On Oct. 8, Anthropic published a research article showing an all-sky ultraviolet map assembled with Claude Science. The Galactic Center sits in the middle, with two bands overlaid: 154-nanometer far-ultraviolet and 232-nanometer near-ultraviolet. The work was led by Johns Hopkins University astrophysicist Brice Ménard, who is also a researcher at Anthropic.
Anthropic calls it the first complete map of the ultraviolet sky. “Complete” comes with a caveat: roughly a third of the sky has never been observed by any UV telescope, and that portion was inferred by a model from data in other wavelengths. The map ships with extra layers that mark each pixel as “measured” or “predicted” and give an uncertainty estimate.
Why a third of the sky was missing
Earth's ozone layer blocks ultraviolet light, so it can only be observed from space. The main dataset comes from NASA's GALEX, which was in orbit from 2003 to 2013 and made about 38,000 independent observations covering roughly two-thirds of the sky.
The gaps have a cause. Regions dense with bright stars can damage the detector, so GALEX deliberately steered around them, and the Milky Way's plane was hit hardest. NASA's Swift, Korea's FIMS/SPEAR and Europe's TD-1 filled in some of it, but holes remain. In Ménard's words in the article:
“Roughly a third of the sky has never been observed in UV.”
What Claude actually did
According to the article, the pipeline has six steps, most of them carried out by Claude Science's agents:
- Search for and download public ultraviolet survey data;
- Subtract the glare around bright stars so that each survey is internally consistent;
- Cross-calibrate, then unify resolution and coordinate systems before merging;
- Inpaint the blank regions using the relationship between ultraviolet brightness and visible, infrared and radio data, including ESA's Planck data;
- Add the ultraviolet contribution of more than 100 million stars, based on Gaia measurements;
- Correct the residual airglow left in individual GALEX observations.
The whole project took a few days, spread across several work sessions over one summer. At each step Claude ran computations lasting hours, iterating through more than a dozen versions along the way. Ménard estimates the traditional approach would typically take weeks of manual work.
For validation, the team masked regions that already had real measurements, let the model predict them, and compared the results with the true values. The deviation was about 10%.
Automated review missed an artifact
The most informative part of the article is a mistake.
An early version of the map showed disc-shaped imprints. They came from atmospheric airglow left in individual GALEX observations. The map had already passed two rounds of review by other agents, and none of them caught it. In the end Ménard spotted it himself and told Claude:
“I can see discs with the imprint of individual observations; can you correct that?”
Anthropic also states in the article that this is only a first version and that the whole process depended on guidance from the researcher.
Where this fits in the Claude Science line
When Claude Science launched this year, its pitch was bringing more than 60 research libraries into one working environment. Since then Anthropic has put out several results: a new enzyme system found in bacteriophages, about a 4x speedup for 30 biomolecular models, and super Yang-Mills amplitudes computed to nine loops. Most of those sit in biology and theoretical physics; the ultraviolet map is the first to land in observational astronomy.
What the results have in common is that Claude handles mostly data moving, cleaning, alignment and repeated iteration, while scientific judgment still rests with people. The disc artifact is one example: two rounds of agent review let it through, and it was finally caught by an astronomer familiar with GALEX data.
Third-party coverage has raised reservations too. One review notes that filling in a UV map by statistical methods is not new: a 2021 paper used neural networks to reconstruct unobserved regions of the far-ultraviolet survey. In addition, the map site refused access at the time of writing, so file formats and layer definitions could not be independently verified. For now, the “first complete UV map” claim rests on Anthropic's own account.
Sources: Anthropic research article, CocoLoop, Crypto Briefing, all-ai.de; we checked the GALEX sky coverage and observation count, the roughly 10% deviation reported in the masking test, and the wavelengths of the two ultraviolet bands.