Japan's Environment Ministry has published two accounts of the same fault, and one of them is far more specific than the other. On 24 August it added a notice to its heatstroke-prevention site saying that some of the WBGT heat-index forecasts it published between 22 April and 22 July 2026 were wrong, that it had temporarily removed the affected values from its database, and that they will be reposted once correct ones have been computed. The notice measures the fault in hours and dates: the period 22 April to 22 July, and the 0, 1, 2, 12, 13 and 14 o'clock hours. The Ministry's own public forecast API measures the same fault in volume. Query it for any date inside that window and a station returns 360 forecast records instead of 492, with those six hours returning nothing at all; across 92 consecutive days and 841 operating stations, that is 10,213,104 forecasts missing from the public archive.
What went are issuance runs, not forecast targets. The API manual defines one field as the moment a run was published and another as the hour being forecast, and since forecasts only ever target three-hourly steps, a 13:00 entry can only be a run. On an intact day a station returns 492 records, and the six deleted hours carry 132 of them, or 26.8 per cent of the day. The boundaries confirm the window rather than merely assuming it: step one day past it to 23 July and all 24 hours come back at 492; step back to 21 April and the API returns nothing, because 22 April is day one of the 2026 service. The baseline year is unbroken — June 2025 returns 492 records on every one of its 30 days. The 841 is the Ministry's own figure too, matching both its station master and the nationwide count in its 14 April operational press release. Only the forecasts went; the observed archive still answers with a full set of hourly readings for dates inside the window.
The Ministry keeps a permanent dated list of notices, and it does log errata there: an entry in June recording a misdelivered alert image on its LINE account, and one in April recording that finalised 2025 values posted days earlier were wrong and had been rolled back. As of 25 August that list's newest entry is dated 20 August, with no mention of this deletion — the two smaller errors are written into the ledger and the largest is not. The notice itself appears on six pages in three different wordings that do not say the same thing, and only the homepage version calls the deletion temporary or records that heatstroke alert issuance was unaffected. The English-language site carries no notice at all, so English-speaking users of the same API have no warning. The Ministry says the correct values will be recomputed and republished, and nothing in the archive contradicts that. What the archive adds is the size of the hole.