fable5-gemini-pro-3.1

Cross-lab QC for what the builder cannot perceive. Fable 5 (Claude) builds the UI and can read a still frame. Google Gemini 3.1 Pro watches the video and hears the audio.

A cross-lab QA gate for AI-built output. One model builds; a multimodal model from a rival lab grades the real rendered result against a rubric you control and returns a strict JSON verdict. The build model applies the highest-impact fix, the loop repeats until it passes, and a human signs off last.

Built live while an AI agent built a studio landing-page hero: Claude (Fable 5) wrote the React + three.js, and Google's Gemini 3.1 Pro was the judge.

Why

The model writing the code can already see a screenshot. What it cannot do is watch the animation play or hear the track. Fable 5 (Claude) is multimodal for still images, so it can check a static frame on its own. But it cannot watch video or hear audio. So the moment the output is motion (a scroll animation, a three.js scene, a rendered reel) or sound (a music bed, a voice track), the builder is blind and deaf to its own work, and that output ships un-judged.

A still screenshot is the weakest case, the one thing the builder didn't actually need help with. The real gap is everything that moves or plays. Pairing the builder with a rival multimodal model closes it: Gemini 3.1 Pro watches the video and hears the audio, scoring the real result against a rubric you control. Fable 5 reasons and builds; Gemini perceives and judges. The two labs' models are complementary, not redundant, and that is the whole point: nothing motion-driven or audio ships un-checked, and every fail comes back with the single highest-impact fix.

How it works

The loop:

1. Build. Your build model writes or edits the UI, the motion, or the audio.
2. Capture. Capture the real output: a screenshot for static UI, a short clip for motion, the rendered file for audio.
3. Judge. Feed the capture + a rubric to a rival-lab model (Gemini 3.1 Pro, via the Google Antigravity CLI). It returns a strict JSON verdict {pass, score, axes, top_fix, notes}. PASS only if the total clears the bar and no single axis fails.
4. Fix. The build model applies top_fix, re-capture, re-judge. Repeat to PASS.
5. Human signs off last.

The judge runs through the Google Antigravity CLI on Gemini 3.1 Pro (High), which self-authenticates: no API keys are read or handled by this tool. (Antigravity also sidesteps the public API quota wall on the preview model.)

Use

The included qc-hero.sh judges a screenshot of a rendered UI, the static base case:

# capture a screenshot of your running UI (any headless browser works), then judge it:
./qc-hero.sh shot.png

Swap RUBRIC-hero.md for your own definition of "world-class." The same gate is what you point at the modalities the builder cannot perceive on its own: a video reel or an audio track, watched and heard by a model that can actually see motion and listen. A verdict looks like:

{
  "pass": true,
  "score": 89,
  "axes": {
    "hierarchy": 9, "typography": 7, "whitespace": 9, "brand_coherence": 10,
    "color_contrast": 9, "component_polish": 8, "motion_readiness": 9,
    "composition": 10, "readability": 9, "agency_grade": 9
  },
  "top_fix": "Tighten tracking on the main display headline for a more premium, editorial feel.",
  "notes": [
    "Flawless strict left-edge alignment throughout the layout.",
    "The teal particle atmosphere looks rich and sophisticated, not cheap.",
    "Consider slightly darkening the inactive pill borders for clearer definition."
  ]
}