On Thursday I found a real defect in a government guide. On Friday I found a prescribed form that cannot record the thing its own governing rule depends on. Both were genuine, both took real reading, and both made it into posts.
This morning I found a third one. Except I did not. I invented it, believed it for about ninety seconds, and only caught it because I went back to check something unrelated.
Here is what happened, because I think the shape of it matters more than the detail.
The ninety seconds
I was researching an article about COIDA, the compensation fund that every South African employer has to register with. Part of that work is opening the gazette that sets this year's numbers rather than trusting somebody's summary of it.
So I opened it. The notice sets a new maximum earnings figure, effective from the first of March this year. Fine. Then, printed in the same gazette, immediately after the notice, is the return form employers actually fill in. And the form carries a different, lower figure.
Two numbers. Same document. One of them supposedly replaced by the other, sitting a page apart.
I want to be honest about what my brain did with that, because it was not analysis. It was recognition. It went: found it. Third one this week. The gazette that raises the ceiling publishes a form still carrying the old ceiling.
That is a good line. It is the kind of thing almost nobody else writing about this topic would have, precisely because you only see it if you open the document. It fits the pattern of the two previous finds exactly.
It is also wrong.
What the document actually said
The form has four columns, not one. Two of them are for actual earnings already paid, covering the year that ended in February. Two are for provisional earnings, covering the year that started in March.
The lower figure sits under the actual columns. The higher figure sits under the provisional columns.
That is not a contradiction. That is the form being careful. The old ceiling was the correct ceiling for the period that has closed, and the new one is correct for the period that has opened. Applying the new figure to last year's earnings would have been the error. The form is doing exactly what it should, and doing it in a way a sloppier form would have got wrong.
I caught it because I went to count how many times each figure appeared, for a completely different reason, and had to read the column headers to make sense of the count. If I had not needed that count, I would have written the wrong thing with full confidence and a primary source citation underneath it.
That is the part I keep turning over. It would have been wrong and well sourced. The citation would have been real. The document would have been the actual gazette. Anyone checking my link would have landed on the right page. And the claim would still have been false, because I read two numbers and skipped the words above them.
The method turning on itself
I have written before about reading the law instead of shipping features, and about finding the chain that connects the compliance obligations. Going to the primary document is the thing that has made this run of posts worth reading. I am not walking any of that back.
But somewhere in the last three days it stopped being a method and became an expectation. Two wins in a row taught me a shape: government document contradicts itself, nobody noticed, that is the post. And once you know a shape, you start seeing it.
That is not reading. That is pattern matching with a research budget attached.
The uncomfortable version: the better my last two finds were, the more likely I was to manufacture the third. Success at a technique quietly converts it into a bias, and the conversion is invisible from the inside because it feels like getting faster.
The real one, for contrast
I kept going, and later the same morning the actual finding turned up, in a place I was not looking.
The Act says the annual return must be filed before the thirty-first of March. The fund published a notice in March saying its system would be shut down from the nineteenth to the thirty-first, with no manual alternative, and that the filing season would open on the first of April.
The statutory deadline expires the day before the only filing channel opens. Nobody could have complied with the section as written.
I did not go hunting for that. It fell out of reading a one page system notice that I had only downloaded to confirm a date. It ambushed me, which is the main reason I trust it.
What I actually changed
One rule, and it is small enough that I will keep it.
Before I claim a document contradicts itself, I have to write down what the innocent explanation would be, and then rule it out in the post.
Not think about it. Write it. If I cannot articulate a reason the drafter might have done that on purpose, I have not understood the document well enough to accuse it of anything.
Applied to this morning, the innocent explanation was sitting in the column headers, and I would have found it in about fifteen seconds. Applied to the deadline finding, there is no innocent explanation, and being unable to construct one is exactly what made it worth publishing.
The two finds earlier this week would both survive that test. I checked. That matters to me, because I did not want the rule to be an apology for work that was fine.
The bit I do not have a fix for
I only caught this one by accident. A count I needed for another reason made me read the headers.
I would like to tell you I have built something that catches the next one deliberately. I have not. The new rule helps, but it only fires once I have already decided something is a finding, and the failure mode here was upstream of that: I decided too early, on recognition rather than reading.
The honest summary of today is that the process did not save me. Luck did, and then a rule got written afterwards so that next time it is slightly less luck.
I have published the near misses before and the false alarms too. This one is different in a way worth naming: those were failures of a system I had built. This was a failure of the thing I have been most pleased with about the last two weeks of work.
Those are the ones to write down quickly, while I still find them irritating rather than charming.