boheling/deltasci/skill

DeltaScience: Two-Perspective Co-Reasoning

Purpose

Generate AI4Science research hypotheses that are:

1. Grounded — every factual claim is tagged with its evidence type, source, and the AI's self-assessed training coverage.
2. Honest about AI's epistemic edges — claims outside the AI's training distribution are emitted as KNOWLEDGE_GAP for the researcher, not fabricated.
3. Explicit about creative leaps — connections the AI is proposing (not citing) are emitted as NOVEL_SYNTHESIS.
4. Falsifiable — every hypothesis ships with a measurable threshold for being wrong.
5. Domain-aware — uses a pluggable lens specific to the scientific field.
6. Two-perspective — alternates between domain expert and ML engineer.

This is the Claude Code skill version of the deltasci Python package. They share the same domain packs and grounding rules.

The three first-class tags

[CLAIM type=<TYPE> coverage=<COVERAGE> source="<CITATION>"]<text>[/CLAIM]
[KNOWLEDGE_GAP category=<CATEGORY>]<question for the researcher>[/KNOWLEDGE_GAP]
[NOVEL_SYNTHESIS rationale="<one-line>"]<the proposed connection>[/NOVEL_SYNTHESIS]

coverage ∈ {well-covered, sparse}. uncovered is not allowed on a CLAIM — for that, emit a KNOWLEDGE_GAP instead.

A round with zero KNOWLEDGE_GAPs and zero NOVEL_SYNTHESES is suspect — it suggests the AI is claiming complete certainty across the entire research idea, which is itself a hallucination signal. Synthesis refuses by default.

Inputs

Outputs

Step-by-Step Procedure

Step 1 — Identify the domain pack

Read <pack_dir>/pack.toml and <pack_dir>/lens.md. If the user's request doesn't name a pack and the domain is unambiguous, pick one without asking.

Step 2 — Run the 4-round dialogue

Follow the prompts in prompts/. The flow is:

Round 1  Domain Scientist  (mechanism, unmet need, prior work, constraints)
Round 2  ML Engineer       (data representation, method, precedents, risks)
Round 3  Domain Scientist  (refinement, evaluation realism, falsifiable prediction)
Round 4  ML Engineer       (revised plan, math, implementation, expected outcomes)

In every round, every factual statement must be one of the three tags. See references/grounding_rules.md and references/coverage_axis.md for the rationale.

After each round, scan the output. If you find untagged factual claims, redo that round (one repair attempt). If a round has zero KNOWLEDGE_GAPs and zero NOVEL_SYNTHESES, prompt yourself: "Am I being honest about my training-distribution edges? Am I being honest about which connections are leaps vs cited?" — and rewrite if needed.

Step 3 — Write the transcript

Write to <out_dir>/transcript.md with each round labeled and all three tag types preserved.

Step 4 — Synthesize the grounded hypothesis

Follow prompts/synthesis_round.md. Produce JSON with:

• title, statement
• domain_grounding, technical_approach
• falsifiability (prediction + threshold + null_outcome — all required)
• feasibility_scores, feasibility_justifications

The CLAIMs, KNOWLEDGE_GAPs, and NOVEL_SYNTHESES are collected automatically from the transcript — do NOT re-emit them in the synthesis JSON.

Hard rules:

• No falsifiable threshold → refuse, output {"error": "no_falsifiable_clause", "reason": "..."}.
• Across all rounds, zero KNOWLEDGE_GAPs and zero NOVEL_SYNTHESES → refuse, output {"error": "no_epistemic_humility", "reason": "..."}.

Step 5 — Render the hypothesis

The final hypothesis.md has THREE evidence sections, in this order:

1. AI-confident foundations (well-covered claims with citations)
2. Likely-reliable, please verify (sparse-coverage claims — cite carefully)
3. Researcher knowledge required — KNOWLEDGE_GAPs + NOVEL_SYNTHESES

Plus an Epistemic summary with counts of each tag type and any warnings (e.g., "sparse claims outnumber well-covered ones").

Step 6 — Report

Tell the user:

• The hypothesis title
• Overall feasibility score
• Counts: well-covered / sparse / knowledge gaps / novel syntheses
• The falsifiability threshold
• Any epistemic warnings
• Where output files are

Configuration

Rules

• Never invent citations. If you can't name a real reference, either tag the claim coverage=sparse and use the most-honest source you have, or emit a KNOWLEDGE_GAP.
• Never use `coverage=uncovered` on a CLAIM. Uncovered material goes in a KNOWLEDGE_GAP.
• Mark your leaps as NOVEL_SYNTHESIS. Combining two well-covered facts into a hypothesis nobody has written down is good — it's the point. But say so explicitly.
• Skip the falsifiability gate is not allowed. A hypothesis without a measurable threshold is a wish.
• The pack lens is read literally. It parameterizes how the domain expert reasons.

Checklist

• [ ] Domain pack identified and lens loaded.
• [ ] 4 rounds completed with proper CLAIM / KNOWLEDGE_GAP / NOVEL_SYNTHESIS tags.
• [ ] At least one KNOWLEDGE_GAP and at least one NOVEL_SYNTHESIS across the transcript.
• [ ] Synthesis produced valid JSON matching the schema.
• [ ] Falsifiability clause has prediction + threshold + null_outcome.
• [ ] Hypothesis.md has all three evidence sections.
• [ ] Outputs written: transcript.md, hypothesis.md, summary.json.