Which part of the response is evidence, and which part describes its limits?
Structured results make research easier to inspect only when their metadata stays attached. A compact response, a partial source and a saved audience can all look like an array of records while carrying different guarantees. Read the actual contract rather than inferring completeness from the visible number of rows or the confidence of a written summary.
Inspect the actual Montage tool response, not only the chat summary. Identify its response family, original records, status, coverage and next cursor. Preserve unknown fields and values, distinguish derived labels, and report whether the displayed rows are only a preview.Identify the response family
Workspace results use the documented signals.v1 envelope with data and meta; search results have their own results categories and metadata. Do not assume one parser or field layout applies to both. Preserve the schema marker and original fields where present so downstream code can distinguish a workspace snapshot from a general search response.
Inspect records and references
Read the actual record identifiers, original links and available evidence fields before using a prose summary as a dataset. A model may display only a small preview or paraphrase the result. Exporting that explanation would not preserve the full tool response. Keep unavailable fields unavailable, and do not fill missing dates or relationships with plausible guesses.
Read metadata alongside the rows
Check status, coverage, freshness and any next cursor before interpreting a count. HTTP success can contain useful partial evidence. A continuation token belongs to the result and query that issued it; do not reconstruct it or mix it with a different request. Keep source limits attached when moving records into another internal analysis step.
Validate the handoff
Compare the records actually delivered with the requested unit and any saved count. Preserve original ordering when required and separate analyst-derived labels from source fields. If a consuming system cannot represent partial coverage or unknown values, fix that contract before dropping the caveats. Never place credentials or short-lived private download URLs into public research output.
Keep the boundaries clear.
- Search and workspace results do not have identical envelopes.
- A prose summary is not a lossless copy of a structured result.
Start with a read-only check: ask your agent to report actual source coverage before collecting records or exporting an audience.
Primary references
Documentation reviewed 2026-11-13. Product capabilities, client availability and plans can change. Third-party marks identify their products, not an affiliation.