The telescope chose the holes

For about fifty years the standard picture of the ultraviolet sky has had a hole in it roughly the shape of our own galaxy. Brice Ménard, who teaches astrophysics at Johns Hopkins and does research at Anthropic, writes that he used to apologise to his students for it: every other wavelength had a complete map to show, and the UV one was full of gaps. On 8 October he published a complete one, built over several days with Claude Science. The caption under it says that about a third of the map, including much of the galactic plane, was predicted rather than measured.

That is an unusually honest caption, and it is also the most interesting sentence in the project. The missing third is not missing by accident. NASA's GALEX mission, which ran from 2003 to 2013 and supplied the largest piece of the data, was deliberately steered away from fields containing very bright stars, including those along the plane of the Milky Way, because pointing at them risked damaging its detectors. So the sky that had to be filled in by a model was selected, by an instrument, for being bright. The model that filled it was fitted on the sky that was not, and then checked against the sky that was not. The number that licenses the whole exercise, an agreement of about 10% on held-out patches, was measured in the one region where the method was never needed.

What was actually computed

Ultraviolet light is absorbed by the ozone layer, so the only way to see the UV sky is from orbit. GALEX imaged roughly two-thirds of it in some 38,000 separate observations; NASA's Swift and South Korea's FIMS/SPEAR added more, and gaps remained. Ménard's instruction to Claude Science was to gather every public UV survey, put them on a common scale, merge them, and fill in the rest.

The first two steps are enormous bookkeeping rather than discovery, which is why nobody had done them: this is a figure a field wants and no individual career needs. Claude orchestrated a team of agents that searched for the surveys, downloaded them, removed the glare that bright stars smear across neighbouring pixels, cross-calibrated instrument against instrument, resampled everything to a common resolution and projected it onto one coordinate system. Many agents worked in parallel on different regions.

The filling in is where the statistics live. The method is inpainting, borrowed from image restoration: learn how each part of an image relates to its surroundings, then reconstruct the parts that are gone. Here it was given more to work with than neighbouring pixels. Where there is no UV data there are still observations at other wavelengths, and the credit line on the figures names Planck and hydrogen-alpha maps among the templates used for gap-filling. So the model learned, on the observed two-thirds, a relation between UV brightness and visible, infrared and radio brightness, and applied that relation where no UV exists, returning an estimate and a confidence for each pixel. On top of that diffuse background Claude added the UV contribution of more than 100 million individual stars, inferred from Gaia's visible-light measurements.

To check it, Ménard asked Claude to take regions that do have UV data, hide parts of them, and reconstruct the hidden values. After several rounds of refinement it recovered them to within about 10% of the real measurements.

Why that check cannot carry the map

This is the oldest trap in missing-data analysis, and it is worth naming precisely, because a learner will meet it in a spreadsheet long before they meet it in a sky survey.

Write \(y\) for the UV brightness of a pixel, \(x\) for everything else we know about that pixel, and \(m = 1\) if the pixel was observed. Imputing \(y\) from a model fitted where \(m = 1\) is safe when the reason for missingness carries no information about \(y\) once you condition on \(x\): missing at random, in the standard taxonomy. It is unsafe when the missingness depends on \(y\) itself, which is called missing not at random, and which cannot be diagnosed from the observed data, because the evidence that would settle it is the part you do not have.

GALEX's pointing rule was a function of brightness. That is missing not at random by construction, and it is the textbook version rather than a subtle case. The held-out test estimates

Those two quantities are the same number only if the relation between UV and the other wavelengths holds as well in crowded, bright, dust-dense fields as in the faint sky off the plane. There is a real physical reason to hope it does, and no way to test it with the data in hand. Meanwhile the specific nuisance the pipeline fought hardest, glare from bright stars bleeding into the pixels around them, is worst exactly where the predictions have to stand alone. The shift from the training region to the application region is not a mild change in the input distribution; it is a move toward the conditions under which every step of the calibration is least reliable.

None of that makes the map wrong. It makes the 10% a statement about interpolation inside the observed sky, which is a different and easier claim than the one a reader will take from it.

There is a second lesson in the project's own account of what went wrong. Looking through processed images one evening, Ménard noticed faint circles in one of the dimmest fields, each slightly brighter or darker than its neighbours. They were the footprints of individual GALEX pointings: each image carries a faint, uneven glow from Earth's atmosphere, and if that glow is not fully removed the pointing's own disc remains visible. Claude had listed this as a known issue at the start of the project. The map had still passed two rounds of review by other agents without it being caught. Once told, the agents traced it correctly and corrected all 38,000 observations in a couple of hours.

That failure is structural, not careless. Claude Science's reviewer agent, as its launch post describes it, inspects outputs for incorrect citations, untraceable numbers and figures that do not match the code that produced them. Every one of those checks asks whether the artifact is faithful to the pipeline. None of them asks whether the pipeline was entitled to its assumptions. A reproducibility check cannot tell you that a background subtraction left a per-pointing residual, because the residual is reproducible. Writing in the same series a week earlier, Matthew Schwartz put the practitioner's version bluntly: look at everything yourself, always ask to see plots, and do not trust automated checks or a claim of good agreement. Here the plot was the oracle, and a person's eye was the instrument.

The strongest case for the map

The best objection is that this is not a generative model hallucinating a galaxy. UV dust emission really is physically coupled to the infrared and radio tracers used to predict it; astronomers have leaned on exactly that coupling for decades. The map reports a per-pixel confidence and, Ménard writes, ships extra layers that label each pixel as measured or predicted. It is offered as an educational resource, not as a catalogue for precision photometry. And the alternative is not a better map. The alternative is the hole, for another decade, because the weeks of pixel-level calibration this needed were never going to outrank anyone's actual research.

All of that is true, and it is why the map is worth having. But notice what the honesty depends on. The mask and the uncertainty layer live beside the image, and the image is what travels. The hero figure in the post is a single colour scale across the whole sphere, in which measured and predicted are indistinguishable by construction. The thing the map was built to show students is the rich structure of the Milky Way at this wavelength, which is to say the galactic plane, which is to say the part that was predicted. A learner looking at the band across the middle of that map is looking mostly at a model's best guess about the one region it could not be tested on, and nothing in the picture says so.

What to take from it

Three habits, in rough order of how often they will save you.

When you are handed a held-out score, ask what made the held-out data available. A score computed on the observed portion of a dataset describes the observed portion. If the rule that decided what got recorded was a function of the quantity you are predicting, the score is in-distribution and your use of it is not, and no amount of refinement closes that gap.

Keep "the model can reconstruct values that were hidden" separate from "the model can supply values that were never taken." The first is a statement about redundancy in the data you have. The second is an extrapolation, and it borrows its credibility from the first.

When a dataset arrives with a provenance mask, treat the mask as part of the data rather than as metadata. An average computed over this map without it is an average over a mixture of measurement and inference, weighted by where a satellite's safety rules happened to fall. Ménard's backlog argument tells you to expect many more artifacts of this shape: resources that were always worth making and never worth a career, now cheap enough to make in an afternoon, and cheap precisely because nobody's reputation is pinned to them.

One caveat about this piece: the egress policy where it was written reached the project's own published account and nothing else, so the map's layer viewer and the mission archives behind it are unopened, and every figure above is Ménard's.

For fifty years the UV sky came with an apology attached. The apology was the most accurate layer on the map: it said, in a shape anyone could see, here is where we have not looked. The complete version is better in every way except that one, and the hole it fills was never a hole in the sky.