Skip to main content
Research never changes a model by itself. It produces a proposal that you decide on.

​
Ask for research

Ask: “Research the retailer lead conversion input.” Claude uses research_variable on one input, or on one relationship between two boxes (“research the link from price to demand”). You get:
  • the focused question that was researched,
  • the sources, with what each one actually reports,
  • a summary that keeps conflicting evidence visible,
  • where possible, a specific proposed change with its arithmetic.
It also says whether retrieval was live or came from a configured search provider. Ask “is live web search configured?” and Claude uses research_status.

​
Local assumptions

Evidence from another company rarely applies in full. The transport factor is the share of a published effect assumed to carry over locally, 50% by default, and you can change it. You can also give local numbers the proposal needs, such as baseline hours per decision.

​
Decide on the proposal

Ask: “Accept the proposal”, “Accept it but widen the high end”, or “Reject it because the sample is too small”. Claude uses decide_proposal.
  • Accept creates a new model version and records your local assumption separately from the published finding.
  • Edit accepts with your own estimate.
  • Reject keeps the model as it was. A short note helps whoever reads the history.
Outside evidence can widen a range or move its middle, but it never makes you more certain than you were. See Research sources.
Next: Write memos and documents