Claim Verification
Turn each material claim into a falsifiable statement, then verify it against the authority capable of proving it.
Read document reconciliation when comparing a current plan, README, runbook, or specification with implementation.
Workflow
- Preserve the claim's original meaning. Split bundled claims into independently testable statements.
- Identify the required evidence class:
- repository state for file, code, test, and configuration claims
- Git hosting state for PR, review, check, and merge claims
- provider state for deployment, configuration, traffic, and acceptance claims
- runtime observation for user-visible or production behavior
- Inspect current evidence directly. Do not use a PR description, stale local checkout, generated summary, or curated repository page as proof of a different authority.
- Classify each claim as
CONFIRMED, REFUTED, PARTIAL, or UNVERIFIABLE.
- Separate verified work that must not be redone from actual gaps.
- Recommend the smallest action supported by the evidence. Do not implement it unless the user also asked for a change.
A supplied claim and receipt can support a bounded verdict without a checkout.
If only A's passing receipt and the existence of revision B are supplied, report
B as unverified. Do not invent a repository, run checks, or require access merely
to state that evidence limit.
Evidence rules
- Cite files and lines for repository claims when practical.
- Record exact branches, revisions, check names, provider resources, timestamps, or URLs when they determine the verdict.
- Treat passing static checks, merge, deploy, live traffic, and correct behavior as separate facts.
- Say what could not be checked and why. Never convert missing access into a passing verdict.
- Re-read drift-prone evidence immediately before a conclusion that depends on it.
- Bind each result to its checked revision and inputs. A result at A does not verify B unless a justified scope comparison establishes that the relevant checked inputs are identical; label that comparison separately from execution at B.
- Keep
not run, not supplied, not found in the inspected scope, and does not exist distinct. Missing evidence cannot establish a failing test, absent coverage, or a completed action.
- Before calling evidence unavailable, check every location the request points to and that you may read, not only the working directory. Report failed reads and what you did not search; do not expand access to unrelated files or credentials.
- Separate the checkpoint described by a receipt from the current task. A record that an earlier step used no model calls says nothing about the current step. Not using a network tool does not prove the client made no network calls or spent nothing. Count sessions, delegated requests, and evaluation runs separately when known; do not invent call counts.
- Bind local edits, committed contents, PR scope, and remote state separately. A partial readback supports only the inspected portion; report the rest as unchecked.
- Similar names do not prove that two resources, branches, or environments are the same. Confirm identity by identifier, URL, or direct inspection.
- When the user disputes a verdict and current evidence still supports it, state the evidence once and record the disagreement. The verdict changes only with new evidence, while the user's direction still decides the next action.
Output
Lead with the supported verdict, then a claim table containing the claim,
verdict, authority, exact evidence, and freshness. Preserve confirmed work and
include actual gaps, unresolved evidence, and a next justified action only where
applicable. Omit empty categories; do not invent remediation when all requested
claims are confirmed.
Read the original acceptance example when checking this behavior.
Example
Synthetic example; not an observed client result.
Input: A supplied receipt records 12 passing unit tests at synthetic revision B, observed on 2026-10-20. No serving revision or runtime evidence is supplied. Is the fix live?
Output: The receipt confirms local checks at B; live behavior is unverified.
| Claim |
Verdict |
Authority |
Evidence |
Freshness |
| The supplied receipt records passing unit tests at B |
CONFIRMED |
Local test receipt |
12 passes at B |
Observed 2026-10-20; applies to B |
| The fix is live |
UNVERIFIABLE |
Serving configuration and runtime |
Neither was supplied |
Currentness unknown |
Preserve the local result. Next obtain the serving revision and an authorized runtime observation; the receipt alone cannot establish either.
Input: An old receipt records all checks passing at A. The current revision is B,
and no B receipt or input comparison is supplied. Is the work verified at B?
Output: UNVERIFIABLE at B. The supplied receipt remains historical evidence for A;
no result at B or relevant input-equivalence review was supplied. Next obtain a
B-bound receipt or inspect the changed inputs within the authorized scope. No
check has been run by this review, and the existence of tests is unknown.
