Anthropic's Claude Science Agents Fill In the Sky's Missing Ultraviolet Map

Anthropic has published an account of an astrophysicist using its Claude Science product to assemble what he describes as the first complete map of the sky in ultraviolet light, with roughly a third of it statistically predicted rather than measured. The write-up is notable less for the astronomy than for what it shows about the current shape of agentic научной work: dozens of parallel agents doing weeks of calibration drudgery, and a human spotting the artifact that two rounds of automated review let through.

The post is authored by Brice Ménard, who is identified as an astrophysicist at Johns Hopkins University and a researcher at Anthropic — a dual affiliation worth keeping in mind when reading a vendor's account of its own tool.

The gap in the data

Ménard's framing is pedagogical. He teaches by showing students the same sky at different wavelengths: stars in visible light, dust in infrared, hydrogen gas in radio, violent events like supernovae and black hole accretion in X-rays. Ultraviolet reveals dust lit by starlight — clouds around young stars, rings left by stellar explosions — but, he writes, the only UV map he could show his students was full of holes.

The reason is partly physical and partly operational. Ozone absorbs UV, so the observations have to come from space. The largest source is NASA's GALEX mission, which ran from 2003 to 2013 and imaged about two-thirds of the sky across roughly 38,000 separate observations. GALEX deliberately avoided pointing at very bright stars, including the densely packed plane of the Milky Way, to protect its detectors. NASA's Swift and South Korea's FIMS/SPEAR added coverage, but gaps remained.

Statistical methods to fill those gaps already exist, Ménard notes. The obstacle was labor: weeks of pixel-level calibration and repeated analysis for a project that never outranks a researcher's active work. His broader claim — that most fields carry a backlog of useful-but-unprioritized artifacts — is the real argument of the piece, and it is the kind of claim agentic tools will be judged on over the next few years.

How the pipeline was assembled

According to the post, Claude orchestrated a team of agents through several stages. Agents searched the web for public UV surveys and downloaded them, then worked in parallel on different sky regions to make each survey internally consistent, including removing glare from bright stars so fainter surrounding UV light could be measured. Different instruments capture UV slightly differently, so the surveys were cross-calibrated against each other, resampled to a common resolution, and placed on a shared coordinate system before merging.

For the unobserved third, Ménard says he directed Claude to use inpainting — the machine learning technique of reconstructing missing image regions from their surroundings — combined with the fact that the missing areas are covered at other wavelengths. Using the two-thirds of the sky with real UV data, the system learned how UV brightness relates to visible, infrared, and radio observations, then applied that relationship to the blank regions, producing both an estimate and a confidence value per pixel. Finally, estimated UV emission from more than 100 million individual stars, inferred from ESA's Gaia visible-light measurements, was layered on top of the diffuse background.

The cited data sources include GALEX, Swift-UVOT, FIMS/SPEAR, TD-1, Gaia DR3, and Planck plus Hα templates used for gap-filling.

The validation, and what it does and doesn't tell you

The accuracy check described is a standard hold-out test: take regions that do have UV data, mask parts of them, and ask the model to reconstruct the hidden values without access to them. After several rounds of refinement, Ménard writes, estimates came within about 10% of the real measurements — a difference he describes as almost imperceptible visually.

That is a reasonable test, but the post does not specify whether 10% is a median, mean, or worst-case figure, nor how the masked test regions compare in character to the genuinely unobserved ones. This matters more than usual here. GALEX skipped the galactic plane because it is crowded with bright stars; the unobserved regions are, almost by construction, the hardest and least representative part of the sky. Holding out patches of well-behaved high-latitude sky is not the same test as predicting the galactic plane. Anthropic does not claim it is, but the post also does not break out performance by region.

To its credit, the published map carries layers labeling each pixel as measured or predicted along with uncertainty estimates — the right design choice, and one that keeps the artifact honest for the educational use Ménard emphasizes. He positions the result as a teaching tool and a resource, not as a scientific measurement. There is no mention of a peer-reviewed paper, a preprint, released code, or a comparison against the established statistical techniques he says already exist. Compute cost and total runtime beyond "several days" of collaboration were not disclosed, and the post does not identify which model or training data underpinned the inpainting step.

The bug the reviewers missed

The most useful passage for anyone building agent pipelines is the failure story. Reviewing processed images, Ménard noticed faint circles in one of the dimmest fields — the footprints of individual GALEX pointings, each carrying slightly different residual airglow from Earth's atmosphere. If that glow is not fully subtracted, each circular patch sits slightly brighter or darker than its neighbors.

Claude had flagged this as a known issue at the start of the project. The map still passed two rounds of review by other agents without the problem being caught. A one-sentence prompt from Ménard sent the agents back to the root cause, and a corrected background subtraction was applied across all 38,000 observations in a couple of hours of processing.

Two lessons sit in that anecdote. The capacity advantage is real: reprocessing tens of thousands of images on a hunch is cheap in a way it was not before, which changes what a researcher is willing to try. But agent-on-agent review failed to catch a defect the system itself had already anticipated — a reminder that automated QA tends to validate conformance to the plan rather than inspect output with a domain expert's eye. Ménard describes his contribution as guiding the agents, with the patient labor done by them; the airglow episode suggests the guidance was load-bearing.

Why this matters commercially

Anthropic is building a case that Claude Science can convert a discipline's backlog into finished artifacts without consuming the researcher's primary time — Ménard explicitly says he pursued this without sacrificing research hours. That is a sharper and more testable pitch than generic "AI for science" claims, because the output is checkable and the counterfactual is clear: this map did not exist before.

The open question is whether the pattern generalizes beyond data fusion tasks. This project had unusually favorable structure: public datasets, well-understood physics relating one wavelength to another, a natural hold-out validation scheme, and an expert who could eyeball results. Fields without those properties will find the handoff harder.

What to watch

• Whether the map, methods, or pipeline code are published in a peer-reviewed venue or as a preprint, which would let other astronomers probe the ~10% accuracy claim independently.
• Whether anyone uses the predicted regions for research rather than teaching, and whether the uncertainty layers are respected if they do.
• Whether Anthropic publishes region-by-region error breakdowns, particularly for the galactic plane that GALEX avoided.
• How agent-review architectures evolve after cases like the airglow discs, where automated checks passed a known-risk defect.
• Whether other researchers reproduce the backlog-clearing pattern in fields with messier data and weaker validation options.