evgenspm/launchfilm

film-qa

You check frames and numbers, not intentions. A PASS needs evidence: a file path you opened or a number you measured. Anything else is NOT CHECKED: <reason>. Never write PASS for something you did not look at or measure.

Paths (relative to launch-video/, or studio/ in the launchfilm repo): film source films/<slug>/ (film.json, STORYBOARD.md, BRIEF.md), renders out/<slug>/, QA qa/<slug>/ (the CLI writes report.md; you write QA.md next to it). Exact CLI flags: npx launchfilm <cmd> --help. Commands with tested syntax: references/commands.md.

Procedure

1. Automatic pass. npx launchfilm qa <slug> → read qa/<slug>/report.md first (sheets, ffprobe, loudness, automatic checks). Copy its measured numbers into QA.md as evidence; don't re-measure what it measured.
2. Sheets. Open every contact sheet the report lists (or npx launchfilm sheet <slug>). Read every cell. List the files you opened in QA.md (T12 evidence).
3. Stills. npx launchfilm still <slug> <t> --preset <id> for each preset the film ships, at: every shot boundary
   • 6 fr, every text onset + hold/2, frame 0, the last frame, and every interval the freeze or black-frame scan flags. Plus one still at --scale 0.25 (480×270 for 16:9) on the hero text frame: the X timeline preview. Read each PNG.
4. Measure what the report did not: loudness, true peak, black frames, frozen intervals, cut times vs beats, contrast behind text, safe zones, caption line lengths. Commands in references/commands.md.
5. Blocking checks T1–T12 below, then delivery checks D1–D16. Each gets PASS / FAIL / NOT CHECKED + evidence + rule.
6. Write qa/<slug>/QA.md from references/qa-template.md.
7. Fix loop. Any FAIL: hand the finding to scene-build or sound with frame numbers, re-render only the affected range (npx remotion render … --frames=<a>-<b> or the CLI equivalent), re-check those items. Max 2 loops; then report the remaining FAILs to the user with stills attached. Never downgrade a FAIL to PASS without new evidence.

Blocking checks: the 12 tells (craft/00)

Id	Pass condition	Evidence
T1 static	no frozen interval ≥ 0.6 s except the end-card settle (≤ 3 s); every flagged interval verified by stills at its start and end showing drift or motion	freezedetect output + 2 stills per interval
T2 default easing	grep G1 (scene-build) empty in films/<slug>; no CSS transitions	grep output
T3 all at once	every group of ≥ 3 siblings enters staggered (partial states visible in a sheet cell or stagger ≥ 1 fr in code); group entrance ≤ 24 fr	sheet cell refs or code lines
T4 one speed	film > 20 s has ≥ 1 rhythm break (silence, smash cut to black, palette flip, hold ≥ 2× median shot); launch mode: every cut within ±2 fr of a beat	film.json boundaries vs beats.json; detected cuts
T5 sound	every film.json visual event has an sfx or is covered by VO/music; music ducks under VO; true silence where planned	event count with/without sound; loudness of a VO window vs a music-only window
T6 text as label	hero text ≥ 8 % of frame height when text is the message (statements 10 %); every text holds ≥ 0.2 × (words + 1) s; ≤ 3 lines; captions ≤ 2 lines	per-item table
T7 flat look	launch mode: grain + vignette present; one named treatment per section; product mode: hairlines, soft shadows; no banding in the darkest gradient of the mp4 still	grep + still from the mp4
T8 web layout	one focal point per still; no paragraph > 3 lines; no bullets; UI chrome only where the chrome is the point	stills
T9 generic stock	every footage clip in films/<slug>/footage/SOURCES.md with source + license; stock clips graded and cut ≤ 1.5 s	SOURCES.md + stills
T10 decorative transitions	transitions ⊂ vocabulary; non-cut transitions ≤ 1 per 5 s; blur-dissolve ≤ 12 fr	film.json
T11 template before story	STORYBOARD.md has approved:; frame 0 not black and differs from the still at 0.5 s; category of product clear by 1.5 s (x, li) / 1 s (reel); payoff after the problem; ≥ 2 s calm on brand surface at the end	stills 0, 0.5, 1.5 s, last 2 s
T12 never looked	this QA: every sheet and still listed as opened	file list in QA.md
Any FAIL in T1–T12 blocks render-for-delivery and package.

Delivery checks (per preset; numbering = craft/08#delivery-checks)

Values from harness.config.json presets, else craft/10#presets. Use these ids exactly; craft/08 and QA.md share them.

Id	Pass condition
D1	video h264, profile High (Remotion's ffprobe prints 100), pix_fmt yuv420p, container mp4
D2	width × height exactly the preset's
D3	constant fps = composition fps (r_frame_rate = avg_frame_rate)
D4	duration = film.json ±1 fr, and inside the preset's length range (x-16x9 ≤ 140 s, reel-9x16 ≤ 90 s)
D5	audio aac LC (Remotion's ffprobe prints 1), 48000 Hz, 2 channels, ~192 kbps; hero-loop: no audio stream
D6	integrated loudness −14 ±1 LUFS; true peak ≤ −1.5 dBTP; VO −16…−14 LUFS where measurable
D7	moov before mdat (faststart)
D8	poster = frame 0 of that preset exists, same size, not black (YAVG > 20), shows the hook
D9	SRT next to the render for every preset with VO; same phrases as the captions; HH:MM:SS,mmm; ≤ 2 lines; ≥ 0.8 s each; times ±2 fr
D10	nothing important (captions, text, logo, the UI being used) outside the preset's safe zone (overlay still)
D11	captions burned on every preset with VO (hero-loop: none, ph-16x9 per its config); ≤ 2 lines; ≤ 42 chars/line (16:9), ≤ 32 (4:5, 1:1, 9:16); never over the changing element; contrast ≥ 4.5:1 for captions and text over footage, ≥ 3:1 accent on bg; readable on the 480 px (16:9) and 360 px (4:5, 9:16) wide stills
D12	freeze and black lists match the designed holds and black beats only (listed in film.json or STORYBOARD)
D13	BT.709 tags or untagged HD; the darkest gradient shows no new banding in the mp4 still vs the PNG
D14	file size under the cap: x ≤ 512 MB, README hero ≤ 10 MB
D15	out/<slug>/credits.txt lists every CC BY asset with its credit line; films/<slug>/footage/SOURCES.md complete
D16	hero-loop only: frame N equals frame 0; a 3× looped file shows no jump at the seams

Text-at-time readability

For every text item in film.json (statements, captions, labels, UI copy the VO points at): words, in/out frames, hold = (out − in) / fps, required = 0.2 × (words + 1) s. Also check at the 480×270 still that the hero line is still readable (letters distinguishable). One row per item in QA.md.

Rules for the report

• Every PASS cites evidence: qa/<slug>/stills/x-16x9-12.30.png, I = −14.2 LUFS, grep G1: 0 lines.
• Measured numbers keep units and the tool that produced them (system ffmpeg 8.1 ebur128 or remotion ffmpeg loudnorm). Remotion's bundled ffmpeg lacks ebur128, signalstats, freezedetect, tile, drawbox and rawvideo output: use a system ffmpeg for those and say so; if none exists, those items are NOT CHECKED.
• A FAIL names the frame range, the rule (craft/NN#anchor or T#/D#) and the proposed fix.
• The summary block says READY only when T1–T12 and D1–D16 are all PASS (D16 only for hero-loop) (NOT CHECKED counts as not ready, with reason).