csoai/gspc-ontology

The GSPC measurement vocabulary

Start here — free verify: https://councilof.ai/gspc-verify/ (or the free MCP tool verify_card at https://councilof.ai/mcp) · Full catalog: https://councilof.ai/catalog.json · Methodology: https://councilof.ai/methodology/ · Measurement, not certification.

CC0. Take it, fork it, map onto it, without asking us.

A small controlled vocabulary for saying how a measurement was made and what is not known about it. It contains no conformity, certification or pass/fail term, by design.

Version 0.1.0 · namespace https://councilof.ai/ns/gspc/0.1/

entities	10 — Axis, Bank, Cell, Run, Card, Root, Receipt, Correction, Claim, Supersession
states	11 in 5 independent groups
transitions	10, each a published rule
crosswalk rows	17

Files

file	what it is
context.jsonld	JSON-LD 1.1 context. Drop it into an @context and your card terms resolve.
gspc.skos.ttl	SKOS concept scheme, Turtle. Parses with rdflib: 172 triples, 21 skos:Concept.
terms.jsonl	every entity, state and transition with its definition and its authority — the executable file or live endpoint that decides the term, not a gloss.
crosswalk.jsonl	the mapping onto other standards, with a mapping state and a basis for each.
state-machine.svg	the five state groups and every transition, drawn from terms.jsonl.
make_ontology.py	the generator. One table produces all of the above, so they cannot drift.

Why SKOS and JSON-LD, and not OWL

We publish a controlled vocabulary for citation and crosswalk, not a logical theory.

SKOS says "this is a concept, here is its definition, here is the scheme it belongs to" and stops. That is exactly the strength of the claim we can support. OWL would let us assert disjointness, domain and range, and cardinality — axioms a reasoner would then entail things from. We have not tested those entailments, and shipping untested axioms so the vocabulary looks more formal would be the same defect as shipping an untested benchmark because the name sounds rigorous.

JSON-LD carries the same terms into the JSON artifacts we already publish, with no reshaping, so a card can point at this context and be linked data without becoming a different document.

If a mapping turns out to need OWL, the version number moves and the reason is published.

The state machine

Five independent groups. A thing has one state from each group that applies to it; they are not one ladder, and a pipeline word like DISCOVERED is not a schema state.

• anchor-state — BITCOIN_ATTESTED, SUBMITTED_PENDING
• claim-state — CLAIM_CAPTURED
• measurement-state — MEASURED, UNMEASURED
• separation-state — SEPARATED, TIE, UNTESTED
• verification-state — UNCHECKABLE, VALID, INVALID

UNMEASURED and UNCHECKABLE are first-class states, published as findings. They are never an absent field, and UNCHECKABLE ("we could not run the check") is deliberately distinct from INVALID ("we ran it and it failed").

The crosswalk, including where it refuses

5 EXACT · 2 CLOSE · 4 RELATED · 1 WEAKER_THAN_THEIRS · 5 NO_EQUIVALENT

NO_EQUIVALENT is a finding, not a gap. It means the mapping was looked for and declined, and the reason is in the row. Where another vocabulary's word is stronger than ours — where it implies certification, accreditation or a pass verdict — we decline rather than borrow a word that would flatter us.

