Community evidence

Contribute to ACBench

  1. Choose the reference protocol and freeze a policy artifact. Record its hash, attribution, license, assistance and exact configuration.
  2. Prepare a suite with pnpm bench:reference --artifact policy.json --seeds /private/seeds.json --out /private/campaign --prepare. Evaluate with pnpm bench:reference --artifact policy.json --seeds /private/seeds.json --out /private/campaign --suite /private/campaign/suite.json --consent-publication --license LICENSE, which writes the consented publication JSON. This uses the generic PlayerThree scheduler and real AC baselines. Retain all failures; do not select only a best run. Trusted grading seeds remain operator-only.
  3. Send the publication JSON and immutable artifact to repository maintainers for review. Never include credentials, private prompts or owner chat. Public requests cannot execute arbitrary configuration modules.
  4. A trusted operator reruns the artifact in the policy sandbox, checks evidence and publication consent, then imports an immutable revision. Submitted claims stay labeled submitted; only operator-reviewed reruns receive trusted-rerun labels.

Artifact performance verification is not model-authorship verification. Assisted and mixed systems belong in community, not an official model-authoring claim. Corrections append a reasoned revision. Withdrawal blocks future delivery, but cannot recall downloads.

Catalogue metadata helps select exact provider/model IDs. It proves neither what executed nor the historical cost of a run.