AI credit analysis software. Built for examination.

Every generation creates a frozen snapshot — model identifier, version, seed, inputs (the cited file spans), output, and any analyst overrides. The snapshot is hashed and stored against the loan's audit trail.

Examiner replay opens the generation snapshot in read-only context. The examiner sees what the model saw, what it produced, and what the analyst changed. Re-running the generation against the same inputs produces a comparable trace, line by line.

The replay surface is part of the platform — not a one-off export. It exists for every memo, every spread row, every classification decision.

Yes — at the span level, not the document level. A paragraph in a credit memo cites the page and line range that supports it. A spread cell cites the figure on the financial statement that backed it. A document classification cites the analyst (or model) decision that placed it in a checklist slot.

All citations are clickable. Hovering or opening a citation surfaces the source document and highlights the span.

CORE is model-agnostic by architecture. Today, the platform routes generations to two pinned model families with stable version identifiers, chosen against measured cite-faithfulness on credit-memo workloads.

Every generation snapshot records the model identifier and version it was produced under. Upgrading the model fleet does not retroactively alter prior outputs — the original snapshot is examiner-replayable on the original model.

Model selection, version pins, and routing rules are documented in the operator handbook and disclosed in implementation.

CORE routes AI like a credit department: section intent, source evidence, credit policy, regulatory context, analyst pair, and supervisor review. The model proposes; the analyst accepts, edits, or rejects; and every decision creates an audit entry.

Where automation-first AI leads with task execution, CORE's posture is examination-first. The platform was built inside an active program-credit operator and tested against examination findings, not against a feature backlog. This is a different center of gravity.

Both platforms have a place. The question for a regulated lender is which posture survives an examination question with no time to prepare.

Yes. CORE ships an AI Architecture Disclosure that covers: model fleet and version pins, prompt families, cite-grounding rules, decline conditions, retraining policy, snapshot schema, and the replay surface contract.

The disclosure is provided to examiners on request and to capital partners during diligence. It is not marketing. It is the document an examiner needs to evaluate the surface they are looking at.