Skip to content

raea check: fewer rounds, shorter checks, and raea's own mix and master - #205

Merged
shinyobjectz merged 6 commits into
shinyobjectz:mainfrom
wackyza:check-fewer-rounds
Nov 20, 2026
Merged

shinyobjectz merged 6 commits into
shinyobjectz:mainfrom
wackyza:check-fewer-rounds

Conversation

@wackyza

@wackyza wackyza commented Nov 19, 2026

Copy link
Copy Markdown
Collaborator

Two Telluride walkaround projects (11-14 and 11-16) took nine renders. Five of them chased check findings that no render could move. The checks' verdicts were right about the piece every time. The work here is in how a finding is routed, when it's found, and how long a check takes. Two measurement bugs are fixed on evidence. No rule's threshold changes.

What changes

  1. raea check says who can fix what it finds (85463a91)
    • A rule an encode can cause is measured again on the composition drawn at the same frames without encoding (raea-app-check-cause). Blotching is the first such rule. The walkaround's phone footage is blotchy in 9 of 28 samples before any encoding, so every export failed and nothing cleared it. Now the picture's own blotching is a note. What the encode added is the error, with a fix raea has: --quality high added to 14 of 28 samples, x264 crf 12 moved the score at most 0.005, and --quality master added none.
    • A finding an earlier export of the same composition had at the same moment ends [also in …]. Its rank stays as measured.
    • A take under assets/ or media/ is a source, not a delivery: its loudness is a note.
    • The text rules move to raea-app-check-text.
  2. raea render holds a video's text to the check's rules as it ends, and keeps the lighting advice off 2D fill (5cd6fbdb)
    • Text off the frame is an error. Text in a UI zone is a warn by the default zones, since the platform's zones are the check's to decide.
    • A frame with no 3D and no footage is the page's own fill: it's measured, not given lighting advice. Frames with footage keep their warnings.
  3. Clipping is measured on each channel (a3eb4e74). The rule read the mono fold, (L + R) × 0.707, where a centred voice reads 3 dB hotter. The walkaround's v1 and v2 masters peak at -1.5 and -1.6 dBTP and no channel nears full scale, yet the fold found 9 to 36 flat-topped runs. Those cost two renders and a voice mastered outside raea. Measured per channel, neither file clips.
  4. raea mixes and masters every render's sound (d50dbe62)
    • Dialogue is a clip on the voice bus, or footage's sound on no bus. Each line is levelled to -23 LUFS over the span of its take it plays (raea-sys-mix-level). Other buses keep their levels, and gainDb is an offset on top.
    • A leveller in the limiter works from 9 dB under the ceiling at 3:1, so the limiter only catches what's left. Live play, the render and seedling's limiter all get it.
    • Without a <Mix>, every render is mastered to -14 LUFS at -1 dBTP.
    • A join into dead silence gets the room's own tone.
    • Result on the walkaround's v1 sound, each line straight from its take: the lines' spread drops from 6.7 to 2.1 LU, no dropout, -14.5 LUFS, -1.6 dBTP, with nothing set by hand.
  5. One read of the picture for flashing, freezes, flash frames and blotching (46d840f1)
    • Four full decodes ran one after another. Run in parallel they take the same time, because decoding is CPU-bound.
    • One FFmpeg now splits the frames into the four filter chains, each with its own options as when it ran alone.
    • tests/parity.rs holds the one read to the four separate reads. The findings JSON is byte-identical on every export in the replay.
    • Checks take 7-8 s where they took 13-20 s, or 20-29 s (was 37-59) when the blotching comparison also runs.
    • Without named pipes (Windows with FFmpeg) the rules read separately as before.

Behaviour to know about

  • Every render's sound is now mastered, and games get the leveller too. MASTER_VERSION is now 4.
  • Live play hears a dead join; only the render's master fills it.
  • A check whose blotching fires renders the sampled frames without encoding: 10-15 s, only then.
  • The guides stop telling agents to level lines with gainDb, master the voice outside raea, or add a <Mix> just to get a master. The rendering guide says to fix every error in one pass, and that a note is for the person.

Test plan

  • Replayed both projects' exports through the old and new checks (scratch copies, before-binary built from main)
  • cargo test for raea-app-check, -check-read (parity incl. the one read), -check-text, -check-cause, -direction, raea-sys-mix, -mix-level, runtime-export, -sound, -encode, io-codec
  • CI's unit test set, cargo check --locked -p raea-cli --tests
  • layers, names, size, one-version, no-void, dupes, windows-paths, rask, sdk-coverage, sdk-props, taxonomy; cargo shear has the same 2 warnings as main

The skill copies of the guides follow in raea-skills.

