# Exact unsent Grok prompt — EM-AI-03

Independent enterprise metamodel review. Do not browse, invent identifiers or claim publication readiness.

EM-AI-03 covers EvaluationPlan, EvaluationRun, BenchmarkSpecification, SafetyAssessment and DeploymentDecision. Complete adjacent drafts cover WM-AI-003 Evaluation, WM-AI-009 Benchmark, WM-AI-008 Safety Assessment, WM-AI-010 Incident Report, WM-SFT-004 Artifact, WM-DAT-001 Dataset, WM-MAT-008 Observation and the decision triad.

Assess this proposal: keep applied plan, run and result inside one WM-AI-003 evaluation instance; reuse WM-AI-009 for Benchmark Specification; reuse WM-AI-008 for Safety Assessment; profile Deployment Decision across rationale, authorized act and issued instrument. Leave reusable Evaluation Protocol and standing Deployment Authorization Instrument as conditional identifier-unassigned candidates only when they have independent lifecycles. Allocate no other identifier.

Separate plan from run and result; benchmark contract from dataset snapshot; capability measurement from safety judgement; assessment from deployment authorization; decision content from act and instrument; incident report from incident evidence. Require pinned artifact, configuration, benchmark, dataset, metric, scorer, environment, context, intended use, population, uncertainty and limitations.

Test a changed benchmark dataset and deployment context that requires new evaluation, assessment and explicit authorization. Reject carrying the old approval forward or treating one high score as universal safety proof.

Return <=900 words with: Verdict; strongest evidence; strongest counterexample; identity/mastership; plan/run/result; benchmark/dataset; measurement/uncertainty; safety assessment; deployment decision; contamination/context drift; incident feedback; time/version/reassessment; scenario; at least 10 invariants; minimum model set; blockers. Explicitly decide whether either conditional candidate needs independent identity.