Repository navigation
Videoform: rules in a taste line's shape, and where to start - #208
shinyobjectz wants to merge 1 commit into
Conversation
|
Checked against #207 sentence by sentence and task by task. Nothing from the videoform plan was lost. All 17 VF tasks are here with their file plans, and the only changes are the two you listed (mined rules in 1. Keep Why that matters: we measured where a check's time goes for #205. On 56 s exports it took 13 to 20 s, almost all of it four separate full decodes of the video (flashing, freezes, flash frames, blotching), about 3.7 s each. Running them in parallel didn't help: four together took 15.8 s against 3.7 s alone, because decoding already uses all 8 cores. Only decoding once helped, which is what #205 did: the same findings byte for byte in 7 to 8 s. A check runs after every render, so it's paid every round. Per-frame SigLIP on every check would cost more than the whole check does now. So the flags should come from cached columns (Fingerprint already caches them by content hash), and SigLIP should run only on shots that changed since the last check, as VF11 already says for videoform checks. Then the one check stays as fast as it is. 2. Gate on consistency, not on how many references there are (VF8b, FP15e). VF8b, from my plan, says that under five references nothing gates. I'd change that. The encoders are frozen: a videoform learns from how consistent its references are, not from training a model, and 3 to 5 very similar videos is enough. Some of the people we're talking to already edit to a formula by hand, so their references will agree tightly, and that agreement is the signal.
FP15e should also give rules a level from their strength, as it does for axes, so a weakly held mined rule is a warning or a note and can't fail a render on something another render won't move. 3. Rule precedence by origin (FP15d). "The nearer one wins" settles two rules on the same check at different levels. At the same level, I'd settle it by origin: a rule the person wrote, then one kept from a note, then one mined from references. That follows the floor's "the person wins". 4. An approved edit becomes a reference (FP15f, VF8c). FP15f routes every label into the fingerprint that judged the piece, which covers learning from notes and picks. What isn't written anywhere: when the person approves a cut made to a videoform, it should be added as a reference, read once like any other (VF3), so the videoform's ranges, shot kinds, structure and rules are remade from its rows (VF8c, in seconds). Then a videoform improves with every approved project, not only from corrections. This fits the three layers already in the plan, which is also why "frozen encoders" in point 2 isn't a limit:
Approved edits feed the first layer, and their notes and picks feed the second. |
Controllers are gone (projects.feature PJ9, #209), so a videoform's mined rules go in its own `rules`, each in a taste line's shape (words, a check, its values, a level), and `raea check PROJECT` holds a piece to the rules of the fingerprint that judges it (VF8, a 15-line hook in apps/direction check.rs). docs/videoform.feature gains a short "Start here": the order (VF1, VF2, VF3, then VF4 with VF12 as the gate before VF1b and VF9-VF11), the new crates, the three hooks into what exists and the bench. docs/fingerprint.md records decision 6: Pro and Studio's monthly pool of clef checks and content intelligence.
285e208 to
f68805b
Compare
This PR is now a small change on top of the videoform plan as Zach wrote it (#207, merged).
The six-way merge this PR used to hold is gone: no rules.ts, no moving taste files, no folding raea judge in. It's kept on
fingerprint-one-fullin case it's wanted later.What changes
rules. Each rule has a taste line's shape: words, a check, its values and a level.raea check PROJECTholds a piece to the rules of the fingerprint that judges it (which), beside the person's, the workspace's and the project's taste. It's a 15-line hook inapps/direction/src/check.rs.videoformfield,raea context,raea check);It's docs only: 3 files, 40 lines added and 14 removed.