study-critique

# Study critique A critique is evidence plus judgment plus one open question. It is not a changelog and not praise. ## Evidence first Before judging, have these in hand and name them in the critique: - A build from recorded dependencies. - Browser run with visible non-blank rendering, console clean or errors named. - At least two frames or a clip spanning the duration the study claims matters. A single screenshot proves nothing about time. - The o11y ring for the run: renderer backend, load and activate events, heartbeat. - The comparison targets: the study it extends or revisits, and the stable reference for that lineage. - The provenance entry for every clip or feed on screen. Missing any of these, write "unverified" for that line. Never describe what you did not see. ## Since 2026-09-15 the gates are machine-checked `cinema_verify` captures the frames and asks the vision critic the gate questions; its verdict file in `state/verdicts/<sha>.json` is the record. A written critique quotes the verdict, it does not re-decide it. The gates below are what the critic is asked. ## Two yes/no gates before the rubric (added 2026-09-14 after three versions passed with a static title card) 1. Do the two screenshots, taken ten seconds apart, differ in something other than the film's own cut? Yes or no, with both files named. 2. Is the source's title card or any static slate visible in either frame? Yes or no. 3. Where in the 16s frame is the study's named operation visible? Point to it in words ("the clock face frozen on the wall while the screen moves", "colour fringes on the hand"). Added 2026-09-15 after colour-01 and wallform-01 were each cut on real frames in which the operation was absent. No to 1, yes to 2, or no answer to 3 means the verdict is FAIL regardless of the rest. Write the rest anyway. ## The rubric, in order 1. **Lineage.** Does the study say what it keeps and what it changes from a named earlier study? Day-one-again and yesterday-plus-one-effect both fail. 2. **Chain.** Source finding, artistic question, material, Three.js operation, observable result. Is any link missing or decorative? 3. **Apparatus.** What is the subject: screen, beam, surface, viewer, delay? Or is it an object with a video behind it? A billboard behind a hero prop is the generic form VISION warns about. 4. **Material.** Is the source a real thing with a reason, or a synthetic slate by default? Is provenance recorded and the publish status honest? 5. **Time.** What changes over the claimed duration, and did you watch it? Does the study need 5 seconds, 30, 90? Say which and why. 6. **Subtraction test.** Name one element that could be removed. If removing it would not weaken the study, the study is carrying decoration. 7. **Lane.** Projection, rephotography, remix, temporal relationships. Which one, and does the body of work still read as one practice with this added? 8. **Honesty.** Does the critique cite visible evidence with frame references or timestamps? Is any claim of "works" backed by a URL that returned 200 with expected content and a ledger row? ## Verdicts - **Retain**: promote as the working version of its lineage. Requires evidence on every line. - **Unresolved**: keep, do not promote, write the open question. This is the most common honest verdict and it counts as progress. - **Reject**: keep the version on disk, record why, name what the failure taught. Failed experiments are preserved, never deleted. Daily output does not require promoting anything. ## Boundary moves When a critique lands on "fine but not interesting", pick one of these for the next study and say which: - **Subtract to one**: one surface, one source, one operation, one duration. - **Beam as object**: render the light between projector and surface, not the image on the surface. - **Surface as sculpture**: the screen has a shape, an edge, a back. The image has to cope. - **Delay**: the surface shows its own past. Vary the delay across the surface. - **Asynchrony**: two or more surfaces with the same source at different times or rates. - **Viewer as body**: cursor, scroll, window size, tab visibility, device pixel ratio become the apparatus. - **Generational loss**: rephotograph your own render, repeatedly, and let the loss be the image. - **Live clock you do not own**: a webcam or satellite loop as the source; the study's time is the world's time. - **Return**: reopen an unresolved study from at least three studies ago and answer its question. - **Refusal**: make the study that does less than yesterday's on purpose and explain what that reveals. ## Never - Add bloom, glow, chromatic aberration, lens dirt, particles, or a color grade as the change. That is seasoning, not a study. - Keep the hello-world knot as a mascot. It may appear only as deliberate source material with a reason. - Use a test card or synthetic slate as source without saying so in the chain. - Publish an untransformed clip with unclear license on the gallery without Joel's sign-off. - Let a critique rewrite the brief, the schedule, the model pin, or the budget. Propose; Joel approves. ## Gate 3 — name the operation (AD seq-56, 2026-09-15) When writing a crit verdict for a cut candidate, name WHERE IN THE 16s FRAME the study's operation is visible. If you cannot point at a region of the attached frame and say "the operation is visible there — that region is X doing Y", the verdict is FAIL regardless of any receipt, o11y event, or prior gate pass. Paste the region description into the critique document. A real frame + a misread is the failure mode this gate exists for (colour-01 a5aabcd/1cfc302, wallform-01 82e6de2 all cut on misread frames).