| name | paper-claim-auditor |
| description | Use when auditing a CFD-AI/SciML manuscript's abstract, introduction, contributions, or discussion to verify that every claim is supported by evidence and scoped correctly. |
| version | 0.3.0 |
| author | CFD-AI Paper Skills maintainers |
| metadata | {"short-description":"Claim-evidence audit for CFD-AI/SciML papers"} |
Paper Claim Auditor
Trigger
Use when asked to avoid overclaiming, check novelty, audit abstract/introduction/conclusion, or map claims to evidence.
Progressive disclosure
- Read
rubrics/claim-evidence-rubric.md before scoring.
- Route to
skills/reviewer-audit-toolkit/SKILL.md when the user asks for a gate decision, rejection risk, submission readiness, or minimum rescue plan.
- Read
rubrics/vocabulary-style-rubric.md when the text contains loaded adjectives such as robust, generalizable, physically consistent, novel, state-of-the-art, interpretable, efficient, real-time, promising, or transformative.
- Use
references/field-terminology-style-guide.md to replace AI-ish prose with field-native CFD-AI/SciML wording.
- Use
references/gold-paper-style-patterns.md and rubrics/gold-paper-closeness-rubric.md when text sounds fluent but unlike a normal CFD-AI/SciML paper.
- Use
templates/claim-evidence-map.md for full audits.
- Use
examples/bad-to-good-abstract.md, examples/ai-ish-to-field-native-prose.md, and examples/generic-ai-to-gold-paper-prose.md when rewriting abstracts or conclusions.
- Read gold-paper files only when the claim domain matches:
- reconstruction:
references/gold-papers/fukami-2019-super-resolution-jfm.md
- field taxonomy:
references/gold-papers/brunton-2020-machine-learning-fluid-mechanics.md
- uncertainty/trust:
references/gold-papers/maulik-2020-probabilistic-neural-networks-prf.md
- wake prediction:
references/gold-papers/lee-2019-cylinder-wake-jfm.md
- CFD opportunity framing:
references/gold-papers/vinuesa-2022-enhancing-cfd-ml.md
Workflow
- Extract all major claims.
- Classify claim type: novelty, accuracy, physical consistency, generalization, robustness, efficiency, interpretability, reproducibility.
- Map each claim to evidence location.
- Score evidence strength: supported / weak / missing / overclaimed.
- If the paragraph is generic or AI-ish, score gold-paper closeness: physical anchoring, rhetorical move order, field-native collocations, and limitation boundary.
- Suggest safer wording or required experiment.
Output schema
| Claim | Type | Evidence location | Status | Risk | Fix |
|---|
Add a rubric score:
| Claim | Score 0-3 | Missing evidence | Safer wording |
|---|
For prose-style audits, add:
| Passage | Gold-paper closeness score 0-3 | AI-ish marker | Field-native rewrite |
|---|
Anti-patterns
- “robust” without robustness tests.
- “generalizable” from random split only.
- “physics-informed” without locating physics in loss/architecture/data/constraints.
- “state-of-the-art” without fair baselines.
- “real-time” without hardware/runtime.
- “AI framework” or “complex dynamics” where the paper should name input, output, flow regime, metric, and diagnostic.
- Generic context paragraphs that do not mention a physical object by the third sentence.
Verification
- Every claim has evidence or TODO.
- No solver/citation details invented.
- Safer wording preserves contribution without hype.
- Claims scored 0-1 are either removed, narrowed, or converted to explicit TODOs.
- Gold-paper closeness failures are rewritten using
examples/generic-ai-to-gold-paper-prose.md.