Markdown edition

# Claude Science helped fill gaps in an ultraviolet sky map

> The UV sky map mixes telescope measurements with predictions; its masked-patch test does not validate regions never observed.

By BIG CHANGE Editorial

Published: 2026-10-09T15:07:06.437Z
Updated: 2026-10-09T15:07:06.437Z
Canonical: https://bigchange.ai/blog/claude-science-ultraviolet-sky-map-estimates

![A generic unbranded space telescope faces a star field with a broad unfilled area beside softly textured regions; the sky pattern is conceptual, not survey data.](https://bigchange.ai/api/media/file/claude-science-ultraviolet-map-hero-v1.png)
Conceptual illustration of an orbital telescope facing a partly unfilled sky map; the star field and gap are not actual ultraviolet survey data. AI-generated illustration by BIG CHANGE.

The full-sky ultraviolet map assembled by astrophysicist Brice Ménard with Claude Science labels each pixel by provenance and supplies uncertainty estimates. Its technical page classifies 28% of the far-ultraviolet band and 27% of the near-ultraviolet band as predicted. Other regions draw on observations from several missions, including GALEX, Swift-UVOT and FIMS/SPEAR; some far-ultraviolet values are inferred from near-ultraviolet data. [Ménard’s interactive map](https://menard.pha.jhu.edu/uvmap/) lets readers inspect the sky and switch between ultraviolet, visible and infrared views.

## The big change

- **What changed:** Claude Science helped Ménard coordinate the collection, calibration and merging of decades of ultraviolet survey data, then fill remaining gaps with statistical estimates. The finished map labels measured and predicted pixels separately.
- **Why it matters:** Students can see structures across the whole sky in ultraviolet light. Researchers can use the same view as a guide, provided they keep the provenance and uncertainty layers attached to quantitative work.
- **What to watch:** The technical paper is still listed as “in preparation,” and the map page says downloadable data products will be added. Independent scientific assessment and accessible files will determine how far the visualization can be reused beyond education.

Ultraviolet light is absorbed by Earth’s ozone layer, so astronomers observe it from space. NASA’s GALEX mission produced about 38,000 images between 2003 and 2013, but it avoided regions around bright stars to protect its detectors. Swift and South Korea’s FIMS/SPEAR mission added observations, yet no single survey had filled every gap. The missing map was partly a labor problem: observations from different instruments needed to be cleaned, calibrated to a shared brightness scale and aligned before they could be combined.

In an [October 8 account](https://www.anthropic.com/research/the-missing-map-of-the-sky), Ménard, a Johns Hopkins University astrophysicist who is also an Anthropic researcher, says Claude Science coordinated agents that searched for public surveys, downloaded data, standardized images and combined datasets at common resolutions and coordinates. The work also required removing foreground glow and correcting artifacts around bright stars. The technical page describes the estimated diffuse ultraviolet component as an ensemble of gradient-boosted decision trees using dust, hydrogen and Gaia starlight data. Claude Science coordinated the workflow; the published description does not say that a language model itself generated the missing sky pixels.

For this project, the reported AI contribution is a workflow: Claude Science managed data collection and processing, while a separate statistical model estimated the missing diffuse ultraviolet. No new telescope observations were added. Ménard describes the result as an educational resource that makes the ultraviolet structure of the Milky Way easier to explore.

Ménard tested the estimates by hiding patches where ultraviolet measurements already existed, asking the model to estimate them, and comparing predictions with the concealed values. His October 8 account summarized the error as about 10%. The technical page gives typical errors of 12–14% for predicted sky and specifies held-out scatter of 0.047 dex in far ultraviolet and 0.057 dex in near ultraviolet. The sources do not explain whether those figures use the same version or summary metric, so the headline “10% accurate” would overstate the precision of the report.

A hidden-patch test checks whether a method can reconstruct data from parts of the sky that resemble its training examples. It cannot directly establish accuracy in places that were never observed, especially where the map relies on coarser inputs or different conditions. The technical page says that the map’s effective resolution is 0.5 to 1 degree in predicted regions, even though the map is sampled on much smaller pixels. The small pixel spacing does not give predicted regions finer resolution than their source data.

The project’s account describes a separate quality-control limit. Ménard noticed faint circular footprints left by GALEX observations after two rounds of review by other agents had missed them. He directed Claude to trace the uneven atmospheric glow and correct it across the mission’s 38,000 observations. The example shows that agents could carry out a broad correction once a researcher identified the defect; it also shows why visual review and domain judgment remained part of the work.

For readers exploring the map, the practical rule is to keep its labels and uncertainty information with any analysis. Ménard’s page cautions that predicted regions cannot support discovery of new sources and should not be treated as an independent dust tracer. It also says the 1.7-arcminute pixel spacing is not the map’s effective resolution everywhere, and that some systematic errors are correlated across regions. The map is a useful whole-sky visualization. Its inferred parts are estimates built from existing data and should be treated as such.

The current research page says the map has not yet been peer reviewed, lists Ménard’s technical paper as in preparation and says FITS and HiPS downloads will be added. Until the paper and products are available for independent inspection, the map’s own provenance labels, stated limits and validation description are the evidence readers have for judging it.

## Sources & further reading

- [The missing map of the sky — Anthropic, October 8, 2026](https://www.anthropic.com/research/the-missing-map-of-the-sky). Ménard’s first-person account of the project, data workflow, inpainting and masked-region check.
- [The full sky in the ultraviolet — Brice Ménard, Johns Hopkins University, October 2026](https://menard.pha.jhu.edu/uvmap/). The interactive map and technical description of data provenance, resolution, uncertainty, reported validation and the paper’s “in preparation” status.
- [Anthropic’s Claude Science creates the first complete ultraviolet map of the sky — The Decoder, October 9, 2026](https://the-decoder.com/anthropics-claude-science-creates-the-first-complete-ultraviolet-map-of-the-sky/). Independent coverage of the announcement; it does not provide an independent reanalysis of the map or its validation.

## Sources

- [The missing map of the sky](https://www.anthropic.com/research/the-missing-map-of-the-sky) — October 8, 2026 first-person account by project participant Brice Ménard; primary source for the reported workflow and masked-patch test.
- [The full sky in the ultraviolet](https://menard.pha.jhu.edu/uvmap/) — Ménard’s October 2026 interactive map and technical description; primary source for provenance, model type, resolution, uncertainty and paper status.
- [Anthropic’s Claude Science creates the first complete ultraviolet map of the sky](https://the-decoder.com/anthropics-claude-science-creates-the-first-complete-ultraviolet-map-of-the-sky/) — The Decoder, October 9, 2026; independent coverage of the announcement, not independent measurement or performance validation.