← ClaudeAtlas

evidence-and-claims-standardlisted

Claim grading: evidence class, provenance, strength, promotion boundaries.
SylphxAI/skills · ★ 1 · AI & Automation · score 74
Install: claude install-skill SylphxAI/skills
# Evidence and Claims Standard Make the strongest claim current evidence supports, and no stronger. Read [references/claim-evidence-method.md](references/claim-evidence-method.md) before assessing a material or disputed claim. Use [references/claim-dependency-revision.md](references/claim-dependency-revision.md) when premises change over time, and [references/assurance-cases.md](references/assurance-cases.md) for a high-impact multi-layer claim. ## Method 1. State the claim precisely enough to be falsifiable. 2. Split compound claims and separate source, review, checks, merge, release, deploy, and live behavior. 3. Identify the authority, subject, revision, observation time, and evidence required for each claim. 4. Inspect current primary evidence; treat memory, summaries, agent consensus, and self-authored status fields as leads only. 5. Test plausible contradiction and staleness paths. 6. Assign `supported`, `refuted`, `unknown`, or `partially supported`. 7. Report residual uncertainty and the narrowest next evidence that could change the verdict. Do not turn unavailable evidence into success or failure. A timeout, missing permission, absent observation, or unrun check means unknown unless the authoritative contract proves otherwise. ## Completion claims Before saying `done`, `fixed`, `converged`, `shipped`, or an equivalent: - bind the terminal condition before evaluating it; - verify the exact candidate or live subject named by the claim; - require evid