🤖 Generated with Claude Code

Zachary Guglin and others added 6 commits November 19, 2026 16:41
…at another render can change

Two Telluride walkaround projects (11-14, 11-16) took nine renders; five chased findings no render could move. The
checks measured right every time; what they didn't say was whose each finding was. Nothing a rule measures, and no
line it draws, changes here: what changes is how a finding is told.

- The picture's own or the encode's (raea-app-check-cause). A rule an encode can cause is measured again on the
  composition drawn at the same frames without encoding (raea render --at, out of the record). Blotching is the first:
  the walkaround's phone footage scored blotchy in 9 of 28 samples before any encoding, so every export failed and no
  setting cleared it. Now the picture's own is a note, and what the encode added is the error with a fix raea has:
  --quality high added to 14 of 28 samples (x264 crf 12 moved the score at most 0.005; master added none), so the
  check says render with --quality master, and the master export then passes with the note.
- Carried. A finding an earlier export of the same composition had at the same moment ends "[also in …]": what changed
  between them didn't move it. Its rank stays as measured.
- A source isn't a delivery: a take under assets/ or media/ checked on its own reports its loudness as a note.

The text rules (clipped, UI zones, reading time, face, cues, brand kit) move to raea-app-check-text, so raea render can
hold its own text to them, and raea-app-check stays under its size limit. Probe gives the video's frame rate.
The rendering guide tells agents to fix every error in one pass, not to repeat a change that didn't move a carried
finding, and that a note is for the person.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… keeps a 2D frame's fill out of the lighting advice

A round sooner: what the render's manifest already knows is said by the render, not left for the check after it (the
11-14 walkaround's price figures in the UI zone waited a render and a check). Text off the frame is an error, as the
check has it. Text in a UI zone is a warn by the default zones, since the platform's are the check's to decide; reading
time waits for the check, which has the transcript. Video renders only: a stills render samples too few frames to time
a text. raea check's issue-to-finding conversion is now Issue::finding, which both use.

Exposure's lighting advice ("a fill, an ambient or the sky's light reaches them nowhere") is a 3D scene's. A frame with
no 3D in the render and no footage on it is the page's own fill, dark or bright by design (the walkaround's brand-dark
logo card warned on every stills render), so it's measured and kept in the manifest, not said. A frame with footage
keeps its warning: a clip that drew black still shows.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ean master as clipped

The clipping rule read the sound folded to mono, which ffmpeg's -ac 1 (and raea's fold, after it) make at
(L + R) * 0.707: a centred voice comes out 3 dB hotter than either side. The 11-14 walkaround's v1 and v2 masters peak at
-1.5 and -1.6 dBTP, no channel near full scale, yet their folds passed 1.0 and the check found 9 to 36 flat-topped runs
over full scale. The agent spent two renders on it and then mastered the voice outside raea; raea check on those same
files, each channel at its own level, finds no clipping. Clipping is a channel's: each is now read as it is (read::channels,
FFmpeg's -ac 2 or raea's own decoders), and a second is the most runs any channel had. A clipped channel is still found,
one that the fold would have hidden among the other too. The dropout and dead-stop rules keep reading the fold.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… leveller before the limiter, room tone at the joins

The person's rule, 11-19: raea does its own mix on everything, then masters it. Until now a render was
mastered only with a <Mix>, and then by a gain and a limiter alone; levels were the agent's to set. The 11-14 walkaround's
lines ran from -13.8 to -26.9 LUFS, so the agent measured each with ffmpeg and set its gainDb, the limiter still took
up to 8 dB off phone speech's peaks, and the joins between lines fell to dead silence (a dropout, the check said). It
then built and mastered the voice outside raea.

- The mix (raea-sys-mix-level, level.rs): each line of dialogue, a clip on the voice bus or footage's sound on no bus,
  is measured over the span of its take it plays (BS.1770, gated) and brought to -23 LUFS, by at most 12 dB. Room tone
  and silent takes (under -45) keep their level; so do music, effects and every other bus, which sit under dialogue as
  their buses set them. gainDb is an offset on top. Live play and the render level the same span the same way (the
  clip's span comes from the SDK's data-clip-from and data-clip-to).
- The master: a leveller in the limiter (from 9 dB under the ceiling, 3:1 on the mix's RMS, 10 ms in, 150 ms out)
  rounds the loud passages so the limiter catches only what's left, in live play, the render and seedling's limiter
  alike; the gain is made up for what it took. Without a <Mix>, every render masters to -14 LUFS at -1 dBTP.
- The joins (joins.rs): a gap the mix drops into dead silence and comes straight back from gets the room's own tone, the
  quietest stretch of sound near it, looped and faded, found at the level the master plays it. A pause that fades is
  the edit's and stays. The render fills them; live play hears the gap.

