The measurement harness

One card per model per axis. Each carries the model, the axis, the graded result, the issuer and a verify path, signed Ed25519 over canonical JSON.

Limit: A card records one graded run. It is not a judgement about the model overall, and it is not a compliance determination.

A recipient recomputes sha256 over the canonical body and checks the signature against the published key, with no call back to us. The verifier is open source.

Cards are linked: each names the previous card's identifier, and the manifest names the head. A card removed or reordered after the fact breaks the chain.

Limit: The chain proves internal order and integrity. It is not an external timestamp — see below.

A deterministic report that maps measured axes to the obligations they bear on, grades the gap on a fixed scale, and states the penalty tier the obligation sits in. No model is consulted to produce it.

Limit: The obligation map currently names a subset of the board's axes, and several of its names no longer match the board's. Where a name does not match, the report says UNMEASURED for an axis that has in fact been measured — it under-reports coverage rather than over-reporting it. Read it beside the live board until the map is re-pointed.

A live check of the published surfaces a result depends on — the board, the card index, the registers, the archive records — reporting each one's actual response.

Candidate benchmark datasets are screened on licence before anything is measured from them. Copyleft and non-commercial licences are rejected.

Limit: Licence is the only predicate implemented today. The canary-string and row-shape checks are described in the tool but are not yet performed, so a clean verdict means licence-clean and nothing more.

Cards are signed on the measurement node, where the key stays. The key is not in any repository and never leaves that machine.

Limit: There is no issuance on a laptop and no self-service issuance. A run is scheduled on the node; the artifacts and the verifier are what travel.

The intended behaviour is that a newly published model is detected and queued for measurement within days of release.

Limit: The recency filter in the current dispatcher does not function — every model the registry returns is admitted regardless of age — so the queue it produces is not the selective one the name implies. Treated as not available until the filter is repaired and re-tested.

The intended behaviour is a third-party timestamp proving a card existed at a given moment independently of us.

Limit: The stored anchor records a success that did not occur: the timestamp authority returned an error and the error was recorded as an ok status. There is no valid external timestamp on any card today. The card chain and the signature are unaffected — but neither of them is a proof of time.

A register of tokenised real-world assets exists, and each entry carries a signed card naming the asset and its public issuing address.

Limit: Every entry in that register reads UNMEASURED, because no question bank exists for these issuers yet. The cards verify as signed documents; they contain no assessment of the asset. Nothing in this register should be read as a rating.