Guardrails

This page exists because quality objections are valid. These controls are mandatory on every engagement.

Failure Mode	Control	Evidence
Hallucinated code	Human-authored test required for each AI-authored function	PR checklist + test diff review
Security vulnerabilities	SAST and dependency scanning before merge	Automated check status + review thread
Architecture drift	Pattern-library constrained AI suggestions	ADR review and architecture gate
Over-reliance on AI	AI-to-human ratio tracking with 70% review trigger	PR metadata + architecture review record
Prompt injection risk	Input/output validation, filtering, sandboxing	Threat model and test coverage