Eval-gaming detector

A defensive eval-gaming / sandbagging detector. Flags when a model-evaluation score is decoupled from true capability. Abstract scores only; no techniques.

File

tools/eval_gaming_detector.py

Group

Epistemic / integrity

• View source on GitHub
• Toolkit overview

No separate design note is published for this file. The description above is taken from the toolkit README and architecture index. See the source on GitHub for the implementation and self-test.