A working build, filmed on a fictional company. Agents map evidence, pre-fill tests and draft findings. A named auditor decides every one.
In brief
Two parts of the build run on this page. Draw a sample or try to tamper with the log.
1 · The problem
Auditors spend most of an engagement matching evidence to requirements, checking samples item by item and writing findings up. Language models are good at exactly that kind of reading.
Internal audit also rests on things a model cannot hold: judgement, objectivity, and a signature someone can be held to. If an agent drafts a finding and an auditor approves it without reading, the audit looks like assurance and carries none of its weight.
I built this system to see whether agents can carry the reading without the auditor losing the decision. It runs internal audits against ISO/IEC 27001 and ISO/IEC 42001, and SOC 2 readiness reviews, on a fictional company.
2 · The rule
At every stage, an agent can only propose. A named person decides. The record keeps the two apart.
An audit committee, an external auditor or a regulator can then see which conclusions came from people, and how closely they checked the machine's work.
At every stage an agent proposes, a named person decides and the record keeps both.
1. Evidence
2. Mapping
3. Sampling
4. Testing
5. Findings
6. Grading
Drawn from the build shown in the film, on fictional data. Any agent can be switched off, and the manual path still completes the audit.
The build also measures whether people are checking. It records how long each proposal was on screen and how often reviewers override the agents. Someone who accepts every proposal in a few seconds is showing automation bias, and the head of internal audit can see it.
Agents do the reading. People make the calls. And the record shows which was which.
The closing line of the film
3 · Try it
Sampling is the one step with no agent at all. Code checks the population first, and a gap blocks the draw until an auditor records a reason.
The sample is drawn with a recorded seed. An external auditor can replay it and get exactly the same items.
The sample can be replayed by anyone who has the seed.
Population check: no tickets in June. The extract may be incomplete.
Seed 4271 drew CHG-0108, CHG-0803, CHG-0903, CHG-1005, CHG-1008, CHG-1107.
Illustrative tickets. Draw again with the same seed and you get the same rows; change the seed and you get a new sample. Untick the reason and the draw is refused, as it is in the build.
4 · Try it
Every action, by a person or an agent, goes into a log where each entry is sealed over the one before it.
The database refuses edits outright. If a record were changed anyway, verifying the chain would show exactly where.
Change one past event and every seal after it stops matching.
a757e428d2af185f35539ed61b584b99b2d776076313de70e3e83f997dd28a16724a2691619d8e81Five events, each sealed over the one before it. Edit any event, then verify.
Illustrative events, real SHA-256 computed in your browser. In the build the database also refuses the edit outright; the chain is the second line of defence, and it shows an auditor or a regulator that nothing was changed after the fact.
A working proof of concept, tested end to end. It has a guided demonstration, an evaluation harness for the agents, and a test that tries to tamper with the database directly. Every agent can be switched off, and the manual path still completes the audit.
It is not a certified audit tool, and it has not been used on a live engagement. It holds requirement identifiers only and never stores the text of the standards. Everything in the film and on this page comes from fictional data.
Who it is for
If you are being asked how AI will change your function, and you want to see a version where the auditor stays accountable, I share the working system on request.
I built this to test one thing: whether agents can carry the reading in an internal audit without the auditor losing the decision. On this build they can, as long as every proposal looks like a proposal and every decision carries a name. Try changing an event in the log above. That small refusal is the part I would want an audit committee to see first.
