That distinction matters. “Complete” does not mean every point in the map was photographed in ultraviolet light. The project combines decades of telescope observations with statistical estimates for places no UV instrument has seen.
A sky assembled from several missions
Most wide-area ultraviolet imaging comes from NASA's Galaxy Evolution Explorer, or GALEX, which operated from 2003 to 2013. Its roughly 38,000 snapshots covered 64% of the sky in far-ultraviolet light and 76% in near-ultraviolet light, according to the project's technical overview. Bright stars and much of the Milky Way's plane were deliberately avoided to protect GALEX's detectors.
Ménard and Claude Science combined that archive with smaller or coarser datasets from NASA's Swift-UVOT, South Korea's FIMS/SPEAR instrument and Europe's TD-1 satellite. The sources had to be cross-calibrated, placed on the same coordinate grid and adjusted so that overlapping observations agreed.
One recurring problem came from faint atmospheric glow left in individual GALEX fields. Ménard wrote in Anthropic's account of the project that he spotted circular footprints in a dim region after two rounds of automated review. Claude's agents traced the artifact to residual background light and reprocessed all 38,000 observations.
How the unseen regions were filled
After the measured datasets were merged, 28% of the far-UV sky and 27% of the near-UV sky still needed values. The gap-filling model was not a language model or an image generator. It was a gradient-boosted regression trained on observed UV regions and 90 features drawn from all-sky maps of dust, hydrogen emissions and Gaia starlight.
The model learned statistical relationships between those inputs and observed UV brightness, then applied them where direct measurements were missing. The pipeline also added measured or estimated ultraviolet fluxes for about 119 million stars in the European Space Agency's Gaia DR3 catalog.
Measured data always take priority in the finished map. Coarser or translated measurements come next, and predictions fill what remains. Published provenance weights and uncertainty layers preserve those distinctions pixel by pixel.
Useful for teaching, limited for discovery
The team tested the prediction system by hiding patches of real GALEX data and asking the model to reconstruct them. The project reports a typical error of about 12% to 14% in filled regions, compared with roughly 5% uncertainty where GALEX took images. Error grows farther from measured areas.
Those numbers come with substantial caveats. The project's own technical page says the filled sky cannot reveal a new star or galaxy because no telescope observed those locations. Fine structure below roughly half a degree to one degree in predicted regions is not real, and the model-filled areas cannot serve as an independent tracer of dust because dust maps were among the inputs.
Ménard presents the map primarily as a teaching resource. The technical page also notes that planned UV missions need realistic all-sky brightness estimates. In ultraviolet light, hot young stars and dust illuminated by starlight stand out, while the dusty plane of the Milky Way can appear dark. The full-sky view lets students compare those structures with visible, infrared and radio maps without confronting the large blank regions in earlier UV mosaics.
Anthropic says the collaboration unfolded over several days, with Ménard setting direction and reviewing successive versions while Claude coordinated many processing agents. The company estimates the same work would have taken humans weeks. That estimate comes from Anthropic, while the technical paper remains a draft in preparation. The project page also warns that the map and its validation have not yet been peer-reviewed by humans.