1---2name: verify-claims3description: "Verifies completion and check results when asked whether work is done, a fix is live, or an old receipt still applies after a change from revision A to B; keeps current proof, historical evidence, and missing checks distinct. Loads before file inspection when the request matches."4license: MIT5metadata:6 source: https://skillmd.com/skills/jovanipink/verify-claims7 author: "Measured Studios"8 version: "0.19.0"9 plugin: "measured-quality"10 invocation: "implicit"11 provenance: "adapted"12 risk_class: "read-only"13---1415# Claim Verification1617Turn each material claim into a falsifiable statement, then verify it against the authority capable of proving it.1819Read [document reconciliation](references/document-reconciliation.md) when comparing a current plan, README, runbook, or specification with implementation.2021## Workflow22231. Preserve the claim's original meaning. Split bundled claims into independently testable statements.242. Identify the required evidence class:25 - repository state for file, code, test, and configuration claims26 - Git hosting state for PR, review, check, and merge claims27 - provider state for deployment, configuration, traffic, and acceptance claims28 - runtime observation for user-visible or production behavior293. Inspect current evidence directly. Do not use a PR description, stale local checkout, generated summary, or curated repository page as proof of a different authority.304. Classify each claim as `CONFIRMED`, `REFUTED`, `PARTIAL`, or `UNVERIFIABLE`.315. Separate verified work that must not be redone from actual gaps.326. Recommend the smallest action supported by the evidence. Do not implement it unless the user also asked for a change.3334A supplied claim and receipt can support a bounded verdict without a checkout.35If only A's passing receipt and the existence of revision B are supplied, report36B as unverified. Do not invent a repository, run checks, or require access merely37to state that evidence limit.3839## Evidence rules4041- Cite files and lines for repository claims when practical.42- Record exact branches, revisions, check names, provider resources, timestamps, or URLs when they determine the verdict.43- Treat passing static checks, merge, deploy, live traffic, and correct behavior as separate facts.44- Say what could not be checked and why. Never convert missing access into a passing verdict.45- Re-read drift-prone evidence immediately before a conclusion that depends on it.46- Bind each result to its checked revision and inputs. A result at A does not verify B unless a justified scope comparison establishes that the relevant checked inputs are identical; label that comparison separately from execution at B.47- Keep `not run`, `not supplied`, `not found in the inspected scope`, and `does not exist` distinct. Missing evidence cannot establish a failing test, absent coverage, or a completed action.48- Before calling evidence unavailable, check every location the request points to and that you may read, not only the working directory. Report failed reads and what you did not search; do not expand access to unrelated files or credentials.49- Separate the checkpoint described by a receipt from the current task. A record that an earlier step used no model calls says nothing about the current step. Not using a network tool does not prove the client made no network calls or spent nothing. Count sessions, delegated requests, and evaluation runs separately when known; do not invent call counts.50- Bind local edits, committed contents, PR scope, and remote state separately. A partial readback supports only the inspected portion; report the rest as unchecked.51- Similar names do not prove that two resources, branches, or environments are the same. Confirm identity by identifier, URL, or direct inspection.52- When the user disputes a verdict and current evidence still supports it, state the evidence once and record the disagreement. The verdict changes only with new evidence, while the user's direction still decides the next action.5354## Output5556Lead with the supported verdict, then a claim table containing the claim,57verdict, authority, exact evidence, and freshness. Preserve confirmed work and58include actual gaps, unresolved evidence, and a next justified action only where59applicable. Omit empty categories; do not invent remediation when all requested60claims are confirmed.6162Read [the original acceptance example](references/continuity-example.md) when checking this behavior.6364## Example6566Synthetic example; not an observed client result.6768Input: A supplied receipt records 12 passing unit tests at synthetic revision B, observed on 2026-10-20. No serving revision or runtime evidence is supplied. Is the fix live?6970Output: The receipt confirms local checks at B; live behavior is unverified.7172| Claim | Verdict | Authority | Evidence | Freshness |73| --- | --- | --- | --- | --- |74| The supplied receipt records passing unit tests at B | CONFIRMED | Local test receipt | 12 passes at B | Observed 2026-10-20; applies to B |75| The fix is live | UNVERIFIABLE | Serving configuration and runtime | Neither was supplied | Currentness unknown |7677Preserve the local result. Next obtain the serving revision and an authorized runtime observation; the receipt alone cannot establish either.7879Input: An old receipt records all checks passing at A. The current revision is B,80and no B receipt or input comparison is supplied. Is the work verified at B?8182Output: UNVERIFIABLE at B. The supplied receipt remains historical evidence for A;83no result at B or relevant input-equivalence review was supplied. Next obtain a84B-bound receipt or inspect the changed inputs within the authorized scope. No85check has been run by this review, and the existence of tests is unknown.