Transparent by design
How a check works
A check is one request and one response. Upload up to 5 evidence files and receive a source-backed assessment for every confirmed rule. There is no job to poll and no second, competing verdict.
The five stages
1. Validation
Files are checked for supported type, size, page count, and readability. Refused files never reach the model.
2. Confirmed rules
The exact natural-language rules you approved are locked into the check snapshot, including their wording, applicability, severity, and guideline source.
3. Independent AI review
The model receives that confirmed snapshot, the evaluation date, and the original evidence files. It must assess every rule exactly once.
4. Citation reference validation
Each decisive finding needs an uploaded document and real page reference. Derived conclusions cite the underlying fact and explain the reasoning.
5. Structured result
The API returns one primary result with statuses, observed values, explanations, confidence, citations, limitations, and reviewer questions.
A rejected file stops the request. An incomplete rule finding or ambiguous value becomes a review item, while valid findings remain in the report.
How rule results add up
Each approved rule is PASS, FAIL, or REVIEW. The headline follows a fixed precedence:
- Any FAIL makes the result non-compliant.
- Otherwise, any REVIEW makes it review required.
- Only when every applicable rule passes is it compliant.
Severity describes importance; it never hides an unresolved or failed rule. Applicability is interpreted as part of the approved rule wording.
Your own requirements
Proposed rules are extracted with their source sentence and page, then remain editable until you confirm them. Only confirmed rules are supplied to the AI review, and their exact snapshot is returned with the result.
Next
Security and privacy explains document handling, or you can run a check.