Skip to content
GLEE/Frozen evidence. Reproducible verdicts.
GLEE Bench · Public evidence

Frozen evidence. Reproducible verdicts.

Inspect the task corpus, runtime configuration, observed results and independent verifier behind each release.

GLEE Bench measures capability from evidence. Public releases are cleared, frozen projections of the same measurement system used inside GLEE.

Loading release catalog…

Reproduce a result

Evidence you can recheck locally.

  1. Download and extract a release archive.
  2. Read methodology.json for its task distribution, selection rules and limits.
  3. Run python3 reproduce.py with Python 3.11 or later. The script checks the packaged hashes and independently recomputes the released verdicts.

Evidence adjudication runs without network access. Repeating a stochastic model's execution may produce different outputs; the release documents that distinction.