csoai-defoneos-mcp
By Nicholas Templeman, CSOAI Ltd (UK Companies House 16939677)
URL: https://councilof.ai
CSOAI — Council of AI blog posts
By Nicholas Templeman, CSOAI Ltd (UK Companies House 16939677) URL: https://councilof.ai
Post 1: 22 Axes, Signed
22 Axes, Signed: How CSOAI Measures AI Without Asking Permission
AI governance is stuck in certification theater. Audits take months. Seals expire. The moment the audit closes, the model changes. Nobody outside the room knows what was measured, how, or when.
CSOAI is a UK-registered measurement body (Companies House 16939677) that does it differently. Every AI we measure produces an Ed25519-signed, Bitcoin-anchored measurement card. The card is public. The signature is verifiable offline. The audit can never silently change.
The board at councilof.ai/board shows our latest 22-axis measurements. Each card carries:
- The axis measured (transparency, safety, fairness, accountability, and 18 more)
- The methodology used (FTS5 search, EUR-Lex corpus, prompt-injection probes)
- The measurement value at the time of signing
- The signature that proves nobody changed it since
The card is a receipt, not a certificate. Receipts are cheap to issue, cheap to verify, and impossible to forge.
Try it yourself: curl https://councilof.ai/api/gspc — every card, every signature, every axis.
Nicholas Templeman is the founder of CSOAI Ltd (UK 16939677). He measures AI so the rest of us can verify it.
Post 2: Stop Certifying AI, Start Receiving
Stop "Certifying" AI. Start Receiving It.
Every AI governance framework in 2026 still ends with the same ritual: an auditor writes a PDF, the vendor displays a badge, and the seal expires 12 months later. By month 13, the model has been retrained three times. The badge says nothing about today's behavior.
CSOAI's GSPC (Governance Signed Public Card) replaces the badge with a receipt:
- Measure the AI on a public axis (transparency, safety, fairness, etc.)
- Sign the measurement with Ed25519 — a one-way cryptographic signature
- Anchor the signature to Bitcoin so nobody can rewrite history
- Publish the card at
councilof.ai/api/gspc/{sha256} - Verify anywhere — offline, no account needed
The card doesn't say "this AI is safe forever." It says "this AI was measured on this date, on this axis, by this method, and here's the cryptographic proof."
If the AI changes tomorrow, the old card stays valid. If someone tampers with the card, the signature breaks. If you want to know what changed, you measure it again and issue a new card.
That's it. That's the whole system.
The board at councilof.ai/board shows 22 axes, hundreds of measurements, and a public-root Merkle tree you can verify yourself.
Nicholas Templeman, founder, CSOAI Ltd (UK 16939677).
Post 3: EU AI Act Article 50, Measured
EU AI Act Article 50, Measured: A 417-Provision Tool That Doesn't Phone Home
Article 50 of the EU AI Act requires AI-generated content to be "marked in a machine-readable way and detectable as artificially generated or manipulated." That is one sentence. It has 47 sub-paragraphs, 89 exceptions, and 281 cross-references to other articles.
CSOAI's eu-ai-act-compliance-mcp is a Python MCP server that ships 417 provisions from EUR-Lex, indexed in SQLite FTS5, searchable in milliseconds. No API calls. No data leaves your machine.
pip install eu-ai-act-compliance-mcp
# 417 provisions, FTS5 search, deadline tracker
Three tools:
risk_scan(system_description)— identifies which articles apply to your AIprovision_search(query)— FTS5 search across all 417 provisionsdeadline_tracker()— every compliance date from Article 113
Why measurement, not certification: CSOAI does not certify your AI. It measures which provisions apply, on what dates, and against what criteria. The measurement is signed. The signature is yours. The auditor can be anyone.
This is the same pattern CSOAI applies to all 22 governance axes. Measurement is portable. Certification is rented.
Nicholas Templeman, CSOAI Ltd (UK 16939677).
Post 4: Day Signed Receipt Beat Compliance Audit
The Day a Signed Receipt Beat a Compliance Audit
A European bank wanted to deploy a fraud-detection AI. The auditor required:
- An ISO 42001 statement of applicability (90 days)
- A SOC 2 Type II report (6 months)
- A vendor risk questionnaire (3 weeks)
That's 8 months of paperwork before one model goes live. The bank passed on the AI.
