Research

Two frames; infer the frame from the request, never ask which:

• report — no decision attached: a primer, survey, deep-dive, or whole-domain landscape map. Load references/report.md.
• opportunity — a decision attached: "should we build / enter / invest", go/no-go, sizing a named idea. Load references/opportunity.md.

Core principle: keep the line between what is known and what is asserted visible. Every load-bearing claim is cited or marked as inference. A research branch that ends without a citation or a next test is organized procrastination, not research.

Run to completion; state assumptions inline. Ask only on a genuine blocker: the frame is undecidable, or the opportunity frame lacks a one-line opportunity statement (fan-out without scope produces disconnected dumps).

Invariants

• A vendor-aligned source is never the sole support for a claim about that vendor's own category.
• When well-triaged sources conflict, both go in the output as a named debate — silently picking one is editorializing.
• Opportunity frame: every assumption carries a falsifiable test; every severity-4 risk is resolved or becomes a kill criterion; the memo ends in go / no-go / pivot, never "more research needed".
• Cut weak citations rather than pad: ten triaged sources beat forty padded ones.

Workflow