researchlisted

# research-factory One pipeline: **FRAME → GATHER → VERIFY → COMPOUND**. A fuzzy question in; a decision-ready, sourced brief out; the library remembers it and the loops learn from it. Three axes compose per run: **mode** (what kind of question) × **tactics** (how to search) × **avenues** (where to search). Rationale and evidence for every rule live in `docs/PLAN.md` and the library's own briefs — this file is just the protocol. ## The run **1. Library in.** `node scripts/related.mjs "<question, expanded with synonyms>" --n 3` — you are the retriever's semantic layer, so expand first. Fresh brief answers it → reuse, don't pay twice. Related briefs → carry their claims in (the run must CONFIRM / EXTEND / CONTRADICT them). Nothing → fresh ground. **Carry-in is for reconciliation, not framing:** draft the engine question from the *user's* question — if a prior brief's conclusions (including ambient-surfaced ones) steer the new decomposition, a later CONFIRM is circular, not corroborating. **2. Mode** — pick ONE; say which and why. Ambiguous/too-broad → ask 1–2 clarifying questions first. Fits none → don't force it; run the closest loosely (the map is what's incomplete). | Mode | Trigger shape | Pipeline | |---|---|---| | [`survey`](modes/survey.md) | "state of the art / best practice for X" | full fan-out (lean engine) | | [`build-vs-buy`](modes/build-vs-buy.md) | "tool/lib/product for X, or build it?" | fan-out, options matrix | | [`decision`](modes/decision.md) | "A vs B