Claude Science's full-sky ultraviolet map still needed an astronomer to spot a glow artifact

The circles Menard saw were in a dim area of an early version of the map ^[12]. Each one was the trace of a single GALEX observation, caused by uneven ultraviolet glow from Earth's atmosphere ^[12]. According to Anthropic, two rounds of review by other agents had passed that map, telescope footprints and all ^[13]. Once Menard pointed it out, Claude traced the cause and corrected the glow across all 38,000 GALEX observations ^[14].

We'd expect an error like this to be faint in any one exposure. It shows up when thousands of exposures are stitched together and the display is stretched for a dim region, where a small background offset is a large share of the signal. A reviewer checking each pipeline step can pass it. A person looking at the finished picture sees circles. We think correcting all 38,000 exposures was the right call, since a patch over the visible circles would have left the same offset in the rest.

The inferred third was tested with a hold-out. Menard says he had Claude hide regions with known ultraviolet data and predict them, and after refinement the estimates came within about 10% of the withheld measurements ^[5]. The two published accounts state that figure differently. Runtimewire says "within about 10%" ^[5]. The Decoder says predictions "averaged about ten percent deviation" ^[6]. A bound and a mean are different claims, and a 10% mean allows some regions to miss by much more. "After refinement" also matters. If the method was tuned against the same withheld regions, part of the 10% is fitted. Neither account says how the test regions were picked or whether the refinement saw them.

For the figure to carry over to the gaps, the test regions would have to resemble the gaps. We doubt they do. GALEX avoided regions around very bright stars, including much of the Milky Way's plane, to protect its detectors ^[2]. The fill uses inpainting plus correlations between ultraviolet light and visible, infrared and radio observations ^[4]. Those correlations can only be learned where ultraviolet data exists. The map also folds in ultraviolet estimates for more than 100 million stars, inferred from visible-light measurements by ESA's Gaia mission ^[10]. That is a second model inside the first. Runtimewire makes the same caution: the test shows performance in selected known regions and does not establish equal accuracy in unobserved sky ^[8].

We think the release handles the inferred region well. The finished map marks measured and predicted pixels separately and includes uncertainty estimates ^[11]. Anthropic presents it mainly as an educational resource and a visualization of existing data, and says it is no substitute for ultraviolet measurements ^[7]. Claude Science, in beta since June 30, 2026, keeps a record of how its outputs were produced ^[16].

Agents ran searches and computations for hours while Menard worked on other projects ^[15]. He says the project took several days, against weeks for people doing comparable work ^[15]. Menard set the task, noticed the flaw, directed the correction and judged what the estimates meant ^[9]. He suspects many scientists have been putting off similar projects that AI could now make possible, The Decoder reported ^[17].