Study

Workflow

1. Frame the study.

   • When the question's shape or sub-questions are ambiguous, ask once, at most a few questions, before framing. Unattended, frame on the safest reading and say so.
   • Write the question in one sentence and name its shape: subject, one thing and how it works or what is known about it, or field, how several sources or approaches answer one question. The shape is fixed here; thin discovery later is a coverage gap, never a reason to change it.
   • List the sub-questions the report must answer; for a field, the dimensions to compare. The stop test in step 4 runs against this list.
   • Set the budget: source count estimate, time, tokens, and whether web access is granted.
   • Read references/evidence-contract.md, the evidence contract. Every record in the study follows it.

2. Choose sources. Follow references/research.md for the scout and collector work orders.

   • Sources the user named are in and seed the scout. Skip scouting only when the user limits the study to the named sources.
   • Scout: a shallow sweep from the seeds that returns candidate sources, one line each, with no close reading.
   • Select collection units. A unit is one system, work, or author's source family at one version or surface. For a field, group candidates by family, a system, author, or school rather than a URL, and take one representative per family with the best primary access, plus every outlier that answers the question differently and every user-named source. Expect three to six units; take the whole field when it is small. For a subject, take the surfaces and versions the sub-questions need. Record each dropped candidate and the reason in the study header.
   • Before collecting, report the frame and selected units. Separate user-named sources from scouted additions, and give each addition's purpose and budget impact. Unless the user requires approval of additions, continue without waiting for a reply; when they do, request it in that report and leave additions uncollected until approved.
   • Collect independent named units first. Then apply any corrections received and reassess additions against the remaining budget. Record additions left out for budget or missing approval as dropped candidates, and deliver partial when consequential coverage gaps remain.