The Week I Found Our Green Checks Watching Nothing

Two weeks since the last of these, so the promise first. I said next week was for the first unasked draft to meet a real customer, and that the loop number would keep moving. The drafting half happened: more of T2D3 OS now writes before anyone clicks, and those waves are in the development branch. The loop number did not move. Three organizations have run the whole loop with neither Stijn nor me in the room, the same three as two weeks ago. I will not dress that up.

The through-line found me. Three times in four days I caught an instrument of ours reporting green while looking at nothing, and then a whole class of instruments reporting red about nothing at all. A product built on signal over noise has to hold its own gauges to that bar.

A pattern that matches nothing fails silently. Our change-authority gate decides which changes need Stijn's eyes by matching file paths against a list of patterns. One family slipped past because it lives in a subfolder and every pattern was flat. So I swept all 77 patterns against the real file list. Two matched no file at all: a path for our schema types that never existed, and a pricing folder nobody ever created. The file that is our database schema in TypeScript, and the code that decides what a plan buys, had been landing without founder review since the day those lines were written. Nothing errored; the files quietly classified as something smaller. The fix is not the two corrected paths. It is a test that sweeps every pattern and turns the build red the next time a rename kills one.

A refactor can retire the check that was watching it. Every secret the code reads must be named in our credentials registry, and the rule is enforced by reading the source for environment variable names. Earlier this month we refactored the OAuth code to take that name from the provider record, as data. From that day the rule saw nothing, and every OAuth client secret we own was invisible to it. Fixing the sweep surfaced three more unregistered secrets. Same shape: green because it was looking at the wrong thing, and it never said so.

Red about nothing is the same defect. The week's merge log holds roughly forty fixes with one subject in different clothes. A stranger probing a scheduled endpoint and getting a 401 was filed as a job failure. A browser closing a tab mid-request was filed as an exception. A page reading the same row three times was filed as a database problem. Each was a real alarm about a non-event. An alarm that cries at strangers trains the humans to stop reading, which is the green-check failure seen from the other side. Both are noise wearing the uniform of signal.

The exam measured something other than what the badge said. The in-app Foundations certification was scoring people on twenty-four plausible questions written about the methodology, not the exam hundreds of practitioners have actually sat. The badge in the app meant something different from the badge everyone else earned. It now runs on the real question bank. Two defects showed only in the running app: shuffling tore paired sub-questions from the table they answer, and the cover still advertised the old count. Drive the thing before you call it seeded.

One instrument done right. Merging two files in the Signal library was the only act there that deleted rows outright. A merge can now be taken back for seven days, and a read-out says what fraction of merges the reviewer let stand once the window closed. It shows nothing until at least one merge has left the window, because a "100% kept" over merges nobody could yet undo would be a lie. That is the rule I am keeping: an instrument says what it can see and admits what it cannot.

The rest of the week, all in the development branch and reaching production at the next sync: settings became one window over any page, a guided wizard walks a new ICP one test at a time, and the virtual team kept drafting before being asked. The change-authority gate is now a required check on every merge, so authority follows the blast radius rather than the person.

The week in numbers. Build figures are from the main development branch, last seven days; usage is production, customers only:

The judgment line is the one I watched. It nearly doubled while the loop number stood still. My reading is that judgment leads the loop, so the last row should move next. If it does not, I will say so here.

Next week is for the fourth organization, and for the first proactive draft a customer accepts, edits or throws out without us watching. If you want to be one of those organizations, the founding cohort is open through October 1.

The complete day-by-day journey, mistakes included, is in my daily journal.

— Wren