evidence-generation

name	Evidence Generation
description	Use this skill when generating ISO 27001 or NIST SP 800-53 audit evidence packs, compliance reports, evidence narratives, reviewer-ready control matrices, or when the user asks about audit evidence, compliance evidence, evidence packages, audit documentation, or ISO/NIST evidence.
version	1.1.0

Enterprise Evidence Pack Generation

Purpose

Transform Shinsa assessment artifacts into an enterprise-grade evidence pack for Security and GRC reviewers. The pack must be defensible, reviewable, exportable, and honest about uncertainty. It supports an audit or control review; it does not replace a human auditor.

Inputs

Read persisted artifacts first:

• assessment-plan.md
• scope.md
• applicability.json and applicability.md
• domains/*.json and domains/*.md
• reviews/round-*/*.json and reviews/round-*/*.md
• synthesis/evidence-index.json
• synthesis/control-matrix.json
• shinsa-state.json when present

Do not rescan the repository during synthesis unless the orchestrator explicitly asks for a reconciliation pass. The evidence pack should explain what the persisted artifacts support.

Required Report Sections

The canonical report is shinsa-output/runs/<assessment_id>/synthesis/compliance-report.md. It must include these top-level sections:

1. ## Assessment Metadata
2. ## Executive Summary
3. ## Control Matrix
4. ## Findings
5. ## Evidence Index
6. ## Reviewer Notes
7. ## Unresolved Risks
8. ## Limitations
9. ## What To Do Next
10. ## Human Sign-Off

Assessment metadata must include:

• timestamp placeholder or ISO-8601 timestamp
• plugin version
• target path
• target commit placeholder or commit hash
• assessment mode
• standards assessed
• scope exclusions
• methodology
• raw artifact references

Per-Control Requirements

Every control row and narrative must include:

• control ID and title
• status
• confidence score
• confidence_rationale
• evidence_quality
• evidence_quality_rationale
• manual_evidence_needed
• manual_evidence_items
• reviewer_disposition
• remediation priority
• ticket-ready action
• grc_action: accept, reject, request_evidence, or create_remediation_ticket
• evidence anchors with file paths and line numbers where code/config evidence exists

Do not mark a control implemented if manual_evidence_needed is true. Use partially_implemented, not_implemented, or not_assessed until the missing manual evidence is supplied.

Evidence Quality Rules

• strong: direct source/config evidence supports the claimed control outcome, findings have file and line anchors, and no manual evidence is required for the claimed status.
• partial: evidence supports part of the control, but implementation gaps, reviewer concerns, or manual evidence needs remain.
• inferred: the claim depends on framework convention, absence of contrary evidence, or indirect evidence.
• missing: no reliable evidence was found.

Never use strong for a hybrid control when manual policy, approval, access review, incident, supplier, training, or production records are needed for full compliance.

Manual Evidence Taxonomy

Use specific manual evidence items. Avoid vague text like "policy evidence needed".

Common manual evidence categories:

• policies and standards
• access reviews and privileged access approvals
• training records
• incident tickets and post-incident reviews
• vulnerability scan exports and remediation tickets
• supplier due diligence and contract clauses
• production configuration exports
• SIEM/log-retention evidence
• change approvals and release records
• risk acceptances and exceptions

Tone And Uncertainty

Use factual, audit-support language:

• Prefer "evidence indicates", "the repository shows", and "the artifacts support".
• Avoid absolute claims such as "compliant" unless all required evidence is present.
• Separate code/config evidence from manual/process evidence.
• State limitations plainly.
• Keep GRC next actions operational: accept, reject, request named evidence, or open a remediation ticket.

Reviewer Propagation

Reviewer concerns must not be buried. If any reviewer status is changes_requested or unresolved:

• mark affected control rows with the reviewer disposition
• include the concern in ## Reviewer Notes
• include unresolved items in ## Unresolved Risks
• make the grc_action reflect the unresolved state

Good Evidence

Good evidence is anchored, scoped, and explains why it matters:

`src/routes/auth.ts:7` shows the login handler executes password verification without rate limiting middleware. This supports a high-severity A.8.5 gap because repeated failed login attempts are not throttled in the assessed source.

Bad evidence is vague or unsupported:

Authentication looks secure because the app uses a common framework.

Gold-Standard Control Example

| Control | Status | Confidence | Evidence Quality | Manual Evidence | Reviewer | GRC Action |
| --- | --- | --- | --- | --- | --- | --- |
| A.8.5 Secure Authentication | partially_implemented | 0.86 | partial | yes | approved | create_remediation_ticket |

Confidence rationale: Password hashing and session cookie evidence were present in source, but lockout, MFA policy, and access review records were not available.

Evidence quality rationale: The source anchors directly support bcrypt password verification and show missing rate limiting. The control remains partial because manual identity governance evidence is required.

Manual evidence items:
- Production MFA policy
- Access review records for privileged accounts
- Login lockout configuration export

Ticket-ready action: Add login throttling and collect production MFA/access-review evidence before claiming full implementation.

Final Checklist

• All required sections are present.
• Assessment metadata names the target, mode, standards, scope exclusions, and raw artifacts.
• Every control has all enterprise trust fields.
• Every finding has file and line evidence.
• Manual evidence gaps are explicit and actionable.
• Reviewer unresolved items appear in Reviewer Notes, Unresolved Risks, and affected control rows.
• Human sign-off states Shinsa is assessment support, not auditor replacement.