Hard, deterministic
rules catch drift.

violations of deterministic code-quality rules (lint, typecheck, format; e.g. Biome, ESLint, SwiftLint) cannot reach the protected branch undetected. Enforcement happens at multiple checkpoints with the same ruleset across layers, so no surprise surfaces between them, and bypassing an earlier checkpoint is caught by a later one

a single test-location convention applies repo-wide (colocated with source, parallel-tree, or other), consistently applied so tests are predictably findable for humans and agents; coverage is enforced via global target *and* per-PR differential threshold

tests are verified to actually catch bugs, not just exercise lines. Mutation testing (Stryker for TS, Muter for Swift) or equivalent mechanism confirms that tests assert behaviour, not merely execution

static and dynamic application security testing with agent-actionable findings; findings, suppressions, and rule disables carry accountability (rationale, named reviewer, expiry where applicable). Tool choice is project-dependent (e.g. Aikido, SonarQube for compliance cases); the concern is coverage across both testing classes, not a specific vendor

Aikido blocks new leaks *and* historical secrets are rotated / cleaned

human glance *or* multi-model review (e.g. Claude Code Reviews $25/PR council-of-experts model)

humans (not agents) test for taste / UX / defect discovery, not just correctness. Delivered either via a dedicated usability research service (Netizen Experience, UserTesting, Lookback, Maze, PlaybookUX) or via your own user base through canary releases and A/B tests. Catches what automated tests can't: is this confusing? Does this feel wrong? Would a real user hit this edge case?

every agent decision is logged with reasoning, retrievable, and reversible at granular level (catches *quiet drift* — subtle wrong actions humans don't notice for weeks)

capability exists *and* is actually exercised. Least frequent of validation types — typically weekly or on major releases