Claude Science Builds First Full-Sky Ultraviolet Map With ~10% Error

Astrophysicist Brice Ménard of Johns Hopkins University has built the first complete map of the whole sky in ultraviolet. Anthropic's Claude Science AI agents did almost all of the work. They downloaded data from several space missions, calibrated it, and filled gaps in coverage using inpainting. For the filled-in regions, the predictions differed from actual measurements by about 10% on average. Ménard himself sees the project as an example of research that "simply would not have been done" without AI.

What the agents actually did

The task looks routine, but it is labor-intensive. The atmosphere absorbs almost all ultraviolet light, so it has to be observed from space. The data comes from different missions, each with its own instruments, coverage areas, and calibration quirks. To produce a single map, you have to collect the datasets, bring them to a common scale, and deal with regions where measurements are missing or unusable.

According to the description, the agents completed this entire path themselves, from downloading data to calibrating it and reconstructing the missing areas. Inpainting fills in missing parts of an image based on neighboring data and learned patterns. It is the same technique used to restore a hidden or damaged area of a photograph. Here, the same principle was applied to the celestial sphere.

How to read the 10% figure

The average error of about 10% comes from comparing predictions with actual measurements. For a scientific map, this is the key metric: it shows how far you can trust regions where data was not observed but reconstructed.

The limitations matter too. The source material gives no details on how exactly the metric was calculated, how the error is distributed across the sky, or where it is noticeably above average. It also does not say which missions made it into the dataset. So far this is a brief summary, not a full methodology. Evaluating the result requires a publication and access to the map itself.

Key project parameters:

• Author: Brice Ménard, Johns Hopkins University.
• Tool: Claude Science from Anthropic.
• Result: the first complete all-sky map in ultraviolet.
• Data sources: materials from several space missions.
• Steps performed by the agents: download, calibration, gap filling via inpainting.
• Accuracy: average deviation of predictions from measurements of about 10%.

Why this matters for the industry

The project shows the practical role of agents in science. They do not pick hypotheses or write reviews. Instead, they take on the heavy engineering work: gathering scattered archives, bringing them into a uniform format, and turning them into a finished product. For many groups, this part decides whether the work gets started at all.

Ménard's phrasing describes a specific class of tasks: work that is useful but does not justify months of human time, so nobody does it. If an agent brings the cost of such work down to a reasonable level, the list of feasible projects grows, especially where large open archives meet.

What you can take away

Research teams should treat this case as a template. It suits tasks where there is a lot of data in public archives and the result can be checked against independent measurements. That is how the 10% estimate was obtained: hide part of what is known, predict it, and compare.

When working with agents, it helps to follow three rules:

• Keep a log of the agent's steps so the calibration can be reproduced.
• In the final map or table, clearly separate observed values from reconstructed ones.
• Publish the error metric not only as an average but also by region.

For readers, the practical takeaway is simpler. Until there is a publication and open data, the map is best seen as a strong demonstration of what agents can do, not a ready reference resource. Once the methodology is out, it will become clear how well 10% works for astrophysicists' tasks.