Appendix D - 3. The semantic-lint stack

Concept

Move a recurring judgment out of review and into a deterministic check that fires at commit time. The stack is what turns "an agent keeps re-introducing this class of mistake" into "the class is now impossible to commit." It has two parts that must travel together: the check itself, and the discipline of writing the check at the level where the invariant actually lives — not one level too shallow, where it fires on syntax and misses meaning.

Mandatory members

• Appendix C - 8. Blocking semantic lints — a blocking check that reads structure and rejects a violation, not a regex over surface text. This is the enforcement muscle: without it the invariant is only a convention, and conventions decay under a fleet.
• Appendix A - 28. Enforce at the right semantic level — the discipline of enforcing at the right semantic level. A lint aimed one level too low passes on a spec-legal variation it should have caught and fires on a legal one it should have allowed. The check is only as good as the level it targets, so choosing the level is not optional polish — it is what makes the check correct.

Complementary members

• Appendix A - 21. CLAUDE.md rule index (the governance document as a mechanism) — a terse index of the standing rules the lints enforce, so an agent can find the rule behind a red check. Aids comprehension; the check blocks with or without the index.
• Appendix A - 3. Dynamic context injection — inject the relevant rule text into an agent's context before it writes the code, so it avoids the violation in the first place. This layers prevention on top of the detection the lint provides — a strict addition, never a replacement for the blocking check.