The only rule that worked is one a machine could check

I tested my own engineering rules against scenarios built to break them, and one comparison came back so lopsided it reorganized how I write rules. The concrete instruction 'reproduce the failure before editing' produced a reproduction in three of three runs; without the rule, zero of six. The same underlying idea phrased as a principle — 'root cause before editing' — produced zero reproductions. Identical intent, opposite outcomes.

The difference between the two phrasings is checkability. 'Reproduce the failure' names a specific action with an observable artifact — either a repro exists in the transcript or it doesn't. 'Root cause before editing' names a virtue. The model can believe itself compliant with a virtue while doing nothing differently, and apparently it does.

The same study kept me humble about the rest of my corpus. Three other rules I believed in came back as ceiling effects: the failure modes they predicted never occurred in any arm, rules on or off, so the scenarios could not discriminate — no harm demonstrated, but no value either. A rule you've never seen fail without is a rule you know nothing about.

The limits are real: one author evaluating his own rules, one project, no external replication. The design fought the bias where it could — blinded grading, scenarios written to falsify — but this is a finding about my corpus, offered as a method more than a result.

The method transfers in ten minutes: take one rule you currently phrase as a principle, rewrite it as a concrete trigger and action with an observable artifact, and run the same task under both phrasings. If you can't tell from the transcript which run had the rule, that is itself the finding.