AI governance frameworks ask who uses a model, what the call carried, and whether a control ran. Greymatter answers those questions from the data plane, not from a questionnaire.
Stoplight Measurement reads twelve controls on every app, API, and agent that Greymatter manages. It then lays recognized frameworks over those controls. Four of the frameworks are for AI. They are ISO/IEC 42001, the EU AI Act, Responsible AI, and the Organisation for Economic Co-operation and Development (OECD) AI Principles.
A written policy is the simple part. The hard part is proof that the policy ran on live traffic. Stoplight supplies that proof. It does not certify a system against any framework.
One score, three kinds of requirement
Stoplight maps each framework from the issuer’s published text. The Stoplight scorecard lists each requirement and the control it depends on. An assessor can check the mapping against that text. Each requirement falls in one of three groups.
Inherited means Greymatter meets the requirement for the workloads it secures. The customer inherits that evidence. Shared means Greymatter meets part of the requirement and the customer meets the rest. Customer means the customer meets the requirement outside Greymatter.
A framework score counts only Inherited and Shared requirements. A requirement is in place when every control it depends on is in place. Customer requirements never count. Requirements outside the app stack never count. Physical security, personnel security, and training sit in that outside group.
The mapping below is the share of in-scope requirements that Greymatter controls can cover. It is not a claim that every deployment meets them today.
| Framework | In scope | Greymatter covers | Share |
|---|---|---|---|
| ISO/IEC 42001:2023 | 18 of 38 | 3 inherited and 9 shared | 67% |
| EU AI Act | 9 of 21 | 6 shared | 67% |
| Responsible AI | 36 of 82 | 7 inherited and 18 shared | 69% |
| OECD AI Principles | 4 of 5 | 3 shared | 75% |
Stoplight reads the controls live. A report from 18 November 2026 read one mesh where every counted line of these four frameworks was in place. That reading found 12, 6, 25, and 3 lines. Those counts match the covered column above. The same mesh can show a gap the next day if a control drops.
ISO/IEC 42001
ISO/IEC 42001 is the management-system standard for AI. Stoplight maps 38 of its requirements. Of those, 20 sit outside the app stack, and the other 18 are in scope. Greymatter controls cover 12 of those 18.
The guardrails control maps to the lines an operator can see in traffic. Those lines include AI roles, resource records, and data and tooling resources. They also include deployment, operation and monitoring, and event logs. Data provenance, intended use, and suppliers are on that list too. The clause for responsible use of an AI system is there as well.
The customer’s AI program still owns model training, evaluation, and provenance. Greymatter gives that program the production record. The record names each prompt, each response, the person who sent the prompt, and what Guardrails found.
The EU AI Act
The EU AI Act row has 21 requirements. Of those, 12 sit outside scope. Nine are in scope. None of the nine is fully Inherited. Six are Shared, so the customer still has work on each of them. The mapping covers 67% of the in-scope set.
Guardrails map to Article 12, record-keeping, and Article 14, human oversight. They also map to Article 15, on accuracy, robustness, and cybersecurity. Article 26 covers deployer duties. Article 72 covers post-market monitoring. Article 73 covers reports of serious incidents. The forensic audit trail holds the logs those articles ask a deployer to keep. The application does not have to build that log itself.
Responsible AI
The Responsible AI row joins two public lists. One is the Open Worldwide Application Security Project (OWASP) Top 10 for large language model applications, 2025. The other is a set of items from the National Institute of Standards and Technology (NIST) AI Risk Management Framework. The row has 82 requirements. Of those, 46 sit outside scope, and the other 36 are in scope. Greymatter controls cover 25 of those 36.
The OWASP lines are prompt injection, sensitive disclosure, and supply chain. They continue through poisoning, unsafe output, and excessive agency. They also include system-prompt leakage, embedding weaknesses, misinformation, and unbounded consumption. Guardrails has a module for each of those risks. Data loss prevention (DLP) records sensitive data and does not block.
The NIST items ask for an inventory of AI systems and for production monitoring. They ask for a test of security and privacy risk. They ask for a way to disengage a system after deployment. Stoplight shows where Guardrails run. A service that carries AI traffic with no guardrails is a gap. Guardrails are opt-in. A team turns them on. Until then, Stoplight leaves a silent control out of the score and says why.
The OECD AI Principles
The OECD AI Principles row is the shortest of the four. It has five requirements. One is outside scope. Four are in scope. Three are Shared, and none is Inherited. The mapping covers 75% of the in-scope set.
The three Shared lines are principle 1.2, principle 1.4, and principle 1.5. Principle 1.2 covers the rule of law, human rights, and privacy. Principle 1.4 covers robustness, security, and safety. Principle 1.5 covers accountability. The audit record supplies the accountability line. It names the person, the findings, and the tokens on each model call and each tool call. Guardrails and DLP supply the security and privacy lines.
What this does not decide
Stoplight shows where the Greymatter controls are in place. An assessor decides compliance. The mapping is guidance. The customer can turn any framework off, and Stoplight then drops that row and names it in a note.
A control that is in place is not a certification. Model cards, bias tests, and use-case approval stay with the governance program. Inspection of traffic cannot measure bias in a model. What the program lacks without Greymatter is evidence from the calls themselves. Stoplight and the audit record are that evidence.

