define-claimslisted

# Define claims and experiments This is the one step that is genuinely the agent's job: deciding *what the paper will assert* and *what runs could provide evidence*. The scripts cannot invent your hypotheses. But once you write them down, the deterministic CLI takes over. **Boundary**: you (the agent) reason about the science; `insightflow validate` checks the structure; `insightflow plan`/`state` later score the evidence. Do not hand-author confidence numbers or a schedule — only the claims and the grid. ## Steps 1. **Extract the claims.** From the research idea / outline / codebase, write `configs/claims.yaml`. One claim per assertion the paper will make. For each: - `statement` — the precise scientific claim. - `target_metric`, `desired_direction`, `minimum_effect_size` — the decision rule (what effect counts as real). - `required_seeds` — how much replication a reviewer would demand. - `importance` (high/medium/low) and `reviewer_risk` (0–1) — what's central and what a reviewer will attack if unsubstantiated. Prefer a few sharp, falsifiable claims over many vague ones. Mark generality / external-validity claims explicitly — those are usually the binding constraint. 2. **Enumerate the experiment grid.** Write `configs/experiments.yaml`: every `(method, dataset, condition, seed)` cell that *could* provide evidence, with `claim_links`, `dependencies`, `expected_cost`, `expected_time`, and a real `command`. **Always include baselin