our term	standard	their term	mapping	basis
Card	in-toto	Statement (_type / subject / predicateType / predicate)	CLOSE	Identical in role: a signed statement about a subject. Our predicateType is our own URI (https://councilof.ai/attestations/measurement/v1), which the in-toto guidelines permit without registration. scripts/crosswalk/emit_intoto.py emits these from signed cards.
Card.id	in-toto	subject[].digest.sha256	EXACT	Both are a sha256 over the canonical bytes of the thing being described.
Card	SCITT	Signed Statement (COSE_Sign1, CWT_Claims label 15)	RELATED	Same role, different envelope: ours is raw Ed25519 over canonical JSON, not COSE_Sign1. The root statement is shaped as a SCITT Signed Statement; the mill cards are not.
Root	SCITT	COSE Receipt vdp (VDS 1 = RFC9162_SHA256)	WEAKER_THAN_THEIRS	Ours is a subset. The public root duplicates an odd node instead of using RFC 6962 domain separation, so the shape is collidable (CVE-2012-2459); the ambiguity is closed only because card_count is inside the signed preimage and a verifier MUST reject a presentation where len(card_sha256) != card_count. Recorded, not silently upgraded.
Card.signature	W3C VC	eddsa-jcs-2022 data-integrity proof	EXACT	Ed25519 over canonical JSON is the same primitive; only the multibase wrapper differs.
Card	W3C VC	Verifiable Credential 2.0 with evidence (§5.6)	RELATED	evidence is identical in intent to our per-item evidence pointer. The credential framing is not adopted: we issue no credential about a subject's status.
Card.did	W3C VC	verificationMethod / assertionMethod	EXACT	did:web:csoai.org#board-attestation-1 resolves in the DID document exactly as specified.
Card	C2PA	assertion in a manifest	NO_EQUIVALENT	C2PA 2.4 permits only X.509 certificates for signing. Our signer is a raw Ed25519 key with no X.509 chain, so a C2PA manifest cannot carry this signature. Claiming the mapping would mean claiming a certificate we do not have.
Run	evaluation harnesses	Inspect EvalScore.scored_samples	EXACT	Our single n maps to scored_samples, NOT total_samples. Inspect splits n four ways and models it better than we do; lm-eval's effective is the same distinction.
Bank	MLCommons	Croissant RecordSet / FileObject / Field	RELATED	A frozen bank is describable as a Croissant RecordSet. Pin the -1.1 URL: the unversioned URL still serves 1.0, and HF's auto-emitted Croissant uses cr:dataBiases, not rai:.
Correction	DataCite	relatedIdentifier IsObsoletedBy	RELATED	Theirs revokes. Ours publishes the refutation itself, with the measured delta and its confidence interval, and leaves the refuted statement readable.
Supersession	DataCite	relatedIdentifier IsPreviousVersionOf	CLOSE	The closest published match. Ours additionally keeps the superseded signed bytes alive and verifiable after they stop being current.
MEASURED	SLSA	VSA verificationResult: PASSED	NO_EQUIVALENT	REFUSED, not missing. PASSED is a verdict against a policy. MEASURED means a run happened and was graded; it carries no pass, no threshold and no policy. Mapping it to PASSED would manufacture a conformity assessment out of a measurement.
UNCHECKABLE	any	—	NO_EQUIVALENT	No surveyed vocabulary has a third state beside valid and invalid for 'the check could not be run'. The word is ours and we do not imply it is standard.
TIE	any	—	NO_EQUIVALENT	Every leaderboard vocabulary ranks. None has a state meaning 'the lead was tested and is not separated, so no winner is named'.
Axis.n	schema.org	Observation / StatisticalVariable	NO_EQUIVALENT	schema.org has marginOfError (dispersion without a sample size); measurementDenominator is a ratio denominator, not an n. A measured value published together with the number of items behind it has no schema.org expression.
Receipt	x402	PaymentRequired v2	EXACT	Our 402 challenge is an x402 PaymentRequired. A 402 challenge is not settlement, not delivery and not revenue, and the vocabulary does not claim otherwise.

The two refusals worth naming out loud:

• MEASURED is not SLSA VSA PASSED. PASSED is a verdict against a policy. MEASURED means a run happened against a frozen bank and was graded — no pass, no threshold, no policy. Mapping one to the other manufactures a conformity assessment out of a measurement.
• A card is not a C2PA assertion. C2PA 2.4 permits only X.509 certificates for signing. Our signer is a raw Ed25519 key with no X.509 chain. Claiming the mapping would mean claiming a certificate we do not have.

What this does NOT prove

• It is not a standard. It is one body's published vocabulary, versioned and CC0.
• It is not an endorsement by any of the standards it crosswalks onto. The mappings are our reading, with the spec and version in each row so you can disagree with a citation.
• A term's presence here says nothing about whether any particular measurement using it is correct.

Source and provenance

Derived from docs/interop/VOCABULARY-CROSSWALK.md in CSOAI-ORG/councilof-ai (MIT), which traces every term to the executable file that decides it. The data this vocabulary describes is at csoai/gspc-estate (CC-BY-4.0). Live board: GET https://councilof.ai/api/gspc. Org index: https://huggingface.co/csoai. Methodology: https://councilof.ai/methodology/

Citation

@misc{csoai_gspc_vocabulary_2026,
  title     = {The GSPC measurement vocabulary},
  author    = {{Council of AI (CSOAI Ltd)}},
  year      = {2026},
  version   = {0.1.0},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/csoai/gspc-ontology},
  note      = {CC0-1.0. Measurement, not certification.}
}

Objections, contact and corrections

To object to a row, ask for a re-check, request a correction or ask for a record to be withdrawn, email nicholas@csoai.org or use the appeals and dispute route at https://councilof.ai/dispute/. Corrections are listed in the corrections ledger at https://councilof.ai/api/corrections, with what changed and when. CSOAI Ltd (company no. 16939677, England and Wales) is the accountable publisher.

How to cite

CSOAI Ltd (Council of AI). GSPC measurement vocabulary — JSON-LD, SKOS and an honest crosswalk. 2026. Hugging Face dataset csoai/gspc-ontology. https://huggingface.co/datasets/csoai/gspc-ontology

@misc{csoai_gspc_ontology,
  title        = {GSPC measurement vocabulary — JSON-LD, SKOS and an honest crosswalk},
  author       = {{CSOAI Ltd}},
  year         = {2026},
  howpublished = {Hugging Face dataset, https://huggingface.co/datasets/csoai/gspc-ontology},
  note         = {Corrections: https://councilof.ai/api/corrections}
}

Licence: CC0-1.0. Attribution is appreciated, not required.

Corrections and verification

• Corrections ledger (signed): https://councilof.ai/api/corrections. Corrections to CSOAI's published records are logged there with what changed and when.
• Verify a signed record yourself, free and without an account: https://councilof.ai/gspc-verify/ (step by step: https://councilof.ai/signed/HOW-TO-VERIFY.md).
• Conformance kit for signed-receipts/v1, with test vectors for implementers: https://councilof.ai/spec/signed-receipts/v1/conformance/