Forecast accuracy log
This is a laboratory, not a scoreboard. Every hour the app writes down what it thinks the weather will be — at several distances into the future, for a dozen rival ways of blending the same sources, including the way it actually shows you. Separately it writes down what the weather did. Everything below is that log read back, with the sample size attached, so you can see which claims are evidenced and which are still collecting evidence.
Nothing here is a verdict until the matched sample behind it is large enough to be one. Where it is not, the page says so rather than picking a winner. One of the rivals is Open-Meteo’s own blend, left completely alone: if our re-blending is making things worse, that line will beat “Consensus (live)” on shared hours — and then we switch.
Scoring only the hours where 4 or more independent model lineages were actually available. This is the line-up the app is designed around, and the only cohort where a holdout comparison can mean anything: a rival that drops a source can only differ from the live blend when that source was there to drop.
reading the log…
Does the wording match the data?
Every couple of hours the app rebuilds the exact sentences it would show at the top of the page, then checks them against the exact numbers they came from. A failure here never means the weather surprised us — it means the page disagreed with itself, which is always a bug.
reading the audit…
Reported issues
Reports are private. Enter the maintainer key to read and triage them.