Twenty years and forty-five minutes, on one axis

The background

A forecasting system files three different things in one drawer: an archive of what the world did, a picture of what it is doing now, and forecasts reaching a little into a future that has not happened. The obvious display is a table, and it hides the two things anyone wants: accumulation — the archive as a long band, the forecasts as short bars crowding the present — and whether the forecasts were any good.

The requirement

The store filling up, and whether a forecast that had run its course turned out right. A table is free keyboard and screen-reader surface and a picture is neither, so a check proving every dataset announced and reachable in order had to pass first.

The options considered

The axis is the awkward part. The archive covers twenty years and a forecast reaches forty-five minutes past the moment it was made, so on a straight timeline a forecast is a hairline at the right edge. Giving each row its own scale is worse: two bars at the same position would be at different moments, and position meaning time is what a timeline promises. So one axis, compressing elapsed time logarithmically from the newest edge, and saying so rather than letting a reader infer duration from tick spacing.

Three faults came from looking at the running page: two lanes landing on every line, four timestamps printed over each other, and a forecast run publishing two datasets over the same interval, so eight stored datasets drew as five. Overlapping bars stack now.

The demo

Open it at the forecast branch and pick a forecast whose window has passed: the comparison draws the forecast, the truth, and the starting field held constant — error alone reads as a verdict.