Claude Science builds a first “complete” UV all-sky map—merging surveys and inferring the unobserved third

Why it matters

【Key points】

On Oct 8, 2026 (20:59 UTC) Anthropic’s research blog published Brice Ménard’s account—Johns Hopkins astrophysicist and Anthropic researcher—of using Claude Science to assemble the first “complete” all-sky map in ultraviolet light: coalesce public UV surveys onto one scale and fill the roughly one-third of the sky no UV telescope has ever observed.

Claude orchestrated agents to gather surveys (chiefly NASA GALEX’s ~38,000 observations), make each survey internally consistent, cross-calibrate instruments, resample to a common grid, then inpaint gaps using relationships learned from visible, infrared and radio maps where UV exists, plus UV estimates for >100 million stars inferred from ESA Gaia optical data. The finished map labels each pixel measured vs predicted and carries uncertainty. A hold-out test—hiding known UV patches and reconstructing them—landed within about 10% after refinement. The write-up also documents failure: leftover Earth-airglow footprints from individual GALEX frames slipped through until Ménard flagged them and agents reprocessed all ~38k frames. Anthropic frames the map mainly as an educational view of Milky Way dust, supernova rings and halo filaments in UV.

【Verifiable】

Anthropic research post: https://www.anthropic.com/research/the-missing-map-of-the-sky

Article published_time: 2026-10-08T20:59:00Z

Official X: @AnthropicAI https://x.com/AnthropicAI/status/2108290395599667700

【Our read】

1. Not a new model—a multi-day, auditable Claude Science case that stretches the research workbench beyond biology-heavy launch demos into astronomy data engineering.

2. “Complete” still means ~1/3 inferred; the ~10% self-test holds only where ground truth can be hidden, not automatically in never-observed sky. Labeling measured vs predicted is the load-bearing disclosure.

3. For readers, the signal is pipeline discipline—messy surveys, cross-calibration, failure rollback—more than another pretty science image.