hypothesis-pipeline-rubriclisted

# Hypothesis Pipeline Rubric ## When to Use Trigger this skill whenever: - A testable claim appears in conversation: directional ("X scales with Y"), mechanistic ("A is regulated by B via pathway P"), comparative ("X differs between groups"), methodological ("method A outperforms method B"), or synthesis ("these two literatures predict the same pattern"). - A proposal from `weekly_hypothesis_generator` surfaces in a briefing and the researcher wants to understand why it passed or was blocked. - The researcher is drafting a hypothesis for a grant Specific Aims page or a preregistration and wants a rubric check before running the formal pass. - A pipeline result needs to be explained to a student or collaborator. The pipeline is type-aware, not domain-aware. It does not hardcode any particular biological subdiscipline; the rubric items it applies depend entirely on the structural shape of the claim (see claim types below). --- ## The Four-Tier Gate The pipeline in `agent/hypothesis_pipeline.py` runs four stages in sequence. Each tier has a different cost, model, and blocking criterion. ### Tier 0 — Deterministic Smoke Filter (free, no LLM) Regex-based. Blocks claims that are: - Placeholder or smoke-test artifacts (`[SMOKE TEST]`, `<PLACEHOLDER>`, `<TO-BE-FILLED>`, `TODO`, `FIXME`, `<INSERT...>`) - Fewer than 30 characters (too short to be a real hypothesis) - Accompanied by notes containing `"backfill"`, `"tool validation"`, or `"smoke test"` A claim blo