EVIDENCE JSON
Proof you can inspect.
One week after each completed test, we publish a JSON bundle with challenge questions, exact prompts, recorded responses, judge reasons, receipts, and explicit score-row status. Verified catches, passes, and unresolved tasks are separate fields; unresolved is not counted as a catch or a pass.
- Live list
The leaderboard page shows pending and available bundles with a live release countdown.
- Archive endpoint
/v1/benchmarks/evidence-archiveslists public bundles and their release times. - Download endpoint
/v1/benchmarks/evidence-archives/{run_id}.jsondownloads an available bundle.
GET /v1/benchmarks/evidence-archives
What GavelTM makes public
- Evidence bundle
Frozen prompts, recorded model replies, expected outputs, judge reasons, and receipt references are released on the public schedule.
- Score-row status
A complete score row has zero unresolved tasks and catches plus passes equal the denominator. Incomplete rows remain public for audit, but they are not certified leaderboard rows.
- Replay boundary
Hashes identify the canonical evidence object. A matching hash means the released JSON has not silently changed.
- Claim ceiling
The public claim is limited to this benchmark pack. GenXis does not expose private authoring machinery or use it as a substitute for public evidence.