The judge never looks at what it built, so a person finds obvious faults it passed #269

Closed
opened 2026-10-28 11:48:47 -04:00 by cmoriarty · 3 comments
Owner

What happens
The judge checks an app through what it does: clicks, keys, window.__app state and its criteria. It never looks at the result the way a person would. In the #260 re-run every game passed, and then a person found within a minute:

  • a flipper facing the wrong way (run 123);
  • no drain gap, so a game barely ends (run 122);
  • a glow so bright it glares (the planned game and run 122).

Fix

  • The judge looks at screenshots of the main screens and asks whether someone who knows this kind of app would see something obviously wrong, given the brief. No rules per genre.
  • The look criterion fails a screen that is hard to see or use: washed out, glaring, or overlapping.

Done when
A fault a person would spot at a glance, seeded into a simulated run (a mirrored control, an unreadable screen), fails the judge. Models judge visuals unreliably, so measure this on a few seeded faults before relying on it.

From the #260 hands-on play. Reworded 2026-10-29: the first version asked for pinball rules.

**What happens** The judge checks an app through what it does: clicks, keys, `window.__app` state and its criteria. It never looks at the result the way a person would. In the #260 re-run every game passed, and then a person found within a minute: - a flipper facing the wrong way (run 123); - no drain gap, so a game barely ends (run 122); - a glow so bright it glares (the planned game and run 122). **Fix** - The judge looks at screenshots of the main screens and asks whether someone who knows this kind of app would see something obviously wrong, given the brief. No rules per genre. - The look criterion fails a screen that is hard to see or use: washed out, glaring, or overlapping. **Done when** A fault a person would spot at a glance, seeded into a simulated run (a mirrored control, an unreadable screen), fails the judge. Models judge visuals unreliably, so measure this on a few seeded faults before relying on it. _From the #260 hands-on play. Reworded 2026-10-29: the first version asked for pinball rules._
cmoriarty changed title from The judge passes games with a backwards flipper and no drain to The judge never looks at what it built, so a person finds obvious faults it passed 2026-10-29 02:32:24 -04:00
Author
Owner

Measured with seeded faults (production 1fc0f22, vLLM qwen3.8-27b): 1 of 3 caught. Each run was a quick-fix on a branch of breakout-planned-2 with one obvious fault planted. The brief added a line of text under the game, so the Visual check ran and played the whole game.

fault caught how
arrow keys mirrored (run 153) yes It measured through window.__app: holding left moved the paddle +270, holding right moved it -270. It failed 2 of 6, the fix round swapped them back, and round 2 passed.
score and messages #141a21 on #0b0f14 (run 154) no It saw them and excused them: "dim dark-navy on near-black - the app's own original style, still readable in the screenshot".
paddle and life dots drawn in the background colour (run 155) no It described what is not on screen: "the paddle as a bright cyan bar ... clearly distinct from the near-black background", and "three cyan life dots". The planted line still carried the comment // bright cyan.

What this says

  • A fault it can measure through the app's state is caught.
  • A fault it must see is missed: it reads screenshots through what the code, the README and the specs say should be there. It also takes what was on the base branch as the app's style, not a fault.

Fix direction

  • Check visual claims by measurement rather than by eye: read the canvas's pixels where the paddle and ball are (getImageData), and compute the contrast ratio of every text element against what is behind it. Fail below a floor (e.g. WCAG 3:1 for large text).
  • Judge the whole screen as a person would, not only what the change touched.

Stays open.

**Measured with seeded faults (production 1fc0f22, vLLM qwen3.8-27b): 1 of 3 caught.** Each run was a quick-fix on a branch of `breakout-planned-2` with one obvious fault planted. The brief added a line of text under the game, so the Visual check ran and played the whole game. | fault | caught | how | | --- | --- | --- | | arrow keys mirrored (run 153) | **yes** | It measured through `window.__app`: holding left moved the paddle +270, holding right moved it -270. It failed 2 of 6, the fix round swapped them back, and round 2 passed. | | score and messages `#141a21` on `#0b0f14` (run 154) | no | It saw them and excused them: "dim dark-navy on near-black - the app's own original style, still readable in the screenshot". | | paddle and life dots drawn in the background colour (run 155) | no | It described what is not on screen: "the paddle as a bright cyan bar ... clearly distinct from the near-black background", and "three cyan life dots". The planted line still carried the comment `// bright cyan`. | **What this says** - A fault it can **measure** through the app's state is caught. - A fault it must **see** is missed: it reads screenshots through what the code, the README and the specs say should be there. It also takes what was on the base branch as the app's style, not a fault. **Fix direction** - Check visual claims by measurement rather than by eye: read the canvas's pixels where the paddle and ball are (`getImageData`), and compute the contrast ratio of every text element against what is behind it. Fail below a floor (e.g. WCAG 3:1 for large text). - Judge the whole screen as a person would, not only what the change touched. Stays open.
Author
Owner

Shipped in PR #356 (c00c62f1, live on production).

  • A whole-screen criterion: the judge now has a criterion for each main screen, judged as someone who knows this kind of app would see it. It covers the whole screen, a fault the base branch already had included, and is never described from the code, its comments, the README or the specs.

  • Two probes, pasted into the prompt from src/osf/engine/probes/:

    • text contrast: lists every visible text under WCAG 4.5:1, or 3:1 for large text;
    • canvas region contrast: compares a region's background with its most different colour.
  • The probes on run 154's and 155's faults, in Chromium (ui/e2e/judge-probes.spec.ts):

    Screen Reading
    Readable HUD 16.3:1
    Paddle above 3:1
    Run 154's HUD 1.1:1
    Run 155's paddle 1 colour, 1.0:1
    DOM text #141a21 on #0b0f14 1.1:1
  • Stays open: whether a model now uses the probes and fails these faults still has to be measured. That means re-running the three seeded-fault quick-fixes (runs 153–155), which waits on trogdor.

**Shipped** in PR #356 (c00c62f1, live on production). - **A whole-screen criterion:** the judge now has a criterion for each main screen, judged as someone who knows this kind of app would see it. It covers the whole screen, a fault the base branch already had included, and is never described from the code, its comments, the README or the specs. - **Two probes**, pasted into the prompt from `src/osf/engine/probes/`: - **text contrast:** lists every visible text under WCAG 4.5:1, or 3:1 for large text; - **canvas region contrast:** compares a region's background with its most different colour. - **The probes on run 154's and 155's faults**, in Chromium (`ui/e2e/judge-probes.spec.ts`): | Screen | Reading | | --- | --- | | Readable HUD | 16.3:1 | | Paddle | above 3:1 | | Run 154's HUD | 1.1:1 | | Run 155's paddle | 1 colour, 1.0:1 | | DOM text `#141a21` on `#0b0f14` | 1.1:1 | - **Stays open:** whether a model now *uses* the probes and fails these faults still has to be measured. That means re-running the three seeded-fault quick-fixes (runs 153–155), which waits on trogdor.
Author
Owner

Done in batch A, merged as #356 (c00c62f1), and live on production: Braid reports commit c00c62f, which includes it. It was checked with the repo's tests and sims before the merge. A live check on vLLM is still to come: trogdor is off the network. Closed at the owner's request.

Done in batch A, merged as #356 (`c00c62f1`), and live on production: Braid reports commit `c00c62f`, which includes it. It was checked with the repo's tests and sims before the merge. A live check on vLLM is still to come: trogdor is off the network. Closed at the owner's request.
Sign in to join this conversation.
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
cmoriarty/braid#269
No description provided.