How I ship AI-built software without breaking things.

See it run

Home, desktop 1280
  PASS  the page answers 200
  PASS  the three margin counts are git's
  PASS  no employer or client name
  PASS  a 620px column with the margin notes
        56px to its right
  PASS  home 1280: no sideways scroll

Home, phone 320
  PASS  the margin notes fold under the column
  PASS  a 20px gutter on both sides

RUNTIME ERRORS
  none. zero pageerror, zero console error,
  zero failed request

VERDICT  62/62 checks passed
WALKED. Every step performed, every
Expect observed.

The path to production

The choices, and why

1. One script is the only way to deploy.

   Early on, changes went straight to production because nothing stopped them. Now every site ships through one script. Staging is the default. Production needs a flag that names the person who approved it, and the script refuses without one.

   Aug 112026, the day the script was written

2. The walk only counts what a person could see.

   A test once passed on a panel that was on the page but hidden. Every text check found its words, and no human could have. So the walker reads a piece of text only if nothing above it on the page is hidden. Hidden text reads as missing.

   Aug 282026, a hidden panel passed every check

3. One error in the browser fails the whole walk.

   A click handler shipped with a mistake that only shows when you click it. The code looked fine to every check that reads code. So the walker listens while it works. One page error, one console error or one failed request fails the walk, even when every check passed.

   Aug 152026, a broken click handler got through

4. Walk the build you just made, not its nickname.

   The staging address is a nickname that points at the newest build. Once it kept pointing at an older build for several minutes, and a walk passed against code that was not there. The script now saves the exact address of the build it just uploaded, and the walker goes to that one.

   Aug 282026, a green walk on the wrong build

5. The script checks more than the code.

   It will not go live while a task for the change is still open, or while a test walkthrough written in the last day has not been walked. A deploy is the last step of the work, so the script checks that the earlier steps happened.

   13refusals in the script today

What broke

1. A fix was reported live for 30 hours and was not running.

   My background programs load their code once, when they start, and then run for days. I changed the code, saw it saved, and called it live. The old code kept running the whole time. Now a backend change ends with a restart, and the script prints the new start time and names any file newer than the program.

   Aug 152026

2. A deleted page stayed up at the address people use.

   I removed a page, deployed, and checked the new build. It was gone. At the public address it was still there, held in a cache set to keep it for seven days. So the last check is always the address a person would type.

   Aug 282026

3. A finished walk kept running for 17 minutes.

   A walk that opened six browsers closed only the last one. It had printed its result and still would not exit, and everything waiting on it sat silent. The walker now tracks every browser it opens and closes them all, and a walk still running after 15 minutes is stopped and marked as failed.

   Oct 72026

What I would do next