Anatomy of a Regulator Dossier

The first-day test

Every records question about automated decisions reduces to one concrete test. An examiner selects one decision — one dispute, one alert disposition, one adverse-action recommendation — and asks to see it. Not the system documentation, not the aggregate accuracy: this decision. What could you hand across the table on the first day, complete enough that the follow-up questions are about the decision rather than about the gaps in the record? Call that document a dossier. Most institutions discover its contents backwards, assembling it under deadline from logs, tickets, and memory. The better order is forwards: know what complete looks like, and build the system so that every decision produces it as a byproduct of deciding. What follows is that anatomy. We build one of these — the Regulator Dossier in the Arcus platform, fifteen sections generated per run — and it serves below as the worked example. The anatomy itself is general — worth holding any tooling against, whoever built it.

Provenance, and a seal on the record

The dossier opens by answering what produced this with specificity: which system, which versions of every component that shaped the outcome — models, prompts, rule corpus, configuration — and which inputs, identified precisely enough to re-derive the decision’s starting conditions. Versions matter because systems change; a record that cannot say which configuration decided cannot explain why two similar cases a month apart diverged.

Then, before any content, the record protects itself: an integrity seal — cryptographic tamper-evidence over the dossier’s contents, so that what is reviewed later is demonstrably what was recorded at decision time. This is the least glamorous section and the one that makes the others worth reading. A record that could have been edited after the fact does not prove what happened; it proves what someone later preferred to have happened. The seal converts “trust our process” into “check the hash.”

The decision artifact and its timeline

The center of the dossier is the decision artifact: the reasoning itself, as a sequence of inspectable steps — what was classified, what evidence was assembled, which elements of the applicable rule were tested and with what result, what challenge was raised and how it was resolved. Each step carries its evidence citations inline, so every claim points at the record that grounds it. This only exists if the system reasons in steps; a single block of generated prose has no anatomy to display — the architectural point we make elsewhere and take as given here.

Two companions complete the center. The compliance timeline lays the decision against its procedural clock — when the matter arrived, when each obligation triggered, when each was met — because error-resolution and adverse-action regimes are as much about when as what. And citation provenance closes the loop on the inline citations: every referenced source — a regulation’s text, a policy version, a document in the case file — must be resolvable, to the exact version cited. An unresolvable citation is indistinguishable from a fabricated one, and in an LLM-era record that distinction is the whole game.

The map and the honesty section

A complete record then situates itself. Standards mapping ties each artifact in the dossier to the governance expectation it evidences — which sections speak to model risk management guidance, which to error-resolution requirements, which to the institution’s own policy — so an examiner working from a framework finds each expectation answered by name rather than by inference.

And then the section that reads strangest and matters most: known limitations, stated in the record itself. What the system could not verify, where evidence was thin, which steps ran degraded, what the reasoning is known not to cover — written into the dossier at decision time, not conceded later under questioning. Supervisory reviewers are professionally suspicious of records that contain no weaknesses; a dossier that names its own is making a credible claim that the rest of it means something. These mechanisms reduce and interrupt failure modes; they do not abolish them, and the record should say so about itself.

The human layer, and the posture line

Last, the record shows its people. Review and attestation captures the human-oversight surface: who reviewed the dossier, what they commented on, what changes they demanded, and their attestation locking the record — so it shows not just what the system concluded but who accepted it, and what they questioned on the way. Oversight that leaves no trace is indistinguishable from oversight that did not occur.

Two threads run through every section. First, a color-coded separation of kinds: every line is marked as evidence of record, regulation, AI analysis, or human decision, so no generated sentence can borrow the authority of the file and no human judgment can hide behind the machine’s. Second, the posture line: the system records a recommendation; the institution makes the determination. That sentence is what the attestation operationalizes — the difference between a tool that participates in accountability and one that quietly absorbs it. A dossier with all nine parts does not make a decision good. It makes it something more useful: examinable — on the first day, across the table, without a caveat.