research-soplisted

# Research SOP When the user asks a research-shaped question ("investigate X", "does X affect Y", "compare A and B"): 1. **Memory lookup**: call `mcp__memory__match_signatures` with the task description. If prior failures exist, read them first. 2. **Literature gap**: call `mcp__memory__query_literature`. If fewer than 3 relevant papers exist, spawn the `librarian` subagent. 3. **Hypothesis generation**: spawn the `researcher` subagent with literature context. Ask for 3-5 hypotheses. 4. **Hypothesis selection**: if you have 3 or more candidates, run `$bt-tournament` (V3.0 successor of `elo-select`). Use `mcp__memory__judge_hypotheses` and `mcp__memory__record_judgement` to rank them; the call dual-writes to the BT comparison ledger. Inspect `mcp__memory__get_bt_leaderboard` and keep the top 2 only when their 95% intervals separate. Otherwise run more comparisons. *Before picking one path among several*, optionally call `mcp__cockpit__narrate("one-sentence reason for the choice")` so the cockpit's activity pane reflects the decision (v5.0; optional, suggestion only). 5. **Preregister when the claim becomes confirmatory**: for exploratory prototypes, the engineer may run and pin results without a prereg as long as the output is labelled exploratory. Before promoting a result to a main confirmatory claim, run `$preregister` and lock metric, direction, threshold, alpha, and `mc_correction='bh'` into `ver_preregistrations`. The reviewer later refuses confirmatory manuscript clai