The idea: a meteorological service that issues rigorous, high-confidence forecasts for weather that has already happened, and refuses to say anything about tomorrow.

I want a reality check on this because I keep going back and forth on whether it's useless or the only honest thing anyone's ever done with a barometer.

The mechanics are straightforward. You take the atmospheric state at, say, noon on the 14th of March 1987, and you run the model forward six hours. Then you publish that as a forecast. The thing is, you don't look at what actually happened on the afternoon of the 14th. That data gets sealed. The forecast stands on its own as a claim about a world we can, in principle, check but choose not to.

What I like about it is that it makes the forecast falsifiable but unfalsified. Which is roughly the epistemic position all forecasting is in anyway, just usually only for about eighteen hours.

Where I get stuck is the archive problem. To seal the observed record you need someone to hold it, and that person now knows. There's a custodian somewhere who could tell you whether it rained in Bristol that afternoon. Does the whole thing collapse if one person knows? I don't think so — plenty of things stay meaningfully open even when someone somewhere has the answer. But I can't articulate why that's fine here and not fine for, say, a sealed envelope with a coin flip in it.

The other snag, and this is the one I actually want help with. If the model is good, the past forecast will mostly be right. So the service converges on being a slow, expensive way to describe history. The value only exists in the gap between the model and the record, and that gap shrinks every year.

Is there a version of this that stays interesting as the models improve? Or is the whole thing load-bearing on us being bad at physics?