Putting Every Content Rule Behind a Check

What was broken

My content system runs on rules. Rules about what gets written, how long it is, where it goes, and what must never be said. The rules live in documents. AI agents read some of them and write the content.

In August I asked a plain question. Which of these rules is actually in force?

Nobody knew. There were a lot of rule documents, and a weekly check page with about 15 rows. Nothing linked a rule to the check that enforces it.

Then I found a clear case. On one day in July I told the system, in writing, to stop posting to one social channel. The note was on disk. But the schedule kept asking for that channel for 45 more days. That made 15 slots due and 0 posts. It was the single biggest reason my posting numbers looked bad. A rule was written down, and nothing enforced it.

What I built

First I had to find the real rules. A folder full of files is not a list of rules. So I opened 79 files and judged each one against a simple test. Does it change what gets written, how, where, or whether it may go out? Each file got a verdict and a written reason. 21 governed content. 20 sat next to content. 38 did not.

Across that pass and an earlier one, 27 documents stated a real content rule. I loaded all 27 into one database. Then I checked each rule for a gate. 13 of them had nothing enforcing them at all.

Then I closed them, one rule at a time. For each rule I wrote a small check that runs every week and gives a number, not a label. Not "fine", but 46 of 46.

Every check follows three habits.

It says when it cannot look. A check that could not open a folder must never report zero and call it a pass.

It ships with a test that plants a fake mistake and makes sure the check catches it. Three checks passed every row on their first run. The planted test showed one of them was wrong. It would have failed an honest price list forever.

It never skips anything quietly. A retired rule leaves the count, and the page says how many left and why.

What changed

By the end of the day, 28 of 28 rules had a working check. That morning it was 16 of 29. The weekly page went from about 15 rows to 48.

The checks found things the rules alone had hidden.

One live job had pointed at a file that did not exist for two months.

The main content job sorted its style files by name. I tested it. With versions 1.9 and 1.10 both present, a name sort picks 1.9. The day a tenth version was signed, it would have run an old one without a word.

A safety alarm had fired on 57 days out of 84. Every time it was the same ordinary word, not a real problem. Because it always fired, a real alarm would have landed next to one everyone had learned to ignore. I narrowed it to whole words only for that one term. It ran clean for the first time since June.

And some bad news turned out to be bad counting. My tracking sheets said three channels were at about 70 percent of the posting target. The live channels showed about 90 percent, which is the target. The sheets were missing posts. So the check now says "not measurable" instead of "fail" until the sheets are fixed.

Six rows still fail. They should. Each one is a real gap, now with a number next to it.

What I would do differently

I would link every rule to its check on the day I write the rule. Writing a rule with no check just makes a note.

I would never trust a status line inside a rule document. Three times in one day a document said something was not built when it was live.

I would sort anything with a version number by version, never by name.

And I would treat every check that passes on its first run as a check to test, not a result to trust.