Turn AI-written code into a scored repair tournament.
Scout gives coding agents a local tool surface: find evidence, judge it, run competing patches, score the result, and export a receipt.
Fast is the default for live demos. Deep is for final proof when latency is acceptable.
7
seeded mistakes
3
specialist scouts
3
patch contenders
Hallucination Scout
Finds fake imports, impossible APIs, and nonexistent helpers.
Example: import exists nowhere
Spec Drift Scout
Finds comments, README claims, and names that lie about behavior.
Example: comment says redacted
Test Theater Scout
Finds tests that pass without proving meaningful behavior.
Example: test only checks called