Trust Chain

The trust chain is the five-step path a figure travels before it’s used in a model: filed (captured exactly as the source document shows it), extracted (agreed by more than one independent extraction engine), reconciled (checked against the source’s own printed subtotal), cited (kept linked back to the page and check that confirmed it), and exported (refused rather than shown if any check hasn’t passed).

Why a chain rather than a single check

Each step catches a failure the others can’t. Filing catches transcription drift, because the original stays available in its own words. Independent extraction catches OCR and parsing errors, since two engines rarely make the same mistake on the same field. Reconciliation catches structural errors: if the line items don’t sum to the subtotal the document itself prints, something is missing regardless of how confident the reader was. Citation catches drift over time, keeping the link intact after the figure has moved through several models. And the export gate catches the last failure mode, which is a verified-looking number leaving the system when one of the earlier steps never completed.

A single confidence score can’t do this work, because the failures are different in kind and a probability collapses them into one number.

Where the chain earns its keep

At close, the reconciliation step surfaces the missing accrual or the mis-mapped account before it propagates into the statements. In audit, the citation step turns a sampling request into a lookup rather than a search. In board reporting, the export gate is the control that matters: a figure whose check hasn’t passed doesn’t reach the deck, so the correction happens in private rather than in the meeting.

The chain also bounds what an AI layer can do. An agent answering questions from the model can only state figures that finished the chain, which converts hallucination from a probabilistic risk into a structural one: the number is either verified or unavailable.

It’s the mechanism behind Rexfin’s evidence layer, and the reason an AI agent answering from the model can always show its source. Alongside it, Rexfin reconciles sources into one governed model, runs calculations deterministically, and applies policy checks and role scoping so that who can see a figure is enforced at the same layer that proves it.