AI assurance · United Kingdom
A record of what your AI agents decided.
Every decision commits to the hash of the one before it. Delete one and the chain breaks, not just the count — and your auditor can prove that themselves, in their own browser, without an account and without asking us.
The verifier runs entirely in your browser. Nothing you paste into it is uploaded.
A quiet log looks exactly like good behaviour
When something goes wrong with an automated decision, somebody asks for the log. What arrives is a list of rows from a database that the organisation under scrutiny controls, in which any row could have been edited or removed before it was handed over.
Nobody rewrites an embarrassing decision to say something else. They delete it. And a log with a fortnight missing reads, to anyone opening it later, exactly like a fortnight in which the agents behaved.
Signing the export does not fix this. A signature proves the rows you were shown are unaltered. It says nothing whatsoever about the rows that were taken out before you looked.
Try breaking it
Record a few decisions, rewrite one, and see what happens. This runs entirely in your browser using the same published algorithm the platform uses — nothing is sent anywhere, and nothing here was prepared in advance.
- No decisions yet. Press Record a decision to start a chain.
Three moving parts
There is no machine learning in this product and nothing is inferred. It writes down what your agents tell it, in a form that cannot be quietly changed afterwards.
- 01
Your agent reports
A few lines of Python or TypeScript. The SDK never blocks your agent and never drops a decision quietly — anything it cannot deliver is surfaced as an error you can alert on, because silent loss is the failure this product exists to prevent.
- 02
The record is chained
Each decision is hashed together with the hash of the previous decision by the same agent. Records are append-only at the database, enforced by triggers rather than by application code — so the answer to "could someone alter this?" is no, not "we do not".
- 03
Anyone can check it
Export the decisions and re-derive the hashes with about thirty lines of your own code. The algorithm is published, and our own verifier runs in your browser so we never see what you are checking.
What it refuses to do
These are the decisions that make the record worth something. Each one costs a feature a competitor will happily sell you.
- It will not give you a compliance score
- A percentage needs a denominator of all obligations, and most obligations cannot be observed by a logging platform at all. Article 14 of the EU AI Act requires that a human can effectively oversee the system. We can count how many decisions carried a review; we cannot tell you whether the reviewer understood what they approved or could have stopped it. A green tick there would hand you documentary comfort about a control nobody has audited.
- It will not classify your system
- Under the AI Act, whether a system is high-risk turns on its intended purpose and deployment context. We can see neither. An auto-assigned risk tier would be a fabricated compliance artefact sitting in a record you would later rely on, so the field stays undeclared — and is reported as a live gap — until you fill it in.
- It will not stop recording when your card expires
- Ingest is not behind a plan tier and never will be. If recording stopped on a billing failure you would end up with a log missing a fortnight, and that gap reads exactly like a fortnight in which your agents behaved. Selling the prevention of that failure while causing it would be indefensible.
- It will not turn away an unregistered agent
- An unknown agent reference is registered on sight rather than refused, because a rejected decision is a dropped decision. It is marked undeclared and flagged on every single response until somebody names an owner and a risk tier.
Assessed against what your auditor will ask about
Every obligation is grouped by who can actually answer it. The count we cannot assess is shown at the same size as the count we can — those are the ones an auditor raises first, and they are the ones every other dashboard hides.
EU AI Act
Articles 12, 14, 19, 26(6) and 73 — record-keeping, human oversight, log retention and serious incidents. Regulation (EU) 2024/1689.
UK ATRS
The Algorithmic Transparency Recording Standard, Tier 1 and Tier 2 — the named owner, the technical specification, and the change control an incident review turns on.
UK GDPR
Article 22 on solely automated decisions and Article 30 on records of processing. We supply the count; whether Article 22 is engaged is a legal judgement about your use case and we do not make it.
ISO/IEC 42001
Evidence toward the AI system event-logging control area. Certification is granted by an accredited body assessing a management system, never by us.
NIST AI RMF
The MEASURE and MANAGE functions. The framework is voluntary and non-certifiable, so this evidences measurement rather than conformity — there is no conformity to claim.
Your own policies
Each decision can carry the policy or prompt version in force at the time, so a change in behaviour can be tied to the change that caused it. That is the question asked after every incident.
Do not take our word for any of this
A verification tool run by the company that stores the records proves very little. Paste any AgentAudit export into the verifier, alter a single character, and watch it tell you exactly which decision changed and at which sequence number. It runs in your browser; nothing is sent to us.
Open the verifier