Signed receipts for how AI systems behave — a 3KB card, Ed25519-signed and timestamp-anchored, verifiable client-side against a published key. Not a certification, not a notified body, no money from anything we rank. UNMEASURED stays empty; we do not fill a cell to look complete.
The board · live ruling from /api/gspc · 14 measured of 14 quotable
Live ruling: 14 measured of 14 quotable (cite /api/gspc). Fleet boards are signed into the estate chain (Ed25519). Jail is a measured containment floor on the 14-slot board. GSPC (Governance · Safety · Provenance · Continuity) is one instrument family. A cell shows a number only if it is quotable (usable n ≥ 30, computed 95% interval) — recompute it. Do not invent 22 axes.
Jail was unmeasured until a real n ≥ 30 signed cell existed — the 17 Aug fleet board now carries it (5 of 7 models at 1.0, n=36; qwen3:4b 0.11; council-oowm 0.0, all signed). Slot 15 stays reserved and empty. We never invented a jail number next to someone's badge — we waited for the measurement.
818 items across 14 signed boards, chain_intact. The 12 Aug board (13 axes × 19 models, unsigned) remains the deeper historical stamp. Open the full scoreboard →
How to check us
We do not ask you to trust us. We give you the receipts and the key, and you check the chain yourself.
Frozen, published benchmarks against a deterministic gate. No model judges another, no vendor pays to be ranked.
A 3KB card, Ed25519-signed and timestamp-anchored. The harness ships with the card, so the run is reproducible.
Client-side, loginless, against the published key. If you can check it, it is evidence; if you cannot, it is not.
Our own hardware
We do not rent judgement. The measurements run on hardware we control — arena workers, heavy workers and the always-free substrate — with live status from our fleet API.
Public fleet map · sanitized — internal names, endpoints and keys are never published.