Documentation

A review starts with a frozen scope.

The authenticated workbench keeps original evidence, review execution and human disposition together.

1. Create a case

Select the correct data class. A synthetic case must contain invented data only. Confidential cases require activation records. Keep one borrower and facility scope per case.

2. Add source versions

Upload a PDF or UTF-8 text version of the draft memo, the existing issue log and supporting evidence. Include the source date and when each version became available. Every version receives a SHA-256 fingerprint.

3. Freeze the snapshot

Select exact source versions, an approved policy version and an analytical cutoff. A later source cannot be silently included. Corrections require a new version and snapshot.

4. Request a review

Select an approved experimental arm, mode and cost ceiling. Execution status is separate from findings and coverage. A provider or processing failure remains visible.

5. Record findings and dispositions

Review the exact memo claim and original source reference. Record an individual rationale and supporting evidence for each action. Material disputed closures and overrides require an independent second review.

6. Export the record

The JSON export includes snapshot and policy fingerprints, checks, findings, limitations and the history of human dispositions. Open and incomplete states remain present.

Private read API

Organization owners can create short-lived keys limited to one data class and selected read permissions. Version 1 retrieves cases, frozen snapshots, review status and audit exports. Keys cannot change a finding disposition. The endpoints are /api/v1/cases/, /api/v1/snapshots/<id>/, /api/v1/reviews/<id>/ and /api/v1/reviews/<id>/export/. Send the key in the Authorization Bearer header over HTTPS. The limit is 60 requests per minute.

Programmatic integration contracts are versioned separately. No endpoint can approve or deny a loan, send an adverse-action notice or impersonate a human reviewer.