We applied our own rule more strictly. The number went up.

The methodology page has said the same thing since it went up: days are comparable by engine count, and a day with an incomplete set is excluded entirely. The rule comes from a failure we had on 24 August. One API key died, every answer that day came from two engines instead of three, and the day’s average went up — a mention by one engine of two scores 50 where one of three scores 33. A partial day reads as a rise that never happened.

What the code did instead

It dropped the incomplete rows and kept the day. A day where one engine missed one question was counted on the rows that engine did reach — an average over the businesses one engine happened to find, which is not the market’s average. The existing test covered a day where every row was incomplete, which the code handled. The case it did not handle was the mixed day, and there was no test for it, so it stayed that way from 4 September, when the index went live, to 14 September.

How it surfaced

While building a check that the daily sweep had actually run, we counted days three ways and got three answers. One of them matched the sentence on the methodology page; the code’s did not.

What changed, in figures

Thailand: 9 comparable days became 8, and the score 1.5 became 1.7. The excluded day sat below the mean, so the looser rule had been publishing the lower number. The same pass found a second gap between what the code did and what it was meant to do; both were fixed in one change, and a test now covers the mixed day.