Unredact ranks PDF redaction guesses

This is less a redaction breaker than a probabilistic document-forensics engine. That makes it genuinely interesting for exploratory analysis, but also exactly the kind of tool that should come with a big yellow caution sign.

• –The pipeline is the whole story: OCR, redaction detection, font inference, width constraints, then LLM plausibility scoring.
• –Because it ranks candidates instead of asserting answers, it’s useful when you already have context and want narrowing, not certainty.
• –The browser-only design is a nice trust signal for a tool this sensitive, but it still depends on a Claude API key for scoring.
• –The project’s own disclaimer matters: it is not meant for legal, journalistic, or law-enforcement use, which tells you how shaky the output can be.
• –It fits a broader trend in AI tooling: use LLMs to amplify weak signals in messy documents, not to pretend uncertainty has disappeared.