Circular No. 1 · 14 October 2026 · gapmap 0.1.0
Labelled, Checked, and Still Wrong Sometimes: Honest Gap-Filling for Survey Maps
The gapmap contributors. Every figure below is computed by the library on this page.
Summary.
gapmap predicts the unmeasured part of a map from complete related layers, labels each pixel, and attaches an uncertainty to each prediction. On 30 fixed simulated skies its block check on measured sky gave a median relative error of 6.2%, against a true error on the gap of 6.6%: the check was optimistic on 23 of 30 skies. Its footprint test found a glow artifact on 40 of 40 skies and raised 3 false alarms on the same 40 without it.
Background.
On 8 October 2026 Anthropic published "The missing map of the sky", in which Brice Ménard, an astrophysicist at Johns Hopkins University and a researcher at Anthropic, describes the first complete map of the sky in ultraviolet light, made with Claude Science [2]. NASA's GALEX imaged about two thirds of the sky in some 38,000 observations between 2003 and 2013, avoiding the brightest regions, including the plane of the Milky Way; teams of AI agents gathered and cross-calibrated it with other surveys, and the missing third was predicted by inpainting from visible, infrared and radio data, with ultraviolet light from more than 100 million stars estimated from Gaia [2][3]. Every pixel is labelled measured or predicted, with uncertainty layers; predictions of hidden regions that did have data came within about 10% after several rounds [2]. Ménard noticed faint circles in a dim field, the footprints of single observations carrying leftover atmospheric glow, which two rounds of review by other agents had not caught [2][3].
Not connected. gapmap, this site and $UVSKY have no connection with Anthropic, Johns Hopkins University, NASA, ESA or the author, and none of them has seen this circular. The sky on this page is simulated; it is not the real map.
The model.
A ridge regression of the map on second-order terms of the standardised layers, fitted on measured pixels; checked against an independent numpy solve on three skies (largest difference 1e-14). A predicted pixel's uncertainty is the block-check error times the square root of one plus its leverage.
The checks.
Block (12-pixel squares in five folds), random (scattered pixels) and gap-like (the fifth of measured sky most like the gap). The model uses no neighbouring pixels, so block and random agree; the gap-like check was meant to come closer to the truth and did so on only 11 of 30 skies.
The footprint test.
For each footprint, the residual step at its edge after a local plane takes out smooth gradients; the null is the same step on same-centre circles at 0.6 and 1.45 radii. The map-level test compares the roughness of the two (log variance ratio, one-sided, alpha 0.01). It found every injected glow, but its false-alarm rate was 7.5% at a nominal 1%, most likely because the normal approximation ignores how neighbouring steps depend on each other.
Offsets.
Fitting one offset per footprint with the model recovers the pattern (correlation 0.65 with the true glow) but taking the offsets out changed the median error on measured pixels from 7.0% to 7.2%, lower on only 16 of 40 skies: footprint-scale sky structure is partly mistaken for offset. In the real map the glow was fixed at its source [2].
Limitations.
The skies are fictional and generated by the same people who wrote the model, so they may suit it. The model knows only the layers it is given. The evaluation seeds were fixed in advance and run once; the library was developed on other seeds. The results above are not a statement about the real ultraviolet map.
Use.
Label every pixel. Report the gap's error as at least the check's error, and say which check. Look at the strongest footprint edges by eye before trusting any map stitched from footprints.