Criteria

​

We list real criteria, not invented ones.

​

MediaChecker SafeGuard does not invent confidence theatre. Each displayed band maps to named evidence rules: what must be true, and what must not be true.

​

How a result is built

​

1. Pixel stages examine the file: compression / ELA-style texture, soft-capture / shaking-hand, phase labels, and an optional cloak proxy.

2. Each stage writes an audit containing inputs, outputs, confidence, and known failure modes.

3. Parallel probabilities compete across possible origins: camera, game, CGI, AI generation, and composite.

4. The label is the strongest supported arm, not a vibe.

5. Displayed confidence is tied to the earned bands below, not a marketing dial.

​

Cloak signals and “Liquid” soft-capture signalssuch as shaking hand, glass, or bloom—stay beside the main score. They are not secretly folded into one ego number.

​

Confidence bands: what must have happened

​

Tribunal rule

​

If the tribunal returns insufficient evidence, the UI band stays at or below 39%, even when one arm edges the others.

​

Cloak rule

​

We may note a possible adversarial texture proxy only when both probability and confidence clear the published thresholds. We do not call it “confirmed Glaze or confirmed Nightshade” without a real labelled corpus.

​

What we refuse to say

​

- A post about AI is therefore AI media.

- Shaking-hand blur, dirty glass, or lens bloom equal synthetic media.

- A pretty percentage is calibrated truth without the table above.

​

The criteria are a public explanation of how the beta presents evidence. They are not a promise that a heuristic result can establish authorship, intent, or provenance by itself.

​

[How it works]

​

[Install the Chrome beta — placeholder link]

​