The walkaround's v1 sound, each line straight from its take: before, its lines spread 6.7 LU and it had a dropout (and
the mono fold's false clipping); now 2.1 LU, no dropout, -14.5 LUFS, -1.6 dBTP, with nothing set by hand. The Kia cut,
whose voice was mastered outside raea, comes out at -14.1 LUFS, -1.2 dBTP. MASTER_VERSION 4. The guides stop telling
agents to level lines or add a <Mix> to be mastered.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… and blotching: the same findings in half the time

Four of raea check's rules each decoded the whole export, one after another: 3.7 s apiece on the 11-14 walkaround,
38 s of FFmpeg on a 56 s master export. Run at the same time they took 15.8 s, the same as in turn: decoding is the
machine's whole CPU, so only decoding less saves anything. One FFmpeg now decodes it once and splits the frames into
the four filter chains the rules were tuned on, each to its own output with its own options as when it ran alone
(raea-app-check-read once.rs): the flashing screen's frames to stdout with their showinfo times on stderr,
freezedetect's lines on stderr, the 32x18 grey and blotching's 2 s crops to named pipes. Each rule is split into its
read and its judgement, and the judgements run on what the one read gave, in the order they ran before.

tests/parity.rs holds the one read to the four alone on a clip FFmpeg makes: the same flashing frames and times, the
same brightness, the same crops and the same freeze. On both Telluride projects' exports the check's findings and passes
are identical to before (its JSON byte for byte), and it takes 7-8 s where it took 13-20, and 20-29 s where blotching's
comparison render also runs (37-59 before). Without named pipes (Windows with FFmpeg) the rules read alone as before,
and without FFmpeg raea's own readers do.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…an: the exposure filter's comment trimmed

rask's size gate counts a crate's planned tasks: the 2D fill filter left raea-runtime-export at 2001 lines with them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@shinyobjectz

Copy link
Copy Markdown
Owner

Thanks, this is careful work, and the evidence for each change is clear. A first review, before the maintainer's decision:

The one read (once.rs) looks right. Every output has its own reader thread, so no pipe can fill while another is waited on. A frame whose showinfo line doesn't come within 60 s stops the read instead of hanging it, and a FIFO FFmpeg never opened is unblocked before the joins. tests/parity.rs holds it to the four separate reads, and the fallback without named pipes keeps Windows as it was.

Per-channel clipping is the right measure. The mono fold reading a centred voice 3 dB hot explains the false flat-tops.

What needs the maintainer's call before merging, because it changes what every existing project sounds like rather than fixing a bug:

  • Every render mastered to -14 LUFS / -1 dBTP without a <Mix>. That matches the loudness raea check already holds a posted film to. But a piece that is quiet on purpose gets lifted, and MASTER_VERSION 4 means every cached export is rendered again.
  • Dialogue levelled to -23 LUFS per take on the voice bus, with gainDb becoming an offset on top of that.
  • Room tone generated into joins that drop into dead silence. That adds sound nobody placed, and some cuts want the silence.

Each of these might be better as the default with an opt-out (a <Mix> prop, say) than with no way out. That's the question I've left for the maintainer, along with the merge. raea-skills #104 should land with this, since it changes the guides to match.

@shinyobjectz

Copy link
Copy Markdown
Owner

Decided: these go in as they are, for 0.1.9.

  • Mastering every render to -14 LUFS / -1 dBTP is the target raea check already holds a posted film to, so without it every export without a <Mix> failed its own check.
  • Levelling each take's dialogue to -23 LUFS and room tone in joins that drop into digital silence are what a dialogue editor does by hand.

If someone needs a render left unmastered, that's an opt-out on <Mix> to add later, not a reason to hold this. Merging with raea-skills #104. Thanks for the careful evidence on each change.

@shinyobjectz
shinyobjectz merged commit 3ffe34b into shinyobjectz:main Nov 20, 2026
4 checks passed
shinyobjectz pushed a commit to shinyobjectz/raea-skills that referenced this pull request Nov 20, 2026
…nd goes only to what another render can change

The copies of sdk/guides/craft/motion/{rendering,sound,delivery}.md from shinyobjectz/raea#205: raea levels each line of
dialogue and masters the piece, so agents don't level lines with gainDb, master the voice outside raea, or add a <Mix>
to be mastered; blotching is the encoder's or the footage's own, and the footage's is a note, never a reason to render
again; fix every error in one pass, don't repeat a change that didn't move a finding marked [also in …], and a note is
for the person.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
shinyobjectz added a commit that referenced this pull request Nov 21, 2026
…rg the one list

It also carries raea-skills #104, whose raea side (#205) is already in main.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants