A fintech hands the regulator the manifest
A credit-decision audit that points instead of narrating. Anonymized, illustrative.
A credit decision under review is not a story you narrate. It is a record you point at — and the record answers the questions the regulator would have asked.
This is an anonymized design-partner scenario; the institution is a fintech, the figures are illustrative, and no actor has agreed to anything. We are pre-revenue, building the POC. With that label on it, here is what a credit-decision audit looks like when the artifact is a per-output manifest, not a slide deck.
Point, do not narrate
A credit decision under review — a limit call, a decline, a pricing decision — starts a regulator question: how was this made? The fintech that narrates the answer is inviting argument; the fintech that points at the manifest is answering the question before it is asked. The manifest (BR-842) carries the model version, the input NodeVersions, the agent trace, the output, and a replay anchor. The regulator questions map onto its fields; the fintech is pointing, not narrating.
The override is logged, not inferred
A credit decision often involves an override — a human raised the limit, a policy exception was granted. The classic audit failure is the override inferred from memory months later. The manifest seals the override as a typed, recorded thing (an AxisOverride is a first-class value), with who, when, and against which NodeVersions. The regulator reads the override from the chain, not from a Slack thread; the override is part of the provenance, not an aside to it.
The fintech that narrates the credit decision is inviting argument. The fintech that points at the manifest is answering the question before it is asked.
- Manifest fields map to regulator questions: model, inputs, trace, override, output, replay anchor.
- Override is logged, not inferred — a typed AxisOverride sealed in the chain, not a memory months later.
- Replay anchor: a third party reruns the computation and gets the identical bytes — evidence, not a story.
What the fintech does not claim
The fintech does not claim the credit decision was correct. The manifest attests provenance and reproducibility, not truth — certified is not true, and the refusal to certify truth is what makes certifying provenance credible. The fintech says "here is how the decision was made, here is the model, here is the data, here is the override, here is the replay hash." The regulator decides whether the decision was sound; the manifest is the substrate for that decision, not the verdict. We are pre-revenue; the manifest fields exist; calibration of the underlying model is gated (Gate 1b) before any accuracy claim.
A credit decision under review is a record you point at, not a story you narrate. The manifest answers the questions the regulator would have asked — and the narration is retired because the record is the audit.
— design-partner credit audit note