Developer docs

API (live)

• GET /health — service status and endpoint list
• GET /leaderboard — measured GovBench 15-dimension leaderboard (JSON, with comparability caveats)
• GET /registry — benchmark registry: ids, axes, item counts
• POST /evaluate — evaluate a single prompt against the reference model

Recompute anything we publish

Every published figure ships with its harness: the AI Act frozen-split harness contains scenarios, the split function, and the scoring code. Clone, run, compare.

Frozen splits

A frozen split is a deterministic, pre-registered train/test partition (sha256 of scenario id). We publish results only on the held-out half, so no one — including us — can tune to the test set.