How It Works

Reasoning and Evidence

How an automated decision shows its work: the reasoning it recorded, the precedents and rules it cited, the alternatives it scored, and the gaps it declares.

Recorded, not reconstructed

When an automated decision is opened for inspection, what is shown was written at the moment the decision was made. No language model runs at explanation time, and the module that assembles explanations contains no inference of any kind. This is the single most important property on this page, and it is a constraint rather than a limitation.

The alternative is available and would be easy. A model asked, at the moment a human opens a record, why a decision was made would produce something fluent, specific, and usually plausible. It would not, however, be a recollection — the model was not there. It would be a reconstruction: an account assembled after the fact to fit the outcome.

In an audit trail that distinction is the whole game. A confidence the system did not actually hold, or an alternative it never actually scored, is a story about a decision rather than a record of one. The two are indistinguishable to a reader and completely different in what they are worth. So an explanation here is a read of stored fields, and if a field was never stored, the explanation says so.

What an explanation contains

Reasoning

The account recorded at decision time — the institutional context the decision was made in, in prose. Not generated later, and not present unless it was written then.

Citations

References to the governing rules the decision was evaluated against. Sourced only: a citation appears because a rule was recorded, never because a rule seemed applicable in hindsight.

Confidence

The score held at decision time and the band it falls in, with an explicit statement of its relationship to the 85% auto-apply threshold — at or above it, applied; below it, routed for review. See Confidence.

Alternatives considered

The other categories that were scored and rejected, with their scores. This is the field most often faked in systems of this kind, and the reason it is trustworthy here is narrow and mechanical: it is only ever populated by the scoring pass itself.

Evidence

The specific inputs the score was built from. See below.

Evidence

Evidence is the set of facts a decision actually rested on. Each item names its kind, so a reader can see at a glance whether a decision was carried by history or by a guess about a vendor's name:

Precedent — a prior decision by this organization on this vendor. These are the only evidence items that link anywhere: each resolves to the recorded decision it refers to, so a reader can follow the chain of reasoning backwards through the organization's own history.
Vendor history — the count and shape of prior encounters with this counterparty.
Amount, date, category, actor — the transaction facts in play.

Evidence is descriptive of what was used. It is not a document store, and this is where a reader's expectations most need correcting.

Declared absence

Where a field is missing, the explanation renders the gap explicitly rather than omitting the section. The reasoning is the same as everywhere else in this system, and it is worth making general because it explains the design more broadly than any single feature does:

A view that quietly drops empty sections is not neutral. It produces a record in which “no receipt was attached” and “every transaction has a receipt” look identical, because absence renders as nothing and completeness also renders as nothing. The reader cannot distinguish them, and will assume the flattering one. Rendering the gap is what makes the presence of everything else meaningful.

This documentation follows the same rule, which is why sections like the one below exist rather than being tactfully left out.

No receipts model exists

Declared absence

There is no receipts table, no documents table, and no file attachment anywhere in the system. Evidence items link only to other recorded decisions inside the application; every other kind of evidence links nowhere, because there is nowhere for it to link to.

So there are no evidence links to receipts, invoices, or bank transfer records. An organization's own policy text may well require that receipts be attached for expenditures over some amount — such rules are common, and OpenBooks will record and evaluate the rule — but the attachment itself has nowhere to live, and the requirement is not enforced by anything holding a document.

This is the largest gap between what these pages describe and what an auditor arriving with an evidence request would want. It is stated here rather than left to be discovered.