Grade Clip

grade-clip

The keystone of the selection/repair/drip suite. The pipeline today is open-loop and produce-only — qc-clip checks only duration/size, visual-cadence is one diagnostic. Nothing ever inspects a delivered pixel-level artifact for upload-readiness. grade-clip is that gate: the on-disk translation of YouTube's un-exportable VVSA ("viewed vs swiped away") front-gate, scored 0-99 off the finished .mp4 and its persisted sidecar plans.

grade-clip.sh <clip.mp4>                       # single / in-chain (after save-local)
grade-clip.sh --backlog [output_dir=output]    # sweep output/<src>/*.mp4 + _triage.json

Output — clip_NN.grade.json (locked SELECTION-SUITE-CONTRACT schema)

{
  "clip": "output/<src>/<title>.mp4",
  "grade": 0,
  "tier": "GOLD|FIXABLE|DROSS",
  "hard_caps": ["letterbox","face_withheld","credit_at_open","blocking_card","dead_tail"],
  "signals": { "frame1_is_face": true, "letterbox_bars": false, "credit_lit_at_open": false,
               "first_visual_change_sec": 1.2, "first_payoff_offset": 2.4,
               "longest_static_gap": 3.1, "opening_caption_words": 5,
               "max_residual_silence": 0.4, "terminal_loop_score": 0.37,
               "claude": {"hook_payoff":7,"open_loop":6,"cold_context":8} },
  "fix_routes": ["broll_open_truncate","shot0_repunch","credit_rerender"],
  "source": "<source-slug>"
}

Algorithm — deterministic retention-proxy floor (no model)

• frame1_is_face — decode frame 0, MediaPipe face detection (reuses fill-vertical's face_landmarker.task); TRUE if a face spans >~2% of frame area. If .vert.fillplan.json is present, factor shots[0].kind != "face" as the face-withheld signal. face_withheld hard-cap fires when no face in frame0 (or fillplan shot0 kind != face). Policy is A/B-able via GRADE_FACE_OPEN: face (default, legacy) keeps the hard cap; broll inverts it (opening ON a talking head takes a soft -8 penalty, no cap, no broll_open_truncate route — the 2026-07-07 perf-style analysis found opens_on_face ranked -0.94 against likes/1k); off disables the frame1 judgment entirely.
• letterbox_bars — sample top/bottom + left/right edge bands; flag a near-constant / very-low-variance band (black bars / blurred pillarbox). A correct full-bleed punch-in has high variance everywhere. letterbox hard-cap (structural → rerun_recommended).
• credit_lit_at_open — edge density in the top ~12% banner at t≈0.6s; flag a dense text strip in the first ~1s (the source credit should appear only in the final CREDIT_TAIL). Approximate. credit_at_open hard-cap → credit_rerender.
• first_visual_change_sec — in-chain: min over (first broll pick t0, first caption swap = chunk0 t1); backlog: first ffmpeg scene-detect change. null if none.
• first_payoff_offset — in-chain: lexical-overlap match of the title's key noun/verb against .chunks.json → t0 of the best-matching chunk; backlog: null.
• longest_static_gap — read .cadence.json max_gap; backlog: compute via scene-detect (visual-cadence logic).
• opening_caption_words — word count of chunk0 within the swipe window; backlog: null.
• max_residual_silence — longest ffmpeg silencedetect silence (sec).
• terminal_loop_score — frame0 vs last-frame normalized-histogram correlation → 0-1 (the >100%-retention loop lever).
• dead_tail hard-cap — trailing silence in the last ~1.5s AND a static final ~1s (low frame delta) → dead_tail (no clean fix route on its own).

Hard caps (letterbox, face_withheld, credit_at_open, blocking_card, dead_tail) any present → grade clamped ≤40. Each maps to a fix_routes entry: letterbox → rerun_recommended, face_withheld → shot0_repunch, credit_at_open → credit_rerender, blocking_card → card_rerender. A broll_plan.picks window overlapping [0, GRADE_OPEN_GUARD_SEC] adds broll_open_truncate (only under GRADE_FACE_OPEN=face).

Grade formula (explicit + commented in grade.py): start near 99, subtract documented penalties for the soft signals (large static gap, late/None first_payoff_offset, high residual silence, weak terminal-loop, too-few opening caption words, no visual change before 3s), then apply the Claude rubric (when present), THEN clamp ≤40 if any hard cap.

tier: GOLD = grade ≥ GRADE_MIN_UPLOAD (60) & no hard caps; FIXABLE = has hard caps but ALL map to a non-rerun_recommended route; DROSS otherwise.

Algorithm — one batched Claude call (the only model use)

When GRADE_SKIP_CLAUDE is unset/0 AND a transcript is available, ONE run_claude_step call rates hook↔payoff coherence + open-loop strength + cold-viewer context, each 0-10, on the opening transcript (first ~10s) + title (build_prompt.py → run_claude_step → parse_reply.py, neutral-5 fallback on unparseable reply). Result lands under signals.claude. GRADE_SKIP_CLAUDE=1 skips entirely. Backlog mode defaults proxy-only — it does NOT call Claude unless GRADE_SKIP_CLAUDE=0 is explicitly set.

Backlog triage

--backlog sweeps every output/<src>/*.mp4 (skips _preview/, source/, _toupload/), writes <clip>.grade.json next to each, and an aggregate output/_triage.json = {generated, n, gold:[...], fixable:[{clip,defect}], dross:[{clip,reason}], by_source:{...}}. Finished output clips usually have NO co-located sidecars, so backlog mode is PROXY-ONLY by direct pixel/ffprobe reads.

Knobs

GRADE_MIN_UPLOAD (default 60), GRADE_SKIP_CLAUDE (1 = proxy-only; backlog default), GRADE_OPEN_GUARD_SEC (2.2), GRADE_FIRST_CHANGE_SEC (3.0), GRADE_PAYOFF_SEC (3.0), GRADE_STATIC_GAP_SEC (5.0), GRADE_SILENCE_SEC (0.8), GRADE_MIN_CAPTION_WORDS (3), GRADE_SILENCE_DB (-30dB), GRADE_SCENE (0.3).

Idempotent — mtime+param .gcmeta signature over the clip + every sidecar that exists. Non-fatal — any error emits a DROSS/empty verdict and exits 0; it never hard-fails the pipeline.