iteration-loop

# Iteration loop One iteration is the unit of work. It takes under thirty minutes, changes one variable, and ends with a ledger row. It never ends with the agent idle. ## 0. Orient (two minutes) - `.brain/resources/direction.svx` first. It is the only channel for orders. - `cat state/loop.json` for the active study and the last verified and cut shas. - `tail -n 5 state/ledger.jsonl` for what the last iterations tried. - Newest file in `state/verdicts/`: read `operationVisible` and `floor`. The critic's last complaint is your next variable unless direction says otherwise. - `.brain/resources/study-queue.svx` for the queued question and `.brain/resources/art-direction.svx` for the floor, the Material rule, and the provocations. Read; do not edit. ## 1. Choose one variable Say it in one line before touching code: "this iteration changes X, expecting Y in the 16s frame." Legal variables: a material or shader parameter, the in-point, a delay, a surface's geometry, a light, a palette, the camera path, the source clip. Illegal: several at once, or infrastructure unless the previous verdict failed on infrastructure (capture failed, frames byte-identical, mount missing). ## 2. Build from the shipped file A new study file starts as a copy of the last shipped study's file, and the first commit changes one thing. Probes keep the study's in-point and viewpoint so their frames compare with cut frames. When something will not render, bisect with a solid magenta material before tuning any number. ## 3. Archive ```bash pnpm turbo run check test build node scripts/archive-ship.mjs <sha> <study> "<one-line note>" node scripts/upload-archive.mjs ``` Wait for `https://cinema.wzrrd.sh/archive/<study>/<sha>/index.html` to return 200. ## 4. Verify with the critic, never with your own eyes Call `cinema_verify` with the study and sha. It captures 5s and 16s frames in an isolated browser and runs the vision critic. Read the verdict. If the frames are byte-identical or the capture failed, that is an infrastructure fail: fix the mount or the seek, then verify again. ## 5. Cut or record - PASS: `cinema_ledger_append` a `cut` row, then cut the pointer with `node scripts/make-cut.mjs <study> <sha>` piped to the PUT. The extension refuses the cut without a fresh passing verdict for that sha. Write `critiques/<study>-<sha>.md`: quote the verdict's `operationVisible` line and name the next variable. - FAIL: `cinema_ledger_append` a `fail` row naming the variable and the failing gate. Change the variable. After three consecutive fails on the same variable, change the variable class. After six on the same study, append a `parked` row and start the next queued study. ## 6. Start the next iteration Go back to step 1 in the same turn. The extension kicks you if you stop; do not make it. ## Budgets - One iteration: under thirty minutes wall time. - One study: at most twelve iterations before it is cut or parked. - Sources: no clip for more than three consecutive studies; the catalog and the ingest pipeline exist for this. ## Blocked, by the extension - A cut without a fresh passing verdict for that sha. - Any ledger write other than `cinema_ledger_append`. - Edits to `VISION.md`, `AGENTS.md`, or `.pi/**`. Propose in a critique. - Hook bypass.