# Paper Claim Audit > Use when verifying a scientific claim or numerical parameter through project artifacts, implementation semantics, primary literature, supplements, and cited predecessors, and show decisive paper passages as inline PDF evidence crops when available. Applies to paper checking, parameter provenance, model-to-code comparisons, and high-cost misreporting risks. Do not use for a routine paper summary that does not ask whether a claim is correct. - Skill: `soyonaoc/paper-claim-audit` (Agent Skill, multi-file: 5 files) - Install (CLI): `npx skillmds@latest add soyonaoc/paper-claim-audit` - Raw SKILL.md: https://api.skillmd.com/api/skills/soyonaoc/paper-claim-audit/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: SOYONAOC (https://skillmd.com/u/soyonaoc) - Updated: 2026-09-22 - Page: https://skillmd.com/skills/soyonaoc/paper-claim-audit --- # Paper Claim Audit Produce an auditable answer whose conclusion is no stronger than its evidence. Treat the investigation as read-only unless the user separately authorizes edits, downloads into a maintained data tree, or external actions. ## Define the claim Turn the request into one or more atomic claims before searching. Distinguish: - the general theoretical definition; - the value stated or inherited by a paper; - the value active in the current project; - the implementation that converts configuration fields into a physical quantity; - the scientific interpretation of that quantity. Resolve ambiguous scope from the active repository when practical. State the chosen model, version, configuration, species, units, and epoch. Ask the user only when a reasonable scope choice would materially change the requested conclusion. ## Build the evidence chain 1. Inspect the current project artifact first when the question is project-specific. Identify the production configuration, executable or library version, validation contract, and any recorded provenance. Do not assume that a README describes the live configuration. 2. Resolve paper identity and version with ADS/SciX, DOI records, arXiv, or the publisher. Metadata services establish identity and publication metadata; they do not establish scientific content. 3. When arXiv source and PDF are available, use a **source-first, PDF-confirm** route. Search the active TeX source for exact equations, units, tables, citation keys, and included files, then verify the decisive passage in the same-version rendered PDF. Source is the efficient semantic reading channel; PDF establishes that the material appeared in the compiled paper and supplies stable page, table, figure, and visual context. Read [references/source-pdf-workflow.md](references/source-pdf-workflow.md) before acquiring or presenting paper evidence. 4. Read the relevant primary-paper section, equation, table, caption, supplement, or author-provided configuration. Trace inherited numbers through citations until the source that actually defines them is reached. A later paper can establish adoption of an earlier model but cannot by itself establish undocumented parameter values. 5. Inspect the implementation used by the active version. Reconstruct the executed formula from source code and configuration, including defaults, feature switches, branch conditions, normalization conventions, unit conversions, and spatial or species dependence. 6. Cross-check the resulting claim against at least one genuinely independent evidence path when available. A README copied from the same configuration is corroboration, not an independent source. 7. Run a focused calculation or validator when it can test the interpretation. Label the result as a reproduction or configuration check, not as a paper statement. For a detailed audit report, use [references/report-template.md](references/report-template.md). ## Label evidence explicitly Classify every central statement as one of: - **Project artifact fact**: value read from the active configuration, output, or data; - **Implementation fact**: behavior established by the relevant source-code path; - **Paper statement**: claim made directly by the authors; - **Cited input**: value or model inherited from another work; - **Model assumption**: selected form, boundary, prior, or simplification; - **Reproduction result**: quantity recomputed from documented inputs; - **Current inference**: interpretation added during the audit. Give an exact locator for decisive evidence: file and line, section, equation, table, figure, appendix, supplement filename, DOI, arXiv version, or source revision. For each decisive paper passage, provide a compact evidence card containing an **inline-rendered** tight PDF crop, a short source transcription, an exact locator, and two explicit fields: **supports** and **does not establish**. A hyperlink to the paper or image file is a locator, not a substitute for displaying the crop. On a headless host, use `scripts/render_pdf_evidence.py`; call the result a PDF page crop rather than implying that a graphical desktop was used. ## Require visible paper evidence When an accessible PDF contains decisive evidence and the user has not opted out of images, the final response is incomplete until the relevant crop is visibly embedded: - Render at least one tight crop for each distinct decisive paper passage. One crop may support several atomic claims when the same passage establishes them. - Mark the exact words, equation, table row, or footnote being discussed when the crop contains surrounding material. Default to translucent light-purple fill with a medium-violet border; preserve and hash an unannotated companion crop so the overlay is auditable. - Use the host's native image attachment or rendering mechanism. In the Codex desktop app, embed a local crop with Markdown image syntax and an absolute path, for example `![PDF evidence: paper ID, page N](/absolute/path/to/crop.png)`. - Place the crop next to its locator, transcription, **supports**, and **does not establish** fields. Keep the public paper link as a citation in addition to the crop. - Before answering, verify that the PNG exists, is non-empty, has been visually opened, is not clipped, and is referenced by an actual image block in the response. A plain file link, source link, DOI link, or statement that a crop was made does not pass. - If acquisition, rendering, visual inspection, or inline delivery is impossible, say exactly which stage failed and mark the visual deliverable incomplete. Do not silently degrade to links only. Keep scientific confidence separate from delivery completeness. ## Handle discrepancies without smoothing them away - Preserve conflicting values and determine whether they reflect version drift, inactive parameters, differing units, different model variants, rounded values, or a genuine inconsistency. - Check control switches before interpreting any nearby parameter. A populated value may be inactive. - Distinguish an exact historical reproduction from a compatibility port, adaptation, or current production choice. - If the evidence chain ends before the claim is established, report the conclusion as unresolved and name the missing artifact. Do not fill the gap from memory, a review, a search snippet, or a numerically convenient value. ## Misreport-prevention gates Before answering, verify all applicable gates: - The source was opened at the relevant passage; a search hit alone is insufficient. - The arXiv source and PDF are the same explicit version, and the cited TeX belongs to the active compilation rather than a comment, discarded draft, or inactive branch. - The PDF was checked for every source-derived claim presented as published content. - The numerical value has units, normalization point, functional form, and validity range. - Every active multiplier, exponent, switch, default, and branch in the implementation has been checked. - Paper identity, version, and citation direction are correct. - General framework behavior is not conflated with the selected configuration or model. - A parameter printed in a file is not called fitted, observed, or recommended unless the source establishes that status. - Any calculation uses the project's declared environment and fails visibly on missing dependencies or inputs. - Evidence crops are visually inspected, tightly bounded, and accompanied by hashes and retrieval/render metadata. A crop is evidence of wording or layout, not proof of scientific truth. - Highlight overlays identify the passage under discussion without replacing the raw crop. Their coordinates, colors, opacity, border, and raw/annotated hashes are recorded. - Every accessible decisive PDF passage is displayed inline in the answer unless the user explicitly requested text-only output. Links alone do not satisfy this gate. - Confidence is reported as **confirmed**, **supported**, or **unresolved**, with the remaining limitation stated next to the conclusion. ## Answer shape Lead with the scoped conclusion, confidence, and physical formula. Then visibly embed the PDF evidence cards before giving source transcription and locator, provenance chain, current configuration and implementation, independent checks, discrepancies, remaining uncertainty, and evidence manifest. Keep quoted text short and paraphrase the rest. Cite public sources near the claims they support and include local clickable file links when working in a repository.