Evidence Work

Keep originals read-only. Start with the cheapest faithful representation: page-marked PDF text or programmatic spreadsheet profile. Keep literal source content in document.md; put summaries, relevance, and limits in notes or topic navigation.

Use targeted research when there is a stated evidence gap. Stop when the gap closes, the time limit is reached, or new results have low value. Use full research only for open or external-fact-heavy questions; expand methods, counterexamples, and citation tracing, then stop after two rounds without high-value additions. Do not present either mode as exhaustive.

For PDFs, record uncertain page/region questions and request only that region for visual review. For workbooks, inspect all sheets programmatically and use the conversion check; do not ask people to review rows or silently clean data. Formula text is not a calculated value, and a missing cache is an explicit limitation.

Register sources by DOI, canonical URL, or content hash. A title match alone is not enough. Mark abstract-only or metadata-only access honestly and do not use it to support claims beyond what it contains. Claims need a specific locator and scope.

Pin the actual versions a consumer adopts. Refresh is an explicit choice, not a side effect of new search results. Freeze only the subset used by the output, including applicable official rules; generate citations from that freeze. New unrelated material leaves old freezes valid. When a known error appears, follow only its claim/pin/freeze dependencies and request review there, while retaining prior files for reproducibility.