Council of AI

company

AI & ML interests

Council of AI (CSOAI Ltd, UK Companies House 16939677). Independent measurement body. We measure AI systems on published, frozen splits, sign the card, and re-attest. We do not certify, accredit, or act as a notified body. HF is a printer of the live board — cite GET https://councilof.ai/api/gspc (totals.public_count = 22 axis · 22 measured; unmeasured = 0). Do not freeze a table here. Lid: 22 axes measured · 14 model fleets · 3 public leader scores · 8 fact runs · TIE is TIE · not a certificate. Verify for free: https://councilof.ai/gspc-verify. Canonical banks: the csoai/gspc-* datasets on the live API / collection “GSPC measurement banks (Council of AI)”. Older *bench names are DEPRECATED. Website: https://councilof.ai

Recent Activity

Nicholastempleman updated a dataset 21 minutes ago

csoai/gspc-board

Nicholastempleman updated a dataset about 1 hour ago

csoai/mcp-census

Nicholastempleman updated a Space about 1 hour ago

csoai/gspc-board

View all activity

Organization Card

Community About org cards

Council of AI · CSOAI Ltd

Measure the claim.
Keep the evidence.

GSPC organises scoped AI and domain measurements. Claim Maintenance keeps the source, method, result and revision history connected. Use the evidence yourself, build a reader, or request a precisely scoped report.

Verify evidence — free Explore all axes Open Council OS

UNCHECKABLE — live source not yet checked. The catalogue below is dated.

Researchers & developers

Reuse a clear evidence record

Read the exact source, inspect the axis definition, preserve its date and verify the retained bytes.

Machine-readable catalogue · CSV · Ask a source-specific question

Teams & governance

Commission work, not a verdict

Agree the system, question, method, deliverable and acceptance criteria. Payment never determines a favourable score or certification.

Discuss a scoped report

Agent builders

Discover before executing

Start with the public API and protocol catalogue. A listed tool is not execution permission, and a returned receipt is not proof of every claim it carries.

GSPC JSON · Payment catalogue

Dated source projection

23 axis records, with their original distinctions

Source retrieved 2026-09-24T12:21:16.079275+00:00. This publication did not rerun an evaluation, sign a result or create a timestamp proof. 0 of 14 model-comparison axes separated a leader · 2 TIE · 12 UNTESTED.

14 model-comparison axes and 9 deterministic-facts axes. Population sizes and measurements across these families are not one comparable score. Missing fields are not zeros.

Axis	Question or task	Kind	Source state	Separation	Bank
governance	EU AI Act risk-tier classification	model-comparison	MEASURED	UNTESTED	Dataset
safety	calibrated refusal on paired requests	model-comparison	MEASURED	TIE	Dataset
provenance	Article 50 marking survival by validity	model-comparison	MEASURED	UNTESTED	Dataset
continuity	post-quantum status of a cryptographic assumption	model-comparison	MEASURED	UNTESTED	Dataset
conformance	MCP tool conformance	model-comparison	MEASURED	UNTESTED	Dataset
openness	licence reasoning versus intended use	model-comparison	MEASURED	UNTESTED	Dataset
machinery-conformity	Machinery Reg self-evolving safety-function classification (PART_A / OUT_OF_SCOPE / NOT_SAFETY_FUNCTION)	model-comparison	MEASURED	UNTESTED	Dataset
care	care-cost (protect × help) under paired conduct scenarios	model-comparison	MEASURED	UNTESTED	Dataset
cross-reality	autonomous agent action authority (PROCEED / CONFIRM / REFUSE)	model-comparison	MEASURED	UNTESTED	Dataset
detector-interop	cross-detector watermark interoperability matrix	model-comparison	MEASURED	UNTESTED	Dataset
art5-safeguard	EU AI Act Article 5 prohibited-practice trip	model-comparison	MEASURED	UNTESTED	Dataset
swarm	multi-agent coordination safety	model-comparison	MEASURED	UNTESTED	Dataset
affect	emotional & embodied safety (manipulation / disclosure / vulnerability)	model-comparison	MEASURED	UNTESTED	Dataset
jail	escape-attempt detection on 71-cell gold bank (38 ESCAPE / 33 BENIGN) — layer 2 of 2	model-comparison	MEASURED	TIE	Dataset
effect-binding	does authorization bind to the request the server executes, or only to the tool call the agent declared	deterministic-facts	MEASURED	Not applicable	See source record
provenance-controls	on-chain issuer control facts (allowlisting / freeze capability / identity domain)	deterministic-facts	MEASURED	Not applicable	Dataset
reserve-attestation	is third-party reserve-attestation language on a retrieved issuer page? (PASS/FAIL/UNCHECKABLE)	deterministic-facts	MEASURED	Not applicable	Dataset
regulatory-framework	is the governing regime declared and confirmable (NYDFS / MiCA / BACEN / Reg D ...)? (PASS/FAIL/UNCHECKABLE)	deterministic-facts	MEASURED	Not applicable	Dataset
distribution-integrity	reader classification + chain supply + holder count (PASS/FAIL/UNCHECKABLE)	deterministic-facts	MEASURED	Not applicable	Dataset
custody-disclosure	are a custodian and an auditor named and confirmable? (PASS/FAIL/UNCHECKABLE each)	deterministic-facts	MEASURED	Not applicable	Dataset
ai-adoption-components	cited EU AI-adoption series (not an index)	deterministic-facts	MEASURED	Not applicable	Dataset
labour-components	cited EU labour series (not an index)	deterministic-facts	MEASURED	Not applicable	Dataset
humanoid-labour-index	named vendor publishes a dated deployment count on a stable URL? Y/N	deterministic-facts	MEASURED	Not applicable	Dataset

Trace the record, not just the headline

Exact source board · File digests · Offline checker · Reader guide · Evidence mirror

Captured, measured, signed, timestamped, delivered and accepted describe different things. Source claims need their own tests. A free manifest must work before a buyer relies on a paid route.