evidence-audit
Skill1 starsChanged 12 days agoApache-2.0View on GitHub
Scope: one claim, not one paper. Reviewing or grading a paper is out of scope, even with its numbers quoted. Adjudicate one specific claim: is it supported, unsupported, or undecidable given the numbers actually available. Use when the input is a claim rather than a paper — a quoted sentence plus its reported figures — and the answer must be a verdict with a stated evidence boundary. Applies three ordered tests (claim anchoring, counter-evidence, availability), a four-tier reversal scale (rank flip / strength collapse / scope narrowing / not a reversal), and the rule that a second number in the same table does not by itself license a second reading. Input: claim text, available numbers with provenance, and the material boundary stating what was and was not read. Output: verdict, tier, attribution as defect versus ambiguity, and the exact citation form allowed. It does not build cards, matrices or proposals; given a whole paper to document rather than a claim to rule on, use paper-card.
6 other projects use a version of this.
Adopted by
Most recently taken up by
- luciferprosun/AIOA-spArkHATtheir own version, first seen Oct 25, 2026
- juanitourquiza/shipframetheir own version, first seen Oct 23, 2026
- Jason-hub-star/model-orchestratortheir own version, first seen Oct 20, 2026
- nino-chavez/agentic-ways-of-workingtheir own version, first seen Oct 13, 2026
- ersinkoc/evidence-audittheir own version, first seen Oct 4, 2026
- crfldrxyz/AuthorOps-Skillstheir own version, first seen Sep 1, 2026
Other versions6 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 6
- ersinkoc/evidence-audit★ 4same name
- juanitourquiza/shipframe★ 4same name
- luciferprosun/AIOA-spArkHAT★ 2same name
- nino-chavez/agentic-ways-of-working★ 2same name
- crfldrxyz/AuthorOps-Skills★ 0same name
- Jason-hub-star/model-orchestrator★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository7 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- literature-radarSkill
- paper-cardSkill
- proposal-forgeSkill
- research-opsSkill
- synthesis-matrixSkill
The file
We split this skill into the rules it gives an agent
2 rules. 0 are common, 2 are distinctive, and 0 are very specific to this project.
0
Common
Most of the 5 edited versions we read keep these.
2
Distinctive
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 47
When later evidence changes a claim's standing, the change has a tier, and the tier is not a matter of emphasis:
- Distinctive: line 62
Every non-trivial ruling also says what kind of problem it is.
“Common” is measured: kept by at least 6 in 10 of the 5 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
evidence-audit
One claim in, one verdict out, with the boundary of the evidence stated in the same breath.
The discipline is entirely in refusing to rule on the cases where the material does not permit it. A verdict of undecidable is a success, not a failure to try.
The three tests, in this order
| Order | Test | Question | If it fails |
| 1 | anchoring | which sentence, in the paper, states this claim | no verdict — the claim is your paraphrase |
| 2 | counter-evidence | does the paper itself contain something that refutes it | unsupported, with the quote |
| 3 | availability | could the available material distinguish the claim from its negation | undecidable, with the missing artefact named |
The order matters: a claim with no anchor cannot be audited at all, and a claim with an anchor plus a direct refutation does not need the availability test.
→ references/three-tests.md
Two rules that decide most cases
- A numeric anchor is not an anchor. If the same table reports another number that would support a second reading, the claim is still a claim: only the paper's own sentence counts as the anchor. Otherwise every paper that reports a full table is automatically immune, and the
undecidablecategory becomes structurally empty. - "Not discussed" is not "not claimed". An absence of discussion is not ambiguity. A factual statement that the paper can be seen to contradict is a defect, not an ambiguity, no matter how little the authors say about it.
→ references/counter-examples.md for the ten rulings that look correct and are not.
Reversal tiers
When later evidence changes a claim's standing, the change has a tier, and the tier is not a matter of emphasis:
| Tier | Name | Test |
| T1 | rank flip | the ordering of two methods reverses |
| T2 | strength collapse | the direction holds, the effect falls inside the noise band |
| T3 | scope narrowing | holds inside a stated boundary, fails outside |
| T0 | not a reversal | restated, re-measured, or re-scoped without changing the claim |
Count after defining, never before. A count of "reversals" produced before the tiers exist overstates: in the one corpus where this has been checked, five of twelve findings moved out of the reversal column (12 → 7).
→ references/reversal-tiers.md
Attribution
Every non-trivial ruling also says what kind of problem it is. Four names, used verbatim:
| Name | Means |
defect-self-data | the paper contradicts its own reported data |
defect-external-fact | the paper contradicts a checkable outside fact |
ambiguity-two-readings | one sentence supports two anchored readings |
ambiguity-term-overload | one term carries two meanings |
The first word decides who fixes it: a defect needs new data or a correction, an ambiguity needs one sentence. Confusing them sends a defect to the authors as a wording question, and an ambiguity back as an error.
→ references/attribution.md
What this skill refuses
- Refuses to rule on a paper. That is
paper-card: a card documents, this rules on one claim. - Refuses to rule without an anchor. A paraphrase is not auditable, and rewriting the claim until it is auditable is how audits produce their own answers.
- Refuses to convert its output into a blocker level or a repair priority. Verdicts live here; stopping decisions and repair order belong to the pipeline's operating layer.
Output contract
claim: "<quoted, with §>"
anchor: "<the sentence that states it, with §>"
verdict: supported | unsupported | undecidable
basis: <the test that decided it, and its result>
tier: T1 | T2 | T3 | T0 | n/a
attribution: defect-self-data | defect-external-fact | ambiguity-two-readings | ambiguity-term-overload | n/a
boundary: <material read, and what was not>
allowed: "<the exact citation form this verdict permits>"The boundary line is not optional: a verdict without its evidence boundary reads as absolute.
Routing
| Task | Go to |
| run the three tests on a claim | references/three-tests.md |
| classify how a claim's standing changed | references/reversal-tiers.md |
| decide defect versus ambiguity | references/attribution.md |
| tag the problem for the register | references/m-taxonomy.md |
| check a ruling against known failure shapes | references/counter-examples.md |
| document a whole paper | paper-card |
| compare several papers | synthesis-matrix |
| act on a verdict inside a proposal | proposal-forge |
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.