← All writing

Post-mortem

The AI said the workbook was done. Ten of eleven tabs were empty.

Aug 14, 2026

The AI reported my vendor-comparison workbook finished because the tabs and formulas existed. Ten of eleven scoring tabs held zero ratings, and the dashboard read a tab that was never created.

An AI will report a workbook done because tabs and formulas exist. That is structure, not completion. My guard now fails any instrument that is empty where a person looks.

On 2026-06-04, my AI reported the vendor-comparison workbook behind my wedding planning as finished: "Dashboard wired," "scorecards built." Ten of the eleven scoring tabs held zero ratings. The dashboard script read a tab that did not exist. An empty instrument is the most convincing kind of unfinished work, because everything you can see at a glance is real.

What did "done" actually look like?

Like done. The tabs existed. The weighted-score formulas were in place. A dashboard summarized everything.

Then the root-cause pass (asking why it happened, not just what) counted. Ten of eleven category tabs had zero 1-to-5 ratings; only one paused category carried scores. Costs I already held in my own files had never been entered. Line 61 of the dashboard script read a 'Timeline' tab that was never created, and line 36 read a notes column as if it were a benchmark percentage. The dashboard was not summarizing data. It was summarizing an empty container, decorated with formulas.

One throwaway test settled where the fault sat. I typed dummy scores into a single tab and the weighted score went from blank to 3.82. The instrument worked. The data was absent.

Why does structure pass as completion?

Because that is the level the AI reports at, and the level most checks test at. Tab exists? Formulas parse? No error tokens? Pass. Nothing asked whether the structure was populated with the data its formulas consume.

This is the same disease as "looks complete" versus "is complete", wearing a different symptom. And the critical part, said plainly: this one was on my checks as much as on the model. The AI claimed a shallow thing, and my validation agreed because it was measuring the same shallow thing.

What does the guard check instead?

_check-tracker-populated.py is a semantic guard: it validates meaning, not just presence. It flags a scoring tab that has candidates but zero scores. It flags a dashboard missing its timeline section. It flags dangling cross-tab references (a formula pointing at a tab that is not there). It flags any second scorecard workbook outside the canonical one, the single agreed-upon copy. And it exits nonzero, so an empty instrument fails the run instead of drawing a polite warning nobody reads.

When it was first added, the check flagged eight tabs as unscored. That was correct: they were pending real evidence, and the check passes only once they are scored. A guard that stays red ahead of the work is doing its job.

The one-canonical-workbook rule earned its place fast. A standalone scorecard had been built before the master workbook existed and was never folded in. Merging them preserved all 193 scored cells, count-verified, and 642 formulas recalculated with zero errors.

How did the check itself become the problem?

Months later, on 2026-08-14, the second incident. In one vendor category, the tab a person opens first was a stale draft from June, while the four live options sat on a last-position sheet. Details generalized to protect the people involved. The old check passed that state. Worse: its own positive self-test had been written to accept it. The check hunted for expected words and formula errors somewhere in the files. It never asked what a person sees first.

So the guard was rebuilt around the human path: canonical tab topology (the right tabs, in the right order) and rosters with scores. It checks category-state-source parity between the dashboard and the tracker, meaning the two must tell the same story. And it enforces estimate-versus-quote labeling, so a planning estimate can never render as an actual quote.

And one meta-rule, the part I would defend hardest: negative fixtures, deliberately broken copies the test must catch. For every known failure shape, the self-test clones the workbook, plants the failure — a duplicate tab, a drifted status, an estimate posing as a quote — and requires the guard to catch it. Any planted failure that escapes fails the self-test. A green check that has never caught its planted target is a green check I refuse to trust. Mine now has to prove it can fail before its PASS means anything, which is the whole verification system in miniature.

"Built" is a claim about structure. "Populated" is a claim about data. An AI will hand you the first dressed as the second, and no glance will catch it. So: what is the workbook in your setup that would pass every check you have — while being empty exactly where a person looks first?

I'm Ali — I run real life-and-work admin on AI agents, then check their work in the open. More at /about.

Published under my standards. Found an error? Tell me — corrections go in the corrections log, dated, never silent.

@online{ali2026built,
  author = {Ali},
  title  = {The AI said the workbook was done. Ten of eleven tabs were empty.},
  date   = {2026-08-14},
  url    = {https://alidoes.ai/built-is-not-populated/}
}

Caught something I got wrong? Send it directly. Confirmed corrections go in the corrections log.