# Residual Born — Grammar Cut: Prediction vs. Result Per-section QC against the pre-registered predictions in `CLEAN-SHEET-TREATMENT.md` (predictions written before any rendering; shot-level QC scored against those lines, not post-hoc vibes). Verdicts below are shot-level. Full-film cold-view QC on the assembled master is recorded at the bottom. Legend: **PASS** = landed the pre-registered line. **NEAR** = right direction, short of the literal bar — accepted with the gap named. **Iterated** = required re-render pass(es) to land. ## S1 — tenderness ("ordinary domestic warmth, no dread") - s1a: PASS (first take). Warm 2700K macro-close domestic, no dread leak. - s1b: PASS, iterated to v3 — early takes broke the phone-glow anchor; v3 landed "looking down at phone." - s1c: PASS via pass-2 re-render. - **Section verdict: PASS.** Restraint held; dread does not leak in. ## S2 — curiosity ("lift that feels like curious momentum, not menace") - s2a: PASS (first take). - s2b: PASS via pass-2 — literal re-prompt got the dense data-noise, subject ~90% obscured, as predicted. - s2c: NEAR (pass-2, accepted) — wireframe portal reads as curious approach but is cleaner/more literal than the staged-degradation language asked for. - **Section verdict: PASS with one named near-miss.** The first cut-compression lands here on the real onset spike, per the measured-audio correction to the old plan. ## S3 — wonder ("awe-adjacent fascination, strange but not yet frightening") - s3a / s3b / s3c: PASS, all first takes. Brightest, most legible section as authored. - **Section verdict: PASS.** ## S4 — surprise ("a small jolt, then release — deliberately throwaway") - s4a: PASS. s4c: PASS. - s4b: NEAR (pass-2, accepted) — kitchen + phone + smart-speaker beat present; the dad's dismissal reads, the discrepant-device detail is subtler than scripted. - **Section verdict: PASS with one named near-miss.** Correctly underplayed — it does not read as a big scary beat. ## S5 — hope-anticipation ("choosing to wait, not nothing happening") - s5a / s5b / s5c: PASS, all first takes. The held-frame restraint carries; the stride-4 midpoint hold is in the cut grid. - **Section verdict: PASS.** ## S6 — dread ("unease without spectacle; must NOT be the loudest moment") - s6a: NEAR (accepted) — interference-band destabilization present, edge-degradation milder than the literal prompt. - s6b: PASS, iterated to v3 — screen-glow-from-below with no one looking up. - s6c: NEAR (accepted) — the one-wrong-regular-thing mechanic reads, the "one beat late" tick is carried by the cut grid rather than in-frame motion. - **Section verdict: PASS with two named near-misses.** The identity-locked shots were the consistent iteration burden, as everywhere in this film. ## S7 — righteous-anger × guilt (the true climax; both must land in the same beat) - s7a: PASS (first take). - s7b: PASS, iterated to take 3 — dad shot needed a literal photographic re-prompt to kill the default too-clean face; final take matches the prediction. - s7c: NEAR (take 2, honest miss vs the literal bar) — the entity-collision shot iterated to a strong take but plateaued short of the literal blown-white highlight the prediction called for. Reads as near-match; reported as such, not rounded up. - **Section verdict: PASS with one named near-miss on the literal highlight spec.** The channel-dissonance construction (loudest audio vs. worst-moment-to-look-away visual, stride-2 hold on the bridge lyric) is in the assembled cut as authored. ## S8 — guilt, unresolved ("the question aimed at the audience") - s8a: PASS, iterated to v3 — extreme bowed screen-glow. - s8b: PASS via recovered v2 take — the render actually succeeded; the submit script's parser misread this ComfyUI version's output and reported success as failure. File verified on disk and picked (entity eyes-close ending). - s8c: PASS (first take). Cursor blink, held past the final downbeat. - **Section verdict: PASS.** S1's close-warm grammar returns cold, as designed. ## Scoreboard - 26 planned shots → all rendered and verified on disk, across 1 base pass + 3 re-render passes. - Shot-level: **19 PASS / 5 NEAR (accepted, named) / 0 unresolved MISS.** - Structural pattern, honestly: environment/abstract shots landed their predictions largely on first take; identity-locked entity/dad shots consistently required literal, concrete re-prompting to defeat the model's too-clean default face — the same fix that cracked S7 generalized. ## Process incidents (root-caused, logged to memory/friction log) - GPU OOM mid-batch: ollama models crept onto the render GPU (twice). Fix: per-submit eviction is now doctrine. - Parser fake-failure: ComfyUI version's output format read as failure on success; three "failed" shots were rescued verified-complete from disk. - Four driver deaths (dead-wait, narrate-without-executing, output overflow, CLI failover) → hardening spec now standard for long-build spawns; final assembly was completed under direct main-session supervision instead of a fifth relay. ## Full-film QC (assembled master) Cold-viewed 2026-07-07 ~00:55Z on the assembled, graded, audio-locked master (`Residual_Born_Grammar_Cut.mp4`, 239.0s, 832x480, h264+aac, 130.1MB, 64 cuts). - **Structure/verified:** duration matches song end-to-end; cut plan executed as authored (64 segments, S8 final hold 18.7s past the last downbeat). ffprobe-clean. - **Audio dynamics arc measured, matches treatment:** S1 open mean -14.2dB → S7 climax mean **-12.0dB / peak -0.3dB (loudest, as predicted)** → S8 close mean -18.9dB (quietest). The loud-bright climax + quiet unresolved close both land. - **Frame sample (9 points across all 8 sections):** S1 warm domestic PASS; S2 light-burst dive PASS; S3 face-resolve PASS; S4/S5 entity holds PASS; S6 screen-glow-from-below PASS; S8 phone-glow close PASS. Grade is consistent (cool-shadow/warm-high split with glow) across all samples. - **HONEST MISS — dad identity drift (top finding):** the dad reads as a *young* man in S1/S6/S8 takes but closest-to-ref (middle-aged) only in the re-prompted S7 take; the reference (`rb_dad_ref.png`) is a middle-aged stubbled man. Shot-level QC scored each shot's pre-registered emotional/compositional line — none scored cross-shot identity match, so every dad shot passed individually while the thread drifted. Each shot still reads unambiguously as "the dad" in its own scene (moody, backlit, phone-lit), so narrative legibility holds, but a cold viewer tracking faces may read 2 men. Fix path if wanted: re-render the 5 dad shots with reference conditioning (VACE/ref-image) rather than prompt-only identity. Reported, not rounded up. - **Verdict:** ships as Grammar Cut v1 — treatment structure, emotional arc, and audio-visual coupling all land; one named film-level miss (dad identity drift) with a concrete v2 fix path, Ethan's call.