CSOAI's GSPC would have done it differently. On day one, we'd publish:
- A signed measurement card for transparency (the model is documented)
- A signed measurement card for safety (the model refuses jailbreaks)
- A signed measurement card for fairness (the model's false-positive rate is balanced across demographics)
- A signed measurement card for accountability (a log of every prediction lives at sha256
…)
Each card is Ed25519-signed. Each signature is verifiable offline. Each one says: on this date, this AI had these properties.
That's not certification. That's a receipt. Receipts don't expire. Receipts don't require a 90-day re-audit. Receipts are cheap to issue and cheap to verify.
The bank could deploy on day 7 — and re-measure on day 30, day 60, day 90 — and each new measurement is its own receipt.
The auditor can still ask for whatever they want. But the bank's lawyers have something better: signed, dated, public proof that the AI was measured, not just claimed.
That's the gap GSPC fills.
Nicholas Templeman, CSOAI Ltd (UK 16939677).
Post 5: 22-axis AI board in 7 lines of curl
A 22-Axis AI Measurement Board in 7 Lines of curl
The CSOAI board lives at councilof.ai/board. It shows every measurement we've issued. The whole thing — including signature verification — fits in a single HTTP request:
curl -s https://councilof.ai/api/gspc | jq '.cards | length'
That tells you how many signed cards exist. Want to verify one?
curl -s https://councilof.ai/api/gspc/sha256/<hash> | jq '{axis, methodology, measurement, signature}'
Want to check the signature?
curl -s https://councilof.ai/api/gspc/sha256/<hash>/verify
The board has 22 axes measured across transparency, safety, fairness, accountability, robustness, privacy, and 16 more. The latest cards, the historical cards, the methodology, and the cryptography — all public, all signed, all queryable.
No login. No paywall. No "contact sales for an enterprise demo."
This is what governance looks like when the artifact is the receipt, not the certificate.
Nicholas Templeman, CSOAI Ltd (UK 16939677).
Post 6: Why compliance is the wrong word for AI
Why "Compliance" is the Wrong Word for AI
The AI industry borrowed "compliance" from finance. Finance compliance is about process: did you follow the procedure? AI compliance got shoehorned into the same word, but AI is about behaviour: what does the system actually do?
Process is checked once a year. Behaviour changes daily.
CSOAI's term for what we do is measurement. Measurement has three properties compliance doesn't:
- Time-stamped. The card says "this AI had property X on date Y." Not "this AI is forever compliant."
- Reproducible. Anyone with the measurement code and the AI snapshot can re-derive the card.
- Signed. Ed25519 means the card cannot be altered without breaking the signature.
The financial industry uses the word "compliance" because their auditors check process. We use the word "measurement" because AI needs to check state.
A measurement card doesn't say "this AI is safe." It says "this AI's refusal rate was 99.7% on 2026-09-15T06:00:00Z, measured against a corpus of 12,847 jailbreak attempts, signed by CSOAI's verification key, anchored to Bitcoin block 870,123."
That's a much more honest statement than any seal on a PDF.
Nicholas Templeman, CSOAI Ltd (UK 16939677).
Post 7: Why MCP matters for AI governance
Why the MCP Protocol Matters for AI Governance
The Model Context Protocol (MCP) is the API standard that lets AI agents talk to tools. A governance body without MCP support is a body that can't reach the agents it claims to govern.
CSOAI ships 68 MCP servers on PyPI. Each one wraps a piece of regulation — EU AI Act provisions, GDPR articles, NIST AI RMF controls — and exposes them as tools an AI agent can call.
Example: an agent that's drafting a compliance report can call eu_ai_act.risk_scan(system_description) and get back the articles that apply. The call is measurement, not certification. The agent owns the output. The signature proves the corpus came from EUR-Lex, not from a vendor's blog.
MCP is the missing layer between AI governance (the policy) and AI agents (the practice). Without MCP, governance is a PDF the agent ignores. With MCP, governance is a tool the agent uses.
CSOAI's MCP estate is open-source, Ed25519-signed, and Bitcoin-anchored. The governance layer is now a runtime dependency.
Nicholas Templeman, CSOAI Ltd (UK 16939677).