels against the retained 35-pixel limit. This bounds snapping, not expressiveness or acting quality.
• Decoded midpoint frames from all eight turns and their contact sheet were inspected. No clipped captions or unintended overlap were observed. An early hand/panel layering issue and an over-fast reach were corrected before this cut. Sampled frames do not constitute watching the complete moving performance.

User review and remaining production

The user has passed visual inspection of this exact directed cut. That approves this pilot's treatment, not new dialogue, a full episode or publication. The next creative artifact is the full annotated narrative. For the future full episode, check whether each action communicates a thought or learning step; whether eyelines, listening, anticipation, holds and settles feel coherent; and whether captions, mouth cues and spoken emphasis line up. Mouth cues remain audio-energy-driven, not viseme alignment. Review the sequence with sound and then muted; compare with the static fallback. Keep voice, motion, factual and accessibility verdicts separate.

The full episode still needs installation/acquisition coverage, fresh scoped execution evidence and complete playback review. The retained source-CLI observation in evidence.json proves only a read-only two-file adopt plan with no writes; no apply, compilation, release or native qualification is claimed. The accepted longer film, rejected audition and native Windows evidence remain intact.