Adversarially Re-Check Your Own Analysis Before You Trust It

When you produce an audit, design review, or confidence report, your first draft is a hypothesis, not a verdict. The cheapest way to find out how wrong it is: run a dedicated pass whose only job is to refute each claim by reading the ground truth again. A recent self-review scored a first draft at roughly 58/100 — and the errors weren’t cosmetic, they were load-bearing.

The most dangerous mistakes invert a safety-critical outcome. In one case a rule meant to flag “idle/done” was ranked above the rules that detect a blocked state, so a system waiting on a destructive confirmation prompt got classified as finished and released early. Priority ordering in any rules engine deserves an explicit test: any blocking signal anywhere must dominate any “all clear” signal, never the reverse.

Also beware citing a reference implementation you haven’t re-verified. It’s easy to describe your design as faithful to some proven original when it actually mirrors an approach that original has since abandoned. Verify against the current source, and label your own inventions as your own.

Practical habit: graph the ground truth instead of hand-tracing it, add a regression test the moment you fix an inverted classification, and treat “what survived the challenge” as the only part you can quote with confidence.