evidence-writing
Skill0 starsChanged 3 months agoApache-2.0View on GitHub
Produce, place, and verify gate evidence that would satisfy an independent assessor. Use whenever recording the result of a gate (G0-G5), writing a verification or assessment record, claiming a control is satisfied, or asked what proves something was done. Also use when reviewing whether existing evidence actually supports the claim it is attached to.
In this repository18 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- ai-impact-assessmentSkill
- artifact-intakeSkill
- backlog-prioritizationSkill
- blameless-postmortemSkill
- control-assessmentSkill
- control-verificationSkill
- dependency-vettingSkill
- eval-designSkill
- exploratory-testingSkill
- red-teamingSkill
- retro-facilitationSkill
- scanner-triageSkill
- secure-code-reviewSkill
- story-splittingSkill
- test-strategySkill
- threat-modelingSkill
- writing-acceptance-criteriaSkill
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 2 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- could be automatic: a setting could enforce it
- Distinctive: line 37
2. Never write evidence you did not produce: Do not describe a scan you did not run, a test you did not execute, or a control you did not …
- Very specific: line 68
Never hand-edit generated evidence: If a scan report or test result is wrong, regenerate it and record why.
- Distinctive: line 119
When checking someone else's claim, verify three things in order:
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Evidence writing
The discipline behind every gate. Consumers: all gate roles, /security-gate, /qa-gate,
/ai-gate, /release, compliance-officer, docs/compliance/evidence-model.md.
The distinction everything rests on
Evidence is an artifact produced by performing a control, sufficient for an independent party to confirm it operated.
| Evidence | Not evidence |
| A workflow run record with results | "CI passed" |
| A SARIF file | "We run SAST" |
| A test result with the assertion visible | "It's tested" |
| A signed attestation | "The build is trusted" |
| A GitHub approval with identity and timestamp | "It was reviewed" |
| A threat model with dispositions | "We considered security" |
| An eval report naming model and prompt versions | "Evals pass" |
| A rollback rehearsal record | "Rollback is supported" |
Prose asserting the work was done is not evidence. This is the line that gets quietly abandoned under deadline, and abandoning it is what turns a compliance posture into paperwork.
Three rules
1. Point at an artifact or say the item is unsatisfied. There is no third option. If you cannot link something, write "unsatisfied — no evidence produced" and stop. That sentence is worth more than a paragraph implying the work happened.
2. Never write evidence you did not produce. Do not describe a scan you did not run, a test you did not execute, or a control you did not verify. A false compliance claim is worse than an open finding — the finding is tracked and will be fixed; the claim is a hole nobody knows exists.
3. Say "unverified", never "passed", for anything you could not check. And report what you did not examine. Overstated coverage is how real gaps survive assessment cycles.
Where it goes
evidence/<issue>/
g1/ dor-checklist.md
g2/ design-note.md threat-model.md control-allocation.md ai-impact-assessment.md
g3/ review-record.md ci-run.json sast.sarif sca-report.json
g4/
qa/ traceability.md test-results.xml exploratory-charter.md nfr-results.md
security/ control-verification.md scan-triage.md findings.md dast-report.html
ai/ eval-report.json guardrail-verification.md redteam-summary.md
evidence/releases/<version>/ change-record.md sbom.cdx.json attestation.intoto.jsonl
approval-record.md deployment-record.md
evidence/incidents/<id>/ timeline.md postmortem.md actions.md
evidence/retros/<id>/ retro.md metrics.json
evidence/assessments/<quarter>/ assessment-report.md sample-selection.md findings.mdIntegrity (AU-9)
- Prefer git-resident — commit history is tamper-evident and attributable via signed commits.
- Prefer CI-generated — a workflow run record cannot be hand-authored after the fact, which is precisely its value over a written claim.
- Never hand-edit generated evidence. If a scan report or test result is wrong,
regenerate it and record why. A hand-edited SARIF is a fabrication regardless of intent,
and
guard-write.shblocks it. - Narrative artifacts (
.mdverification records, charters, assessments) you author yourself — those are legitimate.
What every gate record must carry
Because agents produce most artifacts here, the record must answer a question a human-only process never raises: what produced this, and who is answerable?
# G4 Security Verification — #142
Produced by: security-engineer agent
Model version: <pinned model id>
Date: 2026-08-07
Approved by: <human> on <date>
Issue: #142That, plus signed commits carrying AI-Assisted: trailers and retained transcripts,
satisfies AIC-10 and AU-12. An assessor's real question is "show me how this change came to
be in production, and who is answerable for it" — this makes it answerable in minutes.
Writing a verification record
Structure that survives scrutiny:
## What I verified
AC-3 object authorization → DraftAuthorizer.java:47
Negative test: api_test.py:212 (asserts 403 for a non-owner)
Confirmed the test fails when the ownership check is stubbed out. ✅
## What I could not verify
Performance NFR — no load environment available. UNVERIFIED, not passed.
## What I did not examine
Runtime behaviour under concurrent load. Deferred to /qa-gate.
## Findings
HIGH — drafts.py:88, quota check is non-transactional; two concurrent
autosaves both pass. Reproduce: parallel POST at quota-1.Three sections, always: verified · could not verify · did not examine. The second and third are what make the first credible.
Reading evidence (the assessment discipline)
When checking someone else's claim, verify three things in order:
- Does the control exist? A specific, documented implementation — not an intention.
- Did it operate? Is there evidence it ran?
- Does the evidence support the claim? Read it.
Step 3 is where mature programs fail. Evidence that exists but does not say what the claim says it says is the most common finding in a real assessment, and it survives for years precisely because nobody opens the file. Open the file.
Retention
Per docs/10-definitions.md § Retention. Gate bundles 3 years · findings until closed + 3
years · audit logs 1 year online / 3 archived · agent transcripts 1 year.
Controls
CA-2 · CA-7 · AU-9 · AU-11 · AU-12 · SA-11 · AIC-10.
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.

