How I work: verify first, change second

Phase 1 is read-only

I measure before I touch anything. The output is a table: what changes, at which line, how it will be verified in production afterwards, and what could regress.

You approve the table, not the outcome.

Two inputs that must differ

Reading code is not testing code. Before I quote a number, I run the instrument on two inputs that should disagree and read both outputs.

If they agree when they should not, the instrument is suspect and I say so before using its number. That rule has caught false figures more than once, and some of them were mine.

Phase 2: approved changes, one at a time, each with its measurement

Every change ships with the check that backs it. And the check has to run where it can still fail: a guard placed after a cleanup is not a guard, it is decoration.

What I never do without approval

No publishing, no edge deploys, no visible copy changes. If a markup change forces a change on screen, I stop and ask.

Anything I could not measure ships as pending with the reason, not as an assumption.