Maps of the whole sky exist at visible, infrared, radio and X-ray wavelengths, but not, according to this post, in the ultraviolet: the largest UV survey left about a third of the sky unobserved. The post, published on Anthropic's Science blog on 8 October 2026 under the page heading "The missing map of the sky", is a first-person account by Brice Ménard, an astrophysicist at Johns Hopkins University and a researcher at Anthropic. He describes directing agents in Claude Science, an Anthropic product, to build a complete UV map over several days, mainly as a teaching resource. The post is a narrative essay with no tables. The map's own web page, linked from the post, is a second version of record. It carries the technical description, the validation numbers, the downloadable data and a draft paper, and it states that "nothing here has been peer-reviewed by humans yet." This note reports both, and says which number comes from which.
Why no complete UV map existed
The ozone layer absorbs ultraviolet light, so UV observations must come from space. The post names NASA's GALEX mission (2003 to 2013) as the largest dataset: about 38,000 separate observations covering about two-thirds of the sky. GALEX avoided very bright stars, including much of the plane of the Milky Way, to protect its detectors (section "Why no complete UV map existed"). Other telescopes, among them NASA's Swift and the Korean FIMS/SPEAR instrument, filled some gaps but not all.
The map page gives the coverage more exactly: GALEX covered 64% of the sky in the far-UV band and 76% in the near-UV, and FIMS/SPEAR covered 81% of the sky in the far-UV at coarse, degree-scale resolution (map page, "From incomplete to complete map" and "Datasets"). On this note's reading, the post's "roughly a third of the sky has never been observed in UV" therefore refers to detailed imaging; part of that third has coarse far-UV measurements.
The post's reason the map had not been made is practical. Filling the gaps properly takes "weeks of painstaking work" in calibration and analysis, and astrophysicists set such work aside for research. The author presents the project as an example of a backlog task that AI assistance now makes affordable.
What was done
The instructions and the division of labor
The author's instructions, as the post states them, were to gather every available UV dataset, put the datasets on a common scale, merge them, and fill every unobserved patch (section "Building the missing map"). Claude orchestrated a team of agents. They searched the web for public UV surveys, downloaded them, and made each survey internally consistent, including removing the glare of bright stars. They then cross-calibrated the surveys against one another, put them on a common resolution and coordinate system, and merged them. For the merging step, the author writes that he "gave high-level instructions to Claude on how to do so."
Filling the gaps
For the unobserved sky the author asked for inpainting, which the post describes as a machine-learning technique in which a model trained on many images learns how each part relates to its surroundings and restores missing regions. The post adds that Claude used the parts of the sky observed in UV to learn how UV brightness relates to maps at visible, infrared and radio wavelengths, then applied that relation where UV data are missing, with an uncertainty for each estimate.
The map page describes the filling model differently and in more detail. It is "an ordinary statistical regression (an ensemble of small decision trees), not a language model and not an image generator" (map page, Validation). Specifically, it is a gradient-boosted regression on 90 features derived from Planck dust maps, a hydrogen-alpha (Hα) map, Gaia starlight maps and sky position, trained on the star-subtracted GALEX sky with six-fold spatial cross-validation (Methods, "Learned components"). The map page builds the result by a fixed priority: a measured pixel always wins, a coarser or translated measurement beats a prediction, and every predicted pixel is labeled. The predicted share is 28% of the far-UV sky (most of it already set at degree scale by FIMS/SPEAR) and 27% of the near-UV (map page, "From incomplete to complete map").
Finally, Claude added the UV light of individual stars. The post says "more than 100 million" stars inferred from Gaia's visible-light measurements; the map page gives 118.9 million Gaia stars, with fluxes either measured by earlier UV catalogs or predicted from Gaia data.
The human role
The author's described contributions are the instructions above, a series of planning exchanges, and visual inspection. Over "several days" the two produced "more than a dozen successive versions of the map"; the author planned each step in conversation and Claude then ran hours of computation while he worked on other projects (section "The final map"). The acknowledgments say: "My contribution was limited to guiding them in the process." The post does not name the Claude model used.
What it reports
Accuracy of the filled sky
The post reports that, after several rounds of refinement, the model estimated deliberately hidden patches of observed sky "to within about 10% of the real UV measurements" (section "Building the missing map").
The map page gives larger figures for the same kind of test. Hiding patches of real GALEX sky and predicting them back, it finds "the typical error of filled sky is about 12–14% (0.05–0.06 dex), against ≈5% for the photographed sky, growing with distance from data" (map page, "From incomplete to complete map"). The held-out scatter of the diffuse prediction is 0.047 dex in the far-UV and 0.057 dex in the near-UV, rising to about 0.08 dex several degrees from data (map page, "Photometric checks", which says these numbers come from the paper). In percent, 0.047 dex is about 11% and 0.057 dex about 14% (our arithmetic). A chart on the page shows the far-UV error of filled sky falling from about 45% with GALEX alone to about 12% in the finished map, as the other surveys and the prediction are added. The page captured on 10 October 2026, with the map dated 16 September, already gave the 12 to 14% figure. The post does not say which statistic its "about 10%" summarizes, so the difference cannot be resolved from the documents.
