--- name: report-evidence description: Verify a report's claims, result provenance, evaluation scope, prior-work comparisons, citations, bibliography metadata and external links against permitted evidence. Use for source audits or questions about whether the final report contains the correct results. --- # Report evidence Separate what the text says, what the available records establish, and what remains uncertain. This skill verifies existing evidence; it does not authorise new experiments, external uploads, submission or access to restricted raw data. ## Recurring issues Whenever a user or reviewer identifies a complex or systematic issue that plausibly recurs, formulate or refine a contextual pattern and record the example. Apply the pattern across the draft when that broader check is already authorised; otherwise explicitly offer the wider pass alongside addressing the local request. Follow the [recurring-pattern procedure](../report-review/references/recurring-patterns.md) for scope, contextual matching, counterexamples and completion. This applies even when only one instance has been pointed out; a routine isolated typo does not automatically justify a broad audit. ## Claims and provenance Identify the canonical source/export and the underlying permitted evidence. Construct a claim-to-evidence chain for consequential findings and changed claims: requirement/question, method, data/configuration, observation, interpretation, limitation. Compare the rendered report as well as source when checking the submitted or final version. For numerical results, verify totals, units, denominators, overlap, rounding, population and exclusions. Distinguish failed checks from affected rows, unique entities and confirmed defects. Distinguish observed records from reconstructed context, empty values from absent or unobserved fields, and declared requirements from exploratory assertions. Record assumptions and which outcomes they affect. For latest-result checks, trace report figures to retained summaries, run/configuration records and input identities. Use existing hashes or metadata within the permitted boundary; recency of a filename alone proves little. Separate historical datasets and tools from current assessments. Separate latest supplied data, latest prepared data, latest completed run, and latest published results; a newer input does not establish that it was assessed. Identify which of these the evidence actually establishes. If the executable identity was not retained, say so; an independent recount or bounded replay corroborates outcomes but does not certify the exact full-run binary or constitute a new full run. Keep controlled verification, real observations, synthetic demonstrations, human evaluation and formal acceptance distinct. Missing evidence is not a negative result. Unperformed comparisons do not establish whether their input artefacts exist. Check actual records before assigning complete, partial, proposed or undocumented status. ## Prior work and sources Read [source verification](references/source-verification.md) for a bibliography, link or literature-comparison audit. Use authoritative primary sources and the version relevant to the claim. Verify expressiveness, implementation support, configuration and demonstrated benefit separately. A different representation or placement of information does not itself establish greater expressive power, and a specification's capability does not prove a particular tool enforces it. Check that comparisons use appropriate matching conditions and acknowledge existing capabilities. Preserve completed source-level analysis separately from unexecuted runtime comparisons. Define the useful contribution without manufacturing novelty or usability benefits. ## Reconcile and report Trace consequential changes across methods, results, limitations, conclusion, abstract and appendices. Retain necessary qualifiers after rewriting. Explain sources that are unavailable and the claims this prevents verifying; avoid a blanket assurance of correctness. Record the inspected evidence, scope, finding, correction or retention, and uncertainty. Save source findings incrementally with their exact editions and claim mappings, reusing verified unchanged evidence while refreshing sources affected by new claims or versions. A clear ledger of unverified items is more useful than an unsupported all-clear. When editing is authorised, fix supported errors and recheck affected prose and references after writing passes and export.