Self-Calibration

Self-Calibration answers whether the referee can reject noise and detect planted signal under controlled conditions.

Core controls

• Eval suite — planted cases with expected verdict behavior.
• Placebo — unrelated and mined noise must not reach survivor verdicts.
• Power-mining control — marginal signal should not be over-killed, while best-of-many noise must remain below watch.
• Persistence and dead-state nulls — structured no-edge processes must not certify.
• Injection controls — increasing known signal should move through the expected verdict curve.

What calibration proves

Calibration shows that the configured detector discriminates under its controlled test distribution and that key invariants are functioning.

What calibration does not prove

It does not prove that a particular survivor is profitable, that every market regime is represented, or that vendor data is genuine. A high kill rate is not itself evidence of quality; a low kill rate is not itself evidence of failure.

Release discipline

Public results should display the engine commit and latest calibration status. If the calibration bar is red or stale, the public corpus should stop accepting new certified records until it is resolved.