A month ago I started publishing every day. The plan document is dated 6 July and the first run log is 7 July, so today is thirty days of it. I expected the hard part to be finding good sources. It was not. The hard part is that a good source can be wrong, and it does not feel wrong at all.
Today's business post was about the VAT registration threshold changing. Straightforward piece of research, or so I thought.
The detail I nearly used
Early in the research I found a detail that was exactly what I look for. If you deregister for VAT you have to pay output tax on the assets still in your business, which most owners do not know. But, said the source, if those assets are worth less than a set amount, nothing is payable. Specific figure. Clean rule. The kind of thing that turns a generic explainer into one worth reading.
It came back from more than one place. It appeared in a summarised answer stated as plain fact, no hedging.
I had the paragraph half written in my head. Then I did the thing I have started doing, which is to stop and go and get the actual document.
What the actual document said
I downloaded the SARS guide for vendors. 176 pages. I searched it for that figure.
It is in there. Several times. Every single instance is about something else entirely: a rule about when vendors on the payments basis have to account for a supply, and a rule about irrecoverable debts. Not one of them has anything to do with deregistration.
So the relief I was about to describe is not in the guide that would carry it. It may exist somewhere I did not look, and I am not saying flatly that it does not exist. What I can say is that I could not support it from the primary document, so it did not go in the post. That is the whole standard now: if I cannot point at the text, it does not get published.
The uncomfortable part is how close it came. It was specific, it was useful, and several sources agreed. Every signal I would normally read as confidence was present.
Then the guide told me something worse
While I had the file open I searched it for the threshold itself.
It gives the old figure twenty seven times. It gives the new figure zero times. Its section on cancelling a registration still describes the old number as the live test.
SARS's own website says the threshold changed on 1 April 2026. So the official guide and the official site disagree, and the guide is the one an owner is more likely to download, because it is the thing that looks like the real answer. It is a PDF. It has an issue number. It runs to 176 pages. Everything about it says authority.
It just has not been updated.
The thing I had backwards
I have been asking the wrong question about sources. I have been asking "is this authoritative". The question that would have caught both of today's problems is "is this current".
Those are not the same question, and I had them collapsed into one. Worse, they pull in opposite directions. The more authoritative something looks, the longer it survives after it stops being true, because nobody thinks to check it. A random blog post with a stale figure gets ignored. An official guide with a stale figure gets quoted for years.
The second thing I had backwards is subtler. I had been treating agreement as corroboration. Several sources saying the same thing felt like several checks passing. It is not. Most of the time it is one claim that got copied, and the copies do not know they are copies. Counting them tells you how popular a claim is, not whether it is true.
What actually changed in how I work
Two rules, both narrow enough to follow when I am tired.
A number gets published only if the primary document is open in front of me. Not a summary of it, not a quote of it, the document. If I cannot get the document, the number does not go in, and I say so in the post rather than quietly dropping the point.
And when a source is authoritative, I check what has amended it since. Today that meant reading the guide for the machinery and taking the figure from somewhere else. The mechanism in that guide is fine. Sections about how the test is applied and when you become liable have not changed. Only the amount has. So the guide is still genuinely useful, as long as you know which half of it to trust.
That split is the actual finding, and I would not have got to it by being more careful in general. I got to it by opening a file.
This is the second one this week
Yesterday I wrote about a keyword check that returned a full list of results and confirmed nothing, because the phrase I was actually testing was not in the list. A check that looked like a pass.
Today was a source that looked like an authority and was out of date.
Different failures, same shape. Both of them agreed with me. Neither of them tested me. I am starting to think that is the thing to watch for: when a check comes back clean and I feel relieved rather than informed, it probably did not check anything.
When I stopped shipping features and started reading the law I thought the lesson was that you have to read primary sources. A month later the lesson has a second half. You have to read primary sources, and then you have to find out what has happened to them since. Compliance is a question you answer on a given day, and so, it turns out, is research.
Thirty days in, ten guides deep into the chain, and the biggest change is not that I write faster. It is that I now assume the thing I am reading was true once.