The first complete ultraviolet portrait of the sky is not a conventional photograph. Roughly two-thirds comes from ultraviolet measurements assembled from space missions. About one-third is an estimate produced by relating the observed UV sky to what telescopes see at visible, infrared and radio wavelengths. The project’s strongest design choice is that the finished resource can label each pixel as measured or predicted and attach an uncertainty estimate.[1]
Brice Ménard, both a Johns Hopkins astrophysicist and an Anthropic researcher, led the project with Claude Science. This is therefore a company-published account of a researcher using the company’s system, not an independent benchmark. It nevertheless provides a detailed example of where agentic scientific work is useful: collecting public archives, cross-calibrating instruments, running repeated analyses and maintaining a large processing pipeline under expert direction.[1]
What went into the map
| Measure | Value | Context |
|---|---|---|
| Largest UV survey | GALEX | NASA’s mission operated from 2003 to 2013 and imaged about two-thirds of the sky.Source 1Source 2 |
| GALEX image set | About 38,000 observations | Individual pointings taken under different conditions required calibration.Source 1 |
| Other UV inputs | Swift, FIMS/SPEAR and TD-1 | Separate instruments had to be placed on a common scale and coordinate system.Source 1 |
| Gap-filling context | Visible, infrared and radio observations | The prediction learned relationships between known UV intensity and other wavelengths.Source 1 |
| Stellar layer | More than 100 million Gaia stars | UV contributions were estimated from visible-light measurements.Source 1Source 3 |
Why the ultraviolet sky had holes
Earth’s ozone blocks most ultraviolet light, so an all-sky map requires space instruments. GALEX deliberately avoided very bright stars, especially along the crowded plane of the Milky Way, because they could damage its detectors. Other missions filled some areas, but their coverage, resolution and calibration differed. Merging them is closer to building a consistent measurement system than stitching photographs together.[1][2]
Claude’s agents downloaded surveys, corrected internal differences, removed glare and atmospheric backgrounds, remapped images to one resolution and coordinate system, then combined the sources. These are individually familiar astronomical operations. The novelty in the report is the scale and orchestration: parallel agents performed the long sequence, while the scientist specified methods, inspected outputs and redirected the process.
The missing third was inferred from relationships
The system trained on regions where UV measurements and other wavelengths overlap. It learned how ultraviolet brightness relates to visible starlight, infrared dust emission and radio structures, then applied those relationships where UV data did not exist. This is inpainting with scientific side information rather than an artistic fill. The output is still a model estimate, and its uncertainty should travel with every use.[1]
Three kinds of information in the finished resource
| Layer | What it represents | How it should be used |
|---|---|---|
| Measured UV | Direct observations from ultraviolet missions after calibration | Treat as observational data, subject to instrument and processing uncertainty.Source 1 |
| Predicted diffuse UV | Model estimate for sky regions without direct UV coverage | Use with the measured/predicted flag and uncertainty; do not cite as a telescope detection.Source 1 |
| Estimated stellar UV | More than 100 million star contributions inferred from Gaia visible-light measurements | Useful for a complete educational view; individual estimates remain model-dependent.Source 1Source 3 |
A held-out test gives a bounded result
To test the gap-filling method, known UV regions were hidden and reconstructed without their real values. After several refinements, Anthropic reports predictions within about 10% of the actual UV measurements. That is evidence for the specific reconstruction procedure on held-out observed areas. It is not proof that every genuinely unobserved region has the same error, especially where the relationships between wavelengths differ from the training sky.[1]
The human found the artifact the agents missed
Ménard noticed faint circular footprints in a dim field. They were residual backgrounds from individual GALEX pointings. Claude had recognised the issue early in the project, yet the map passed two agent review rounds without the circles being caught. After the human identified the pattern, the system traced the cause and reprocessed all 38,000 observations. The episode is a useful model of scientific oversight: the expert did not redo the pipeline, but supplied the anomaly judgement that made the pipeline better.[1]
The collaboration produced more than a dozen map versions over several days. That makes the result less like one spectacular prompt and more like ordinary scientific computing accelerated by an agent team: formulate, process, inspect, find a failure, revise and repeat.
The project’s wider significance is not that AI looked at the sky and discovered a complete truth. It is that a scientist could finish a neglected but valuable data product by delegating the repetitive computational labour while retaining responsibility for calibration, anomaly detection and interpretation. The first complete UV map is strongest when its completeness does not conceal how much of it remains inference.
Sources & further reading
- The missing map of the sky Anthropic · 8 Oct 2026 · Accessed 10 Oct 2026
- GALEX: Galaxy Evolution Explorer NASA Science · Accessed 10 Oct 2026
- Gaia mission European Space Agency · Accessed 10 Oct 2026