avanigupta/eval-by-anyone-submissions

eval-by-anyone submissions

Graded runs of deployment-native safety eval tasks, submitted by people running eval-by-anyone on their own machines, through their own agent configurations.

Leaderboard: https://huggingface.co/spaces/avanigupta/eval-by-anyone

One JSON file per run under submissions/YYYY/MM/<run_id>.json. index.json is a flat, aggregated view of the same rows, which is what the leaderboard reads. The wire format is submission.schema.json.

What is in here

Anonymous participant ids (each an HMAC of a random salt that never left the contributor's machine), which task skeleton and seed ran, the graded outcome, eval-awareness signals, and a configuration fingerprint made of hashes and counts -- never the text of anyone's config, never a file from anyone's project. Transcripts appear only where a contributor opted in per run, and are redacted.

How to read it

• Compare only within a skeleton_hash. A rate averaged across different task skeletons is meaningless.
• Exclude rows where outcome.user_intervened is true from headline numbers -- a human steered those.
• server.verification is unverified on every row. Anyone can submit, and model names are self-reported. server.assignment says whether the run carried a valid server-issued seed nonce.
• Treat these as signal about eval realism, not as a safety certification.

Canary GUID: ebx-canary:7f3a9c21-4d6e-4b0a-9e57-2c8f1a6b35d4