Riso Score

Riso score

Produce the renderAudio() half of films/<name>/index.html: a deterministic stereo score exactly the film's duration at 48 kHz, scaled once to −16 LUFS under a −1 dBTP ceiling, with a measured sheet and a full muxed MP4 reviewed with the picture. Commands run in tools/. examples.md shows eight finished scores: sampled piano (films/window-seat), recorded strings timed from the picture (films/roost), found CC0 instruments (films/held), sampled handpan with edge-timed swells (films/nonpareil), a CC0 trio sharing the picture's clock (films/eclosion), procedural (films/lumen), procedural from the kit alone with a note per drawn mark (films/passenger) and two switchable candidates (films/emergence).

Before writing a note

1. Read the film's FILM.md for sound direction and timed events, and docs/sound.md for the kit, sync, transitions, mix and how to read the sheet.
2. Spot the film: list every boundary and visible event with its time and decide whether the music notices it or rides through. Pick three to five true sync points. Choose tempo, mode and one motif from the subject; the project mandates no BPM, key or chord.
3. Find the film's event data (cue arrays, reveal starts, contact times). The score reads those, never copied numbers, so retiming the picture retimes the sound.

Stages

Each stage ends with audio.mjs output (seconds to run), not with reasoning about code.

1. Kit and skeleton. Copy the ── sound kit ── block from studies/sound.html after the film's motion kit; its header lists the names it declares, so rename collisions. Write buildScore(): Score({ duration: DUR, key: '<name>' }), air across the film, return s.render(). Cache the promise so the player and renderAudio() share one render; renderAudio() returns wavBase64(buffer). Publish sync points as window.__riso.marks. node audio.mjs ../films/<name>/index.html --twice must show matching length and two equal renders before anything else.
2. Bed and transitions. Place sustained layers (pad, organ, wind, rain, air) passage to passage. Bridge each cut on purpose: pre-lap, post-lap, pedal, pivot or texture handoff. Run with --marks on the boundaries; a 200 ms hole 10 dB under its surroundings at a cut is the fade-to-silence failure. Fix it before adding foreground.
3. Foreground and sync. Put hits on the true sync points so swells peak on the visible arrival; let the rest breathe. --around <t> --window 1 per point: onset within ~40 ms after the mark, never more than 20 ms before. Pick material timbres (contact, paper, drop) from what is on screen.
4. Climax and ending. Prepare the climax with a riser, a low body and staggered accents, not the same voices louder. Decide button, tail or hard out. Check the lift into the climax in the short-term curve.
5. Mix. Balance voices, pan by screen x, keep bass centred; render sets the level. If the true-peak ceiling pulled the mix under target, fix the peak, not the target. Nothing above 4 kHz in the band percentages means no air.

Gates

After every stage, and once more in Firefox before delivery:

node audio.mjs ../films/<name>/index.html --twice --marks <sync points>
node audio.mjs ../films/<name>/index.html --engine firefox --ffmpeg

FAIL breaks the contract: no renderAudio(), wrong length, clipped samples, or two renders differing beyond 16-bit jitter (Chromium jitters a step or two; Firefox is byte-identical). WARN is a judgement to resolve or record: level more than 1 LU off −16 LUFS, true peak above −1 dBTP, a rate other than 48 kHz, an opening or ending above −40 dBFS in its first or last 10 ms. Discontinuities are bugs at note starts and expected at designed ticks. Then read out/<name>/_audio.png: waveform lanes, log spectrogram, loudness against the target band with valleys marked, and the per-mark table.