Regulated AI needs more than a log file.
AI agents are already operating in financial services, healthcare, and legal workflows. When something goes wrong — a bad trade, a missed diagnostic flag, a contract clause the agent should not have touched — your organisation faces a question it currently cannot answer.
Existing audit logs are mutable. Post-hoc summaries are reconstructions. There is no standard way for a regulator, an insurer, or a counterparty to independently verify your account of what your AI actually did.
This is not a gap most teams notice until they are sitting across from a regulator or a claims adjuster. By then, the only evidence available is the operator's own logs — an inherent conflict of interest.
A cryptographic record that anyone can verify.
Cogitator wraps around an AI agent and produces a cryptographic witness root — a single tamper-evident fingerprint — after every run. Any third party can recompute that fingerprint from the recorded run and check it against a value they hold, detecting whether the committed record was altered after it was sealed.
Before any tool call is dispatched, Cogitator checks it against a declarative policy file. Allowed calls are executed and recorded. Blocked calls are also recorded — proving not just what the agent did, but what it tried to do and was prevented from doing.
The result is a self-contained, independently verifiable bundle: every agent step, every tool call, every policy decision, sealed cryptographically.
Every action taken, recorded in a tamper-evident chain.
Every blocked call recorded, not just permitted ones.
Policy decisions committed into the audit record, before dispatch.
Built for audit, procurement, and legal scrutiny.
- Regulatory evidence A concrete artifact satisfying EU AI Act Articles 12 and 9 tamper-evident record-keeping requirements, and aligned with FCA, SEC, and FDA SaMD audit trail expectations.
- Independent verification No need to trust operator infrastructure. Any auditor, regulator, or counterparty can rerun the verification themselves from the same inputs.
- Pre-call control Every tool call is checked against a declarative policy before any side effect — allowed, blocked, or observed — and each decision is recorded as part of the run.
- Byte-stable replay A recorded run can be re-executed from its transcript and checked for divergence down to the individual tool call, with deterministic inputs producing the same witness root.
- Liability clarity In the event of a dispute, both parties work from the same verifiable record. Not competing log extracts. Not operator-provided summaries.
Across the organisation and across the table.
Designed against real regulatory requirements.
| Framework | Requirement addressed |
|---|---|
| EU AI Act | Articles 12 and 9 — tamper-evident record-keeping and risk management systems for high-risk AI. |
| FCA / SEC | Mandatory audit trails for high-autonomy AI systems acting in financial markets. |
| FDA SaMD | Software as a Medical Device audit trail guidance — verifiable record of AI decision steps. |
| IEC 61508 | Safety-critical systems — policy layer as a capability firewall with committed evidence of blocked commands. |
| IETF / SCITT | The underlying witness protocol has been submitted as an IETF Internet-Draft and is designed to be carried as a SCITT Signed Statement for transparent, timestamped registration. |
Start with evaluation. Deploy with confidence.
Cogitator is released under Apache 2.0. The protocol specification is also Apache 2.0 and free to implement independently.
PROTOCOL SPEC — Apache 2.0. Free to implement, cite, and build on.
Regulated deployment enquiries, procurement questions, and compliance assessment support are welcome directly.
george.g@tuta.iogithub.com/noctem-o/cogitator