Every AI field id, in template order: the chunking plan a caller walks through annotationIds.
AI fields this dry run actually asked for — min(aiFieldCount, cap).
Grouped and redaction-aware, the same projection the run log carries; raw details never leave the server.
Answered fields, deterministic lane first, then the AI lane in template order.
The filled PDF, base64, never stored anywhere. Null when it was not asked for or was too large to return.
Annotations with no prompt and no sub-prompts: no question to answer, left for a human.
The cap left AI fields unanswered; the preview is a sample, not the form.
Answers the renderer refused to draw (off-page, degenerate box, unencodable text).
Boxes this preview put no ink into, counted per box.
AI fields the template has in total, so the caller can say "10 of 47".