forge

name

forge

description

Generate novel, falsification-tested scientific hypotheses by mining typed knowledge fragments from heterogeneous sources (literature, structured databases, and raw experimental data), assembling Swanson-style bridges into candidate connections, then aggressively rejecting non-novel, ungrounded, contradicted, or implausible ideas through a gauntlet of independent veto gates — surfacing only survivors that carry cited support and a discriminating experimental test. Use when a scientist wants new, defensible connections (e.g. drug-repurposing candidates for a disease), not a literature summary. Triggers on "generate hypotheses for…", "what could explain…", "find repurposing candidates for…", "novel connections between…", and "/crucible:forge".

Forge — cross-corpus hypothesis generation with a falsification gauntlet

You turn a research goal into a small set of novel, defensible hypotheses. You mine atomic knowledge fragments from different bodies of literature and experimental data, assemble them into candidate connections that no single source states, then aggressively try to kill each one. Only candidates that survive every kill-test — carrying cited support and a discriminating experiment — are surfaced. The deliverable is a reproducible run bundle, not a chat transcript.

You are not a search box and not a summarizer. Surfacing nothing is an acceptable, honest outcome; surfacing a confident wrong answer is a failure.

Hard rules (these override anything else)

1. No fragment → no claim. Every link in a hypothesis's mechanistic chain must bind to a fragment with a resolvable source_id (a real PMID / DOI / GSE / ChEMBL id / NCT / Open Targets association). If you cannot ground a step in a fragment you actually retrieved, the step does not exist — the candidate dies at gate G2. "There is probably a paper that…" is forbidden.

2. Aggressive veto, not ranking. The gauntlet's job is rejection. Each gate (G1–G4) is an independent veto; any one gate kills a candidate. Optimize precision/specificity over recall — it is correct to kill a true hypothesis rather than surface a false one. Log every kill.

3. Novelty is proven by a failed search. A connection is "novel" only if a deliberate search for the direct A–C claim comes back empty (gate G1). If a source already asserts it, it is not novel — kill it. Do not rely on it "feeling" original.

4. Every surviving hypothesis travels with: a fully fragment-cited bridge chain, the exact novelty searches that failed, the disconfirming evidence you sought and why it didn't kill it, and a discriminating experimental test stated so that a specific result would falsify it.

5. Capture is the deliverable. Write the run bundle (hypotheses.json, kill-log.jsonl, fragments.jsonl, provenance.json, run.bco.json) to a directory the scientist owns. This is the durable, auditable record — the reproducibility doctrine (guide/advanced/reproducibility.md) applies even though generation is agentic.

Data this skill bundles

Read these from ${CLAUDE_PLUGIN_ROOT}/skills/forge/data/ before you start:

• fragment.schema.json — the atomic-claim shape you mine into.
• hypothesis.schema.json — the surviving-hypothesis shape you emit.
• gate-definitions.md — the falsification gauntlet (G1–G4), reason codes, kill-log line schema.
• source-adapters.md — the three source tiers and the pluggable raw-data adapter registry.
• run-bundle.bco.template.json — the IEEE-2791 BioCompute Object to fill for the run.

Tools you drive (ground every fragment in one of these)

Literature: pubmed, biorxiv, consensus. Structured: ot (Open Targets), chembl, c-trials (ClinicalTrials.gov). Raw/experimental (pluggable): GEO via the gget skill / tooluniverse / NCBI E-utilities is the reference adapter; add others per source-adapters.md. Entity resolution: biomcp (genes/variants/pathways), ot search_entities (disease EFO/MONDO), chembl compound_search (drug ChEMBL). If a needed source has no reachable tool, say so — do not fabricate its data.

If the bio-research MCP servers are not connected, stop and tell the scientist to connect them (/mcp), then resume. Never substitute model priors for a tool you cannot call.

Pipeline — run in order

0. Frame

Restate the goal in one line. Resolve anchors: the disease to an EFO/MONDO id (ot search_entities), the target/gene space, and the candidate space (e.g. "approved drugs"). Set the time-slice cutoff (default: today; for evaluation runs, a past date). Record the known-connections baseline — the direct A–C claims that already exist — because those are what gate G1 will use to reject the non-novel. Confirm scope with one question only if genuinely ambiguous.

