Evidence interpretation
August 14, 2026 ยท View on GitHub
| Artifact | What it proves | What it does not prove |
|---|---|---|
config.json | frozen route/profile/limits for one run | provider availability |
journal + HEAD | ordered, tamper-evident durable controller events | external job correctness alone |
budget.jsonl | idempotent USD reservations, spend and release | market price beyond frozen schedule |
failure-pool.json | tasks frozen before candidate rewards | representative benchmark score |
external evaluator summary.json | normalized one-trial reward, usage and raw job path | official leaderboard result |
| proposal gateway receipts | locked request/response identity without model text | proposal quality |
| candidate build receipt | source/build/capsule content identities | improvement |
| crash-resume receipt | one launch/observation/commit after real process kill | all possible crash timings |
audit result | stable-demo invariants at the final state | sealed promotion or SOTA |
Per-run state is private and is not committed. A public release audit records hashes and aggregate facts without publishing credentials, concealed assignments, model reasoning, task trajectories or private provider responses.