The artifact a person caught
The post's clearest account of where the agents fell short is a calibration error (section "Trial and error"). In one of the dimmest fields, the author noticed faint circles, each slightly brighter or darker than its neighbors. They were the footprints of individual GALEX observations, each carrying a slightly different residual glow from Earth's atmosphere. Claude had listed this as a known issue at the start, yet the map "had still passed two rounds of review by other agents without the problem being caught." After the author pointed it out, the agents corrected all 38,000 observations in a couple of hours. The map page records the same episode: "Automated review missed them twice; a person looking at the picture did not." The post says the circles were then gone. The map page's caveats, in the 21 October capture, say footprints of single GALEX fields "remain at the few-% level": near-UV fields observed after May 2010 are 2 to 7% fainter at their edge than at their center. That residual is an edge-to-center falloff, not the field-to-field level offset the post describes, so the two statements need not conflict (our reading).
What the map shows
The finished map, in the author's description, shows dust clouds around young stars, loops traced by gas bubbles from exploding stars, faint high-latitude filaments of dust, and the Magellanic Clouds (section "The final map"). The post presents it as an educational resource that can "join the other canonical maps" in teaching.
Novelty and verification
The post's title calls the result "the first complete map of the sky in UV light." The map page's 10 October capture describes it as "the first detailed version with no holes in it." Earlier work produced partial maps: Murthy (2014), whose GALEX background maps set this map's zero point, made maps of the diffuse UV radiation "over the entire observed sky" (arXiv abstract). The map page treats the FIMS/SPEAR far-UV maps, which include the Galactic plane, as an input. Neither the post nor the map page compares the result with an earlier all-sky composite, and no outside assessment was found beyond news coverage that restates the post.
On checking, the map page is explicit: "nothing here has been peer-reviewed by humans yet", and "the validation reports come from the same system that built the map." The draft paper linked from the map page, dated 14 October 2026 and described as "in preparation", could not be retrieved for this note: the site blocks scripted access and no archived copy was found. The map page also limits use of filled sky. It "cannot hold a discovery," should not be searched for sources or used as an independent dust tracer, and users who need real photons are told to keep only pixels with provenance codes 0 to 2.
The instructions given
The post quotes the task only in summary and gives one verbatim instruction, the author's message about the artifact: "I can see discs with the imprint of individual observations; can you correct that?" It does not publish the prompts, the agents' configuration or the review agents' instructions. It says Claude itself "had listed this as a known issue at the start of the project", so a written note of the problem existed and was not acted on until a person raised it.
Limitations and open questions
- Not yet reviewed. By the map page's own statement, the map and its validation have not been peer-reviewed by humans, and the validation was done by the system that built the map. The draft paper was not available to this note.
- Two accuracy figures. The post's "about 10%" and the map page's 12 to 14% describe similar held-out tests; this note could not establish whether they measure the same thing.
- Predicted is not measured. About 28% of the far-UV map and 27% of the near-UV map are predictions from dust, Hα and starlight templates. The map page warns that this sky cannot reveal new sources and is not an independent dust tracer.
- One project, told by its author. The post reports one success by the person who directed it. It gives no account of time or cost beyond "several days," names no Claude model, and does not say whether other approaches were tried and abandoned.
- Map versions change. The map page has been updated since the post (the map is dated 16 September in a 10 October capture and 14 October in a 21 October capture), so numbers should be cited with the map's date.
Research questions
- What does the post's 10% figure measure? Comparing it with the draft paper's validation section, once available, would show whether it is the same held-out statistic as the map page's 12 to 14% or a different one.
- How would the filled regions compare with new UV imaging? Any future UV imaging of the Galactic plane or the bright-star holes would test the prediction where it matters most, far from GALEX data.
- Why did two rounds of agent review miss the footprint artifact? Recording what the review agents were asked to check, and testing them on maps with seeded artifacts, would show whether review by agents catches visual calibration errors.
- How much human time did the project take? A log of the author's exchanges and inspection time, set against the "weeks of painstaking work" the post cites, would measure the saving the post describes.
- Will the map hold up in peer review? The outcome of review of the paper, and any changes it forces to the uncertainty model, would settle how far the map can be used quantitatively.
How it connects
- Long-running Claude for scientific computing describes another Science-blog project in which agents worked for days under a scientist's occasional supervision, there checked against a reference code. This connection is the note's own.
- Introducing our Science Blog set out the blog's aims, including accounts by researchers of how AI changes scientific practice. This connection is the note's own.
References
- Ménard, B. (2026, October 8). Using Claude Science to produce the first complete map of the sky in UV light (page heading "The missing map of the sky"). Anthropic. https://www.anthropic.com/research/the-missing-map-of-the-sky record
- Ménard, B. (2026). The full sky in the ultraviolet (map, data and methods page). https://menard.pha.jhu.edu/uvmap/. Read through Internet Archive captures of 10 October 2026 (snapshot 20261010154619) and 21 October 2026 (snapshot 20261021070444); the live page blocks scripted access. record
- Ménard, B. (2026). Full-sky ultraviolet maps from harmonised GALEX, Swift-UVOT, FIMS/SPEAR, TD-1 and Gaia data (draft paper, in preparation). Linked from the map page; not retrieved for this note. record
- Murthy, J. (2014). GALEX diffuse observations of the sky: The data. Astrophysical Journal Supplement Series, 213, 32. arXiv:1406.5680. Abstract read for this note. record