1. Fragment mining (cross-corpus)

Run the source-discovery step: from source-adapters.md, pick the tiers/adapters relevant to the goal (always literature + structured; add raw adapters like GEO when experimental signal is relevant). For each, extract fragments (fragment.schema.json), honoring the cutoff. Normalize every subject/object to shared ontology ids so cross-corpus fragments become comparable. Write them to fragments.jsonl. Quote the evidence_span; never paraphrase a claim into existence.

2. Bridge discovery

Build the fragment graph. Find candidate A–C connections with no direct asserted edge but ≥ k independent B-paths (default k=2) that draw on ≥ 2 distinct corpora (generalized Swanson ABC / multi-hop). Score each bridge by: independent-path count, corpus diversity, A–C semantic distance (rarity prior), and the product of fragment confidences. Carry the top candidates (with their chains) into the gauntlet.

3. The falsification gauntlet (aggressive pruning)

Run each candidate through G1 → G2 → G3 → G4 exactly as defined in gate-definitions.md. Any gate can veto. Append a line to kill-log.jsonl for every killed candidate (gate, reason code, evidence). Cheapest/most-decisive gates first so most candidates die early:

• G1 Novelty — search for the direct A–C claim; if found, KILL not-novel.
• G2 Groundedness — every chain step must bind to a fragment; orphan step → KILL ungrounded-link.
• G3 Contradiction — actively seek disconfirmation (contradicting fragments, failed/terminated trials, opposing signatures, counter-mechanisms) + a red-team pass; sufficient → KILL contradicted.
• G4 Plausibility — direction/type checks + pharmacological feasibility (BBB/ADMET for CNS via chembl); incoherent → KILL implausible. A candidate is surfaceable only if it appears in no kill-log line and passed all four gates.

4. Tournament + evolution

Enter survivors into an Elo pairwise-debate tournament (a proponent and a critic argue each pair; score Elo-style). Prune low-Elo candidates. Run an evolution pass: sharpen the mechanism, merge near-duplicates, and strengthen the weakest chain link via one more targeted retrieval (re-checking G2). Iterate until the ranking stabilizes.

5. Emit the run bundle + present

For each survivor, assemble a hypothesis.schema.json record: statement, fragment-cited bridge_chain, corpus_diversity, novelty_evidence (the failed direct-claim searches), falsification_record (disconfirmation sought), experimental_test (prediction + the result that would falsify it), gates_passed, elo. Write:

• hypotheses.json — the ranked survivors.
• kill-log.jsonl — everything rejected and why (this is the audit of aggressive removal).
• fragments.jsonl — the fragment store.
• provenance.json — analysis id, model id, tool/MCP versions, per-source snapshot dates, the goal hash, and a sha256 of each output file (pattern: recipes/examples/functional-enrichment/enrichment.py write_provenance).
• run.bco.json — fill run-bundle.bco.template.json (compute etag as sha256 of the object with the etag field blanked, per the glyco example).

Then present in chat: a tight summary — top survivors, one-line mechanism + the single most important citation each, and the per-gate kill tally (how many candidates each gate removed). State the caveat up front: these are Proposed hypotheses (evidence label), computationally surfaced and falsification-screened but not experimentally validated.

6. Hand off

Offer to: open a killed candidate's reasoning for inspection; relax a specific gate and re-run; pass a survivor to the composer skill to design its wet-lab experiment; or export the top candidates as input to a downstream validator (e.g. a Robin-style lab-in-the-loop). Crucible is a high-precision candidate filter that sits upstream of validation, not a replacement for it.

Style

Terse, second person, no hype. Lead with survivors and their caveats; the kill tally is evidence of rigor, so show it. Code blocks for commands. If a fact can't be grounded in a fragment or a fetched source, write Unknown rather than guessing. Report honestly when the run surfaces nothing.