Hypothesis Crucible (Claude Plugin)
A Claude Code plugin in this repo’s marketplace that generates novel, falsification-tested hypotheses: it mines atomic knowledge fragments from different bodies of literature and experimental data, assembles them into candidate connections no single source states, aggressively kills the weak ones through a gauntlet of independent veto gates, and surfaces only the survivors — each with a fragment-cited mechanism, a proof-of-novelty search, and an experiment that could falsify it.
| Type | Claude Plugin (bundles the forge skill + /crucible:forge command) |
| Supplier | Scripps AI Enablement (this repository) |
| Availability | GA — shipped in the sci-ai-enabler marketplace |
| Pricing | Free / OSS |
| Capabilities | Read/Write — drives bio-research MCP servers; writes a reproducible run bundle to disk |
| Verified | works · 2026-07-27 |
| Security | cleared · 2026-07-27 — first-party this repo, plugin dir + marketplace registration confirmed |
How to install
- Claude Code — add this repo as a plugin marketplace, then install:
/plugin marketplace add scripps-ai-enablement/sci-ai-enabler /plugin install crucible - Prerequisite MCP servers — connect the
bio-researchservers the skill drives (pubmed,ot,chembl,c-trials,biorxiv,consensus,biomcp,tooluniverse) via/mcporclaude mcp add …. GEO (raw-data adapter) is reached via theggetskill /tooluniverse. The plugin runs without them but every fragment must be grounded in a real tool call, so an unconnected server means that source is skipped rather than fabricated.
What it does
Given a research goal (e.g. “repurposable approved drugs for Alzheimer’s disease”), the forge skill runs a six-stage pipeline: frame the goal to ontology anchors; mine typed, provenance-carrying fragments across literature, structured databases (Open Targets, ChEMBL, ClinicalTrials.gov), and a pluggable tier of raw experimental sources (GEO is the reference adapter); discover Swanson-style A–C bridges with no direct edge but multiple independent cross-corpus B-paths; run every candidate through a falsification gauntlet (G1 novelty by active negative search, G2 groundedness, G3 contradiction/red-team, G4 plausibility) where any gate can veto and every kill is logged; rank survivors in an Elo debate tournament; and emit a run bundle (hypotheses.json, kill-log.jsonl, fragments.jsonl, provenance.json, run.bco.json).
Primary use cases: drug-repurposing candidate generation; explaining an observation via cross-domain mechanism; filtering a large pool of raw hypotheses down to a defensible few before committing experimental resources. It optimizes precision over recall — it is designed to surface few, well-supported ideas and to reject aggressively.
Notes
Distributed as a plugin (crucible/) in this repository. The reasoning is model-driven (a SKILL.md procedure), so runs are auditable via the emitted IEEE-2791 BioCompute Object rather than byte-reproducible; deterministic scoring of a captured run lives in recipes/examples/hypothesis-crucible/eval/score.py. Positioned as a high-precision candidate filter upstream of experimental validation, complementary to end-to-end systems like Robin, Co-Scientist, and OpenScientist.
Sources
Installed this tool?
Share feedback — install path, OS, errors, workarounds. The form opens with this tool pre-selected and a link back to this page.