By name: review-architecture description: Review planned or implemented architecture for software engineering, ML, analytics, data science, AI systems, hardware-integrated systems, and enterprise/system designs using ISO/IEC/IEEE 42010, SEI ATAM, ISO/IEC 25010, NIST SSDF, OWASP ASVS/SAMM, NIST AI RMF, and NASA/JPL Power of Ten. Use before implementation, before large refactors, before publishing, or when assessing maintainability, testability, reproducibility, security, safety, and compliance-style readiness.
Review Architecture
Use this skill as an architecture review gate.
Read ../../references/governed-engineering-framework.md for substantial reviews. Default to adversarial mode when the design touches security, AI/ML, data correctness, safety, hardware, public APIs, or production operations.
Review Focus
Prioritize findings over praise:
- ISO 42010 gaps: missing stakeholders, concerns, architecture views, rationale, interfaces, or correspondences.
- ATAM risks: quality attribute scenarios that fail, sensitivity points, tradeoff points, and unstated assumptions.
- ISO 25010 quality gaps: functional suitability, performance, compatibility, interaction/usability, reliability, security, maintainability, portability, flexibility, or safety.
- Correctness and data leakage risks.
- Broken or missing contracts between data, model, API, UI, hardware, and operations layers.
- Hidden state, global configuration, and irreproducible execution.
- Excess coupling, oversized modules, and abstractions without payoff.
- Package compatibility, supply-chain, deployment, and rollback risks.
- Security and privacy: credentials, unsafe deserialization, untrusted files, PII, shell execution, network calls, trust boundaries, and authorization.
- AI/ML risk: dataset lineage, evaluation validity, drift, monitoring, explainability, fairness, model rollback, prompt/data injection, and human oversight.
- Test strategy: unit, integration, property, notebook, smoke, adversarial, load, and operational checks.
- Performance and scale: memory growth, unbounded loops, batching, streaming, and parallelism.
Output Shape
Lead with issues ordered by severity. Include a framework anchor for each issue. End with open questions, residual risks, and a gate decision: pass, pass with conditions, or block.