Trust the Commit Message? Don't.

The report card written by the student

Most of the software work on my own systems now runs through AI coding agents, and most of those agents leave a note behind explaining what they just fixed. It reads like a report card. There's just one problem worth remembering: it's written by the same agent that did the work, about its own work, before anyone independently checked it.

What happened when I actually checked

During a recent review I stopped taking those notes at face value and checked three of them directly against the real code and the real config data instead. All three were wrong in a way that mattered:

• One claimed a scheduling calculation "now computes a real value" -- true only in the one case that almost never happens, because the edit screen always pre-fills the old value first.
• One claimed a pricing bug was fixed -- the ordering fix was real, but a second factor was still being ignored, silently mispricing a third of the shipped defaults.
• One claimed a status only advances "on genuine success" -- true, except the thing it was calling success didn't meet the bar the business had actually set.

Every one of those would have shipped clean, with a passing test suite, if the commit message had been taken as proof instead of a claim.

The habit this builds

A green test suite and a confident commit message both describe what the author had in mind when they wrote the code -- not necessarily what the business actually needs to be true. That gap is exactly where AI-assisted work needs a human check the most, not the least, because the write-up sounds so reasonable. I don't take my own AI-generated self-reviews as evidence anymore. I take them as a claim, and I go read the actual code before I believe it.