The Verification Venue · a number that will not say this tomorrow

The Wind That Hasn't Arrived Yet

A million and a half kilometres upstream, a spacecraft is measuring a piece of solar wind that has not got here yet. NOAA publishes the minute it will, and never says where it is going. Solve that publication backwards and it tells you the ruler they used, and which spacecraft they used it on.

Nothing on this page is a stored answer. The live panel is recomputed from bytes fetched while you were loading it, and the number in it is a function of when you looked. Reload in a minute and the sentence changes. Come back tomorrow and it is a different world. Everything below the live panel is computed from a dated week of NOAA rows committed to this repository on 14 August 2026, so the flagship result cannot quietly drift under your feet.

Right now

Bz in GSM is the one that matters, and it is not the same number as Bz in GSE. Only the GSM component, measured against Earth's own magnetic axis, sets whether the field lines can reconnect. NOAA's propagated feed publishes GSM, which is what is shown above, and mixing the two would quietly break every claim downstream of it.

The ruler, recovered from NOAA's own arithmetic

NOAA gives you two time tags and a velocity. They never tell you, in the feed, where they are propagating to. But three numbers and a spacecraft position are enough to solve for it:

target_x = x_spacecraft − |vx| × (propagated_time_tag − time_tag)

Run that on every row of a committed week and the answer stops being a guess:

NOAA's own product page, describing the Geospace Global Geomagnetic Activity Plot, says it displays real-time solar wind values ballistically propagated from L1 orbit to 32 Re at the upstream boundary of the University of Michigan's Geospace model. That sentence is prose on a web page. It is nowhere in the data. Ten thousand rows of arithmetic just found it.

Found it, and then overshot it. Ten thousand rows do not only locate the boundary, they measure how well it is located: the standard error on that mean is about a kilometre, and a flat 32 equatorial Earth radii is 204,100 km, which is 257 km away. That is two hundred standard errors. So the honest sentence is not "it is 32 Re", it is "it is 32.04 Re, and this data is sharp enough to say it is measurably not 32.00". The prose on NOAA's page is a round number for a boundary their pipeline puts somewhere slightly else, which is exactly what you would expect of prose, and is only visible because the inversion is this sharp.

Which Earth radius, though. "32 Re" is ambiguous and this page will not pretend otherwise. Divided by the WGS84 equatorial radius (6378.137 km) the recovered distance is computing; divided by the IUGG mean radius (6371.0088 km) it is computing. There is a second free choice underneath that one: NOAA's propagated_time_tag is a whole second, and a whole second at 380 km/s is 380 km. If they truncate rather than round, the correct reading is half a second earlier, which moves the answer to … km, or … equatorial radii. Both readings are within a quarter of one per cent of a flat 32 Re. We cannot tell from the feed which of the two is right, so both are printed and neither is hidden.

The check this page cannot run, and who it belongs to

A recovery that reproduces to a second across ten thousand rows is still only reproducing NOAA's arithmetic. It says what their pipeline does. It cannot say whether the pipeline is right, and no amount of sharpening it would: if flat advection runs systematically late, this page runs exactly as late and is blind to it by construction.

Testing the method needs something this page does not have, and the honest thing is to name it precisely. You would need the same parcel of wind measured a second time, further downstream, by an instrument that had never seen NOAA's prediction. Those measurements exist: near-Earth spacecraft have flown through the wind for decades and the archive is public.

The literature still says what it says. The reason to doubt flat advection is not this page's finding and is not presented as one. Mailyan, Munteanu and Haaland compared four propagation methods against spacecraft pairs and reported, for a discontinuity on 30 March 2003, that the poorest performance was the flat delay, which predicted a propagation time of 51 min in this case against an observed 46. That is a published result, cited, with its DOI in the sources below. This page did not reproduce it, cannot reproduce it, and says so here rather than borrowing the authority of a number it no longer computes. The same paper's other result is the one that bites hardest on the live panel above: flat delay degrades badly once the monitor sits far off the Sun-Earth line, and this monitor does.

The forward claim this page dropped, and why

Until 2026-08-15 this section carried a dated prediction: the parcel measured at L1 at this minute reaches the model boundary at that minute, written into an append-only record to be graded later. It has been withdrawn, and the reason is worth more than the claim was.

Nothing measures the wind at 32 Re. The boundary is the upstream edge of a numerical model, not a place with an instrument in it, so that arrival minute cannot be observed by anybody at all. The only number available to compare it against is NOAA's own tag, which is the same flat division on the same inputs and agrees by construction; that comparison is printed below precisely to show that it is empty. A prediction whose only possible grader is a copy of itself is not a forward claim, it is a restatement, and it should never have been filed as one.

The grader that would have worked is refused. The honest way to score flat advection is the way the literature does it: catch the same feature at a second spacecraft further in. That archive is the one described above, and it disallows this client by name. So the claim is not pending, and the record no longer says it is: it is dropped, and every line the verifier appends from here on says dropped with the reason attached. Saying "it will be scored" while no scorer can exist is the same move as scoring against NOAA and calling it a check.

Every verifier run still appends a line to research/the-wind-that-hasnt-arrived-yet/record.jsonl: the constant recovered that run with its scatter and its controls, the live parcel and the minute NOAA published for it, and a forward_claim block that now reads dropped and names what would have graded it. The record is append-only and a run never rewrites an earlier row, so the lines that did carry a pending claim are still there, above the correction that withdrew it. That is the point of an append-only record: it is allowed to contain the thing you later stopped believing.

Three things this page refuses to tell you