diff --git a/skills/raea/motion/delivery.md b/skills/raea/motion/delivery.md
index 3b935bf..54e5295 100644
--- a/skills/raea/motion/delivery.md
+++ b/skills/raea/motion/delivery.md
@@ -18,8 +18,8 @@
`raea check out/launch-9x16.mp4 --manifest out/render.json --shape 9:16 --platform tiktok`.
- Captions: burn them in for feeds that play without sound, unless the delivery controller says captions off; then
burn none, and leave the words file beside the export for the platform's own captions.
-- Sound to -14 LUFS integrated with true peak at or below -1 dBTP: one `` per
- composition does it (`motion/sound`). A spec that names another number wins.
+- Sound to -14 LUFS integrated with true peak at or below -1 dBTP: raea's mix and master do it on every render
+ (`motion/sound`). A spec that names another number wins: `` sets it.
- Render the export at `--quality master` (motion graphics and type) or `--quality high` (a film), and versions made
for notes with no `--quality` flag; name it in lowercase letters, digits and hyphens, with its shape and length:
`spring-launch-9x16-15s.mp4`. Each version is a new file beside the last, never written over it.
diff --git a/skills/raea/motion/rendering.md b/skills/raea/motion/rendering.md
index cbae3d7..99fda4b 100644
--- a/skills/raea/motion/rendering.md
+++ b/skills/raea/motion/rendering.md
@@ -8,7 +8,10 @@
It measures clipped text, text under a platform's buttons, text over the face, captions on too briefly to read,
caption contrast against the pixels behind them, frozen stretches, flashing past the limit, cues that miss their word, loudness, a sound that
stops dead or (with `--words`) fights the voice, and encoder blotching in the darks. Each finding is ranked P0 to P3
- with its fix; it exits 1 while a P0 or P1 stands, and `-o check.json` writes them for the critique. An
+ with its fix; it exits 1 while a P0 or P1 stands, and `-o check.json` writes them for the critique. Fix every error
+ in one pass before rendering again. A finding that ends `[also in …]` was in the earlier export too: what you changed
+ didn't move it, so don't repeat that change; try another, or say why it stands. A note is for the person, not a
+ reason to render again. An
exception in the composition (an unclamped value going NaN, keyframes out of order, a phrase `word()` can't find, a
missing prop) fails the render and names the line.
- **Watch it as a viewer would once the piece is close.** `raea check` measures; it can't judge a frame cropped too
@@ -21,7 +24,10 @@
- **Without `--manifest`**, `raea render` writes the export alone, without the measurements `raea check` reads. Render
another deliverable from its own file the same way (`raea render NAME.tsx -o NAME.mp4 --manifest NAME.json`) and
check it.
-- **Loudness** comes out at -14 LUFS integrated (±1) with true peak at or below -1 dBTP when `` is in the tree.
+- **Loudness** comes out at -14 LUFS integrated (±1) with true peak at or below -1 dBTP: raea levels each line of
+ dialogue (the `voice` bus, or footage on no bus) to the same loudness and masters the piece, so don't measure lines
+ or set `gainDb` to level them, and don't master the voice outside raea. `gainDb` is for a deliberate change, on top
+ of raea's level. `` sets another target.
## Why
@@ -45,9 +51,10 @@ looking again, finds more than a still per change and costs less.
A video's bitrate follows its size and frame rate (about 20 Mbit/s at 1080p30, 80 at 4K30; VideoToolbox spends less on
flat pictures), and `--kbps N` sets it. `--quality draft|high|master` (or 0..1: 0.6, 0.85, 0.95) asks for constant
quality instead, each frame getting the bits it needs: `master` keeps thin type and hard vector edges clean at 4K
-(about 31 Mbit/s on flat motion graphics). Blotching in dark areas is the encoder, not the composition: export dark
-material as PNG frames (`RAEA_READBACK=rgba raea render piece.tsx -o piece.frames/`) and encode them with more bits for
-the darks. A `NAME.frames/` folder is never committed, and raea's storage dry run lists it a day after
+(about 31 Mbit/s on flat motion graphics). Blotching in dark areas is the encoder's or the footage's own, and on a
+render made in a project `raea check` says which: the footage's is a note to pass on, never a reason to render again;
+for the encode's, render with `--quality master`. Dark material can also go out as PNG frames
+(`RAEA_READBACK=rgba raea render piece.tsx -o piece.frames/`). A `NAME.frames/` folder is never committed, and raea's storage dry run lists it a day after
the render unless something names it.
Which to ask for. A version made for notes (a cut to review, a shot that will be joined, a preview) takes no
diff --git a/skills/raea/motion/sound.md b/skills/raea/motion/sound.md
index 2a6fd6f..c5a9eff 100644
--- a/skills/raea/motion/sound.md
+++ b/skills/raea/motion/sound.md
@@ -36,11 +36,14 @@
- Narration over the picture, when the footage has no voice for it, is
`` from `raea/speech` (Pocket TTS, on this computer; `voice` picks a
Pocket voice or a WAV of one).
-- Master the sum with one `` per composition. It brings the whole mix to -14 LUFS
- integrated and limits true peaks under the ceiling, which is what the floor's loudness defect asks for.
+- raea mixes and masters every render: each line of dialogue (the `voice` bus, or footage on no bus) comes to the same
+ loudness, a join into dead silence gets the room's tone, and the whole mix comes to -14 LUFS integrated with true
+ peaks under -1 dBTP, which is what the floor's loudness defect asks for. Don't level lines with `gainDb` or master
+ the voice outside raea. One `` per composition sets another target.
- Put the sounds in the tree in time order. The engine keys a sound's stretch by its place in the tree, so adding a
sound in the middle of the list can split the ones after it.
-- Balance relative to the voice. The approved pieces, mastered by `` to -14 LUFS, sit like this:
+- Balance relative to the voice. The approved pieces, mastered to -14 LUFS, sit like this (each a `gainDb` on top of
+ raea's level for the voice, or on the sound's own level for the rest):
| sound | with a voice | without one (promo, deck, game) |
|---|---|---|
@@ -110,7 +113,6 @@ ends up shouting over the narrator while another is inaudible.
- `loop` on an `