name: cyber-exercise-adjudication
guardrail: true
description: Adjudicates authorized exercise outcomes against independently held ground truth, redacted evidence, and explicit status boundaries. Use when comparing observations, resolving discrepancies, or preparing evidence-led security-review handoffs.
metadata:
labels: [cybersecurity, adjudication, white-team]
triggers:
keywords: [exercise adjudication, independent ground truth, outcome adjudication, evidence discrepancy, exercise finding status]
Cyber Exercise Adjudication
Priority: P0 (CRITICAL)
Compare evidence to pre-held expectations without self-approved promotion or compliance claims.
Structure
cyber-exercise-adjudication/
├── SKILL.md
├── references/adjudication-record.md
└── evals/evals.json
Workflow
- Confirm adjudicator independence from operator and evidence collector.
- Load approved objective, scope, expected observation, and ground-truth version.
- Compare redacted evidence by time, source, and limitation; record discrepancies.
- Assign
confirmed, suspected, blocked, not-tested, or false-positive only with rationale.
- Preserve unresolved conflicts and route them to an accountable owner.
- Emit sanitized findings to
artifacts/security-review.md when that artifact is in scope.
Rules
- Independent ground truth is held before observation review; do not rewrite it to fit results.
- A blocked runtime or missing evidence remains blocked, not passed.
- Adjudication describes exercise evidence; it does not certify compliance or measure real-world efficacy.
- Production changes, credential use, and live target interaction remain separately gated.
Adjudication records carry
engagement_scope_ref, skill_version, source, observed_at, finding_status, evidence_refs, limitations, and accountable_owner.
Anti-Patterns
- No operator grading: Separate execution from adjudication.
- No pass by silence: Missing observation is not success.
- No certainty inflation: Keep suspected and not-tested statuses.
- No raw leakage: Redact before handoff.
References