Answers-object, records-array, generic JSONL, and batch JSONL formats normalize into one adapter.
AI RELIABILITY REGISTER · MODEL-RUN-0.9Test the answer.
Test the answer.
Keep the caveat.
Any model provider can submit the same 140 source-linked questions. Peplyst records the exact benchmark and evaluator versions, coverage, and deterministic pass rate without turning the result into a medical-quality claim.
Every run records the benchmark, adapter, and deterministic evaluator versions used.
A score is a regression signal—not medical validation, overall model quality, or a vendor endorsement.
PUBLIC SYSTEM SUMMARY
Comparable runs, explicit limits.
Scores are deterministic fact-coverage and prohibited-claim regression checks, not independent medical validation, overall model quality, or a vendor endorsement. Answer text and operator identity are excluded.
No public model runs yet.The adapter and benchmark packets are ready. A first external run requires a provider choice and credentials.
REPRODUCIBILITY CONTRACT
Three versions travel together.
The public register never publishes model answer text or operator identity. It exposes only aggregate outcomes and the metadata required to interpret them.
140 source-linked, timestamped test items.
Normalization, validation, duplicate detection, and payload hashing.
Required-fact coverage and prohibited-claim checks.