CLAUDE.md rules ignored: 74-line file slips past prompt-only agent

The experiment targets a common practice in AI coding agent operations: writing architecture conventions — no cross-layer dependencies, no mutable globals, mandatory return type annotations, a maximum line count per file — as dozens of lines of instructions in CLAUDE.md, .cursorrules, or system prompts. This Prompt-Only approach has a known failure mode, because the model's attention is consumed by implementing functional requirements and soft constraints are probabilistically skipped, especially when a smoke test passes and nothing flags the breach. Mode A demonstrated exactly that: the agent followed local syntactic rules like banning imports and using tuples, but expanded tariff.py to 74 lines, nine over the limit, and submitted confidently because only a functional smoke test ran at commit.

Mode B attacked the problem in two stages. At session start, a deterministic AST audit listed the 14 violations in the initial fixture and told the agent that all five rules, including MAX_RULE5_LINES = 65, would be mechanically blocked at COMMIT_AND_SUBMIT. That hard gate let the agent plan line budgets from Turn 1, compressing intake.py to 35 lines and tariff.py to 60. The result was full compliance without a single retry, with lower output tokens, shorter elapsed time, and lower estimated API cost than Mode A. The practical guidance is to move any deterministically checkable rule into an AST linter or pre-commit hook, inject the initial violation list into the prompt, and reject commits with line-numbered errors — a pattern framed as preventing architecture drift.