Reproduce, Then Fix

A fix you can't watch fail is a guess, and a fix you can't explain is a coincidence. Both regress. The discipline that prevents this is one loop, applied to every bug regardless of size: reproduce red → minimize → explain the mechanism → fix the cause → watch red turn green → keep the repro. Each step exists because skipping it produces a specific, well-known failure: patches that mask symptoms, "fixed" bugs that were never the bug you saw, and regression tests that would pass with the bug still present.

The loop

Six steps, in references/the-loop.md: reproduce red, minimize, explain the mechanism, fix the cause, watch the same red turn green, keep the reproduction.

Three of them are the ones people skip, and each skip has a signature:

• Red first. A test written after the fix has never been seen failing, so nothing establishes it would have caught this. If it cannot be made to fail against the unfixed code, it is not a reproduction.
• Explain the mechanism. A fix that works without an explanation is a coincidence you have decided to trust. Until you can say why the wrong behaviour happened, you cannot say the fix addresses it rather than perturbing it.
• Fix the cause, not the observable. Suppressing the symptom — a catch, a retry, a guard on the value that was already wrong — closes the ticket and leaves the defect, now harder to find.

Certifying the red