When a film is code, every frame must be a pure function of t: no Math.random, no Date, no state carried from one frame to the next. A renderer can then seek to any time, out of order, and average several sub-frame samples into motion blur that is real rather than painted. One stray Math.random makes the samples of a frame disagree with each other, and the blur turns to noise. Seeded randomness is fine; the rule is the same one vibe-to-generator sets for generators.

The shutter

Default: a 180° forward shutter, 4 samples a frame. Grain and bloom are added in the render, after averaging, and the grain is seeded so a re-render matches the last one. Never fake blur, grain or bloom in the page.
Fast windows raise the count: a whip pan stretches the shutter to about 600° with up to 24 samples. The smear is what makes a whip read as one move (launch-video-cuts, "into motion").
Never cross a hard cut. If a cut falls inside the shutter, slide the window back so it ends on the cut. A clipped window renders the fastest frame of a whip sharp, which reads as a stutter on exactly the frame the eye is following.
Counters and HUD digits read the output frame time, not the sample time, or blur smears the numbers.

The grid comes from the track, not from the brief

Everything is placed in beats (B(k) = k × 60 / bpm), so one tempo change retimes the film. A generated track's tempo is a hope until it is measured:

Give the beat tracker the intended tempo, or it can lock to half or two thirds of it.
Fit only the stretch the film uses. Generated music drifts, and only that stretch has to be exact.
Drop alignment: music.offset = drop_seconds − planned_drop_beat × 60 / bpm. Check the result against the listed downbeats.
Prefer a music model whose schema takes a tempo. In the Kaiju Postal build, both takes asked for 172 BPM measured 172.0.

Over a 50 s film, a 0.15 % tempo error puts the last hits about 70 ms late. The sync budget is 0–30 ms, never early (sound-motion-sync), so a grid that is "nearly right" fails by the end card.

Delivery gates

Hand over nothing until every file passes all of these:

Gate	Why
Frame count equals the film's	a mux with `-shortest` silently drops the last frames when the audio ends a few ms early
H.264 Level ≤ 5.2 at 4K, ≤ 4.2 at 1080p	a slower x264 preset without `-level` writes Level 6.0, which plays in ffmpeg and fails in QuickTime, Safari and on phones
Under 1 GB at 4K	a 52 s 4K master with grain is 0.96 GB at CRF 12.4 (preset slower, ~50 dB PSNR); CRF 11 at preset medium reaches 1.37 GB for a 51 dB difference no one can see
A full decode on the hardware decoder	what QuickTime and phones use; a software decode passing proves nothing
Loudness written for the register, −1 dBTP	launch-video-sound; the engine's −14 LUFS default is not a standard (scenario-kinetic-reconciliation)

Render 1080p until the film passes, then the 4K master once, then run the gates again on the master. Make the 1080p and 9:16 files from the master, and lay the 9:16 one out again per launch-video-formats, not as the film centred on black.

Gotcha

Headless Chrome has capture traps of its own, separate from the render traps in gotchas. A 1920 × 1080 viewport paints only about 990 px, so use a taller viewport and clip. captureScreenshot returns CSS pixels unless clip.scale equals the device scale factor, so a "4K" render comes back at 1080p. Large 3D layers are captured half-drawn without --run-all-compositor-stages-before-draw. Each 4K worker needs about 11 GB of memory; six of them stalled a 36 GB machine.

Sources

Scenario, scenario-kinetic-film skill (zip received 2026-10-30, no licence file, so it is cited here and none of its code is copied): SKILL.md (rules table, common mistakes), references/render.md (pipeline, traps, encode settings, gates), references/sound.md §1 (beat grid, offset), references/method.md §4.
Related: launch-video-review, launch-video-music, cinetic-reconciliation, hyperframes-reconciliation.