Skip to content

Survey checks

Each survey answer goes through checks C1 to C7. While retries remain (advanced.survey.check_retries), an answer that fails a check is sent back to the model with every failure listed; the last answer is kept as it is, and each check it still fails becomes a review issue. The accepted answer is then compared with the measured facts.

Label names are present and unique, labels have kinds, regions use them.

Names compare regardless of case. A region may also use other, which pydelity adds to every profile.

Each region is on a page of the PDF, with a box inside the page.

Rotation ranges lie inside the PDF and do not overlap.

Page-numbering ranges lie inside the PDF and do not overlap.

Every repeated line is accounted for exactly once.

Each repeated line ID appears once, in an artifact or in non_artifact_lines, and no other ID appears.

Writing-direction ranges lie inside the PDF and do not overlap.

The answer lists at least one language, as valid BCP-47 tags.

A tag is valid when langcodes finds every part of it in the IANA language subtag registry; special tags such as zxx are valid.

Record where the accepted answer differs from the measured facts.

A page can differ only where the probe measured a value: a confident text angle for rotation, or a printed number with a value for page numbering. The profile’s value for a page comes from the first range inside the PDF that covers it. The comparison never changes the answer and never sends it back to the model.

  • page_number_differs: List the pages whose measured printed number the profile does not give.
  • rotation_differs: List the pages whose confident measured rotation the profile does not give.