. · * · .
Claudestronomy
Grants for finding things in public sky data.
Some things about this:
- Telescopes have outrun analysis. Years of full-frame images, light curves and catalogues sit in public archives, searched once, under one set of assumptions.
- Anyone with a laptop and an agent can search them again. People already are (see below).
- We give compute, Claude credits and a small stipend to people who want to do this seriously.
- The data is free. All of it. TESS, Gaia, ZTF, DECam, HSC, WISE.
- If a candidate earns telescope time, we help you ask for it.
- Nulls count. A calibrated non-detection with a completeness curve is a result.
· . + . ·
What we fund
- Slow movers. Distant solar system bodies that move less than a pixel per month.
- Shift-and-stack TESS full-frame images. Cross-match against ZTF, DECam, HSC, Pan-STARRS.
- Planet Nine, Planet Y, sednoids, the inner Oort cloud.
- An object at 550 AU drifts about 0.13 TESS pixels across a 27-day sector. One pixel is 21 arcseconds. This is why nobody has looked properly.
- Transits and variables. Signals the automated pipelines skipped.
- Single transits, long periods, faint hosts, blended neighbours, eclipsing binaries nobody catalogued.
- Vet against Gaia. Run false-positive statistics. Preregister what the next sector should show.
- Cross-archive oddities. Objects that only look strange when two surveys are joined.
- Infrared excess with no optical source. High proper motion with no parallax. A transient with one detection in each of three catalogues.
| Archive | What is in it | A first project |
|---|---|---|
| TESS | A full-frame image every 200 s since 2018; 2-min light curves for 200k+ stars; thousands of candidates with no follow-up | Re-search one sector's FFIs for what the pipeline missed |
| Gaia DR3 | 1.8 billion sources with parallax, proper motion, colour | Vet every TESS candidate host for an unresolved companion |
| ZTF | Nightly northern sky to r ≈ 20.5 since 2018 | Precover or kill a slow-mover candidate along its exact parallax track |
| DECam | Deep southern imaging to r ≈ 24 | Plane fields never linked across nights |
| HSC-SSP | Subaru imaging to r ≈ 26 over 1,000+ deg² | Faint-tail completeness beyond 400 AU |
| WISE | All-sky infrared, 10+ years of epochs | Cold objects with no optical counterpart |
* · . · *
Some things people have already found
- TIC 4206066. A candidate 1.4 Earth-radius planet on a 3.18-day orbit around a late-K dwarf 116 light-years away. Pavel Rabtsevich, October 2026.
- Seen in TESS Sector 98, traced back to Sectors 32 and 6.
- Claude Code ran the search, fits, nearby-star and TRICERATOPS checks. Separate Codex and clean-agent sessions audited it and removed one overclaimed statistic.
- Nine transits predicted for Sector 110 (Oct 31 to Nov 26), with depths and frozen code published before the data exist. FPP about 0.03. Still a candidate. Write-up.
- Two eclipsing binaries nobody had catalogued. Mehran Jalali, October 2026.
- Two red dwarfs, each about a seventh of a solar mass, 500 light-years away, orbiting every 3 h 26 min, closer together than the Sun is wide, with a brighter third star travelling with them.
- Two half-solar-mass stars 740 light-years away, eclipsing twice per 36.7-hour orbit and losing up to 40 percent of their light each time.
- TESS light curves joined with Dark Energy Survey, Legacy Surveys and Gaia imaging. Thread.
- Planet 26. A Planet Nine campaign across TESS, HSC, DECam and ZTF. Will DePue and agents, October 2026.
- Twelve parallel agent lanes on a shared GPU node. An append-only results log. A candidate review board. No lane may claim a search until it recovers injected objects.
- Result so far: a calibrated null. TESS alone covers about 1 percent of the plausible population at one false candidate per tile; the deep archives about 2.5 percent. No candidate survived cross-survey vetting.
- ExoMiner++. NASA Ames, March 2026. An open-source vetter that flagged about 7,000 TESS targets as planet candidates. Nobody has the hours to follow them up. That gap is the labour we want agents doing. NASA.
. + · + .
Grants
Draft numbers.
- Scout. $5k in Claude credits, 200 GPU-hours, six weeks.
- One archive, one question, one sector or field. Deliverable: a recovery test on known objects, then a candidate list or a clean null.
- Survey. $25k credits, 2,000 GPU-hours, $5k stipend, four months.
- A full search of one survey, or one object class across two. Deliverable: injection-calibrated completeness, preregistered follow-up predictions, a public write-up.
- Campaign. $100k+ credits, a dedicated node, stipend, six to twelve months.
- Multi-survey, multi-agent programs with a coordinator and a review board. By invitation after a Survey grant.
Rolling. Two pages is enough: the question, the archive, which known objects you will recover first, and what a null would be worth. Decision within two weeks. We fund the method, not the hoped-for headline.
Some things we believe
- Agents make it cheap to produce candidates and cheap to fool yourself. Every rule below came from a project above learning it the hard way.
- Recover known objects before searching for new ones. If the pipeline cannot find Sedna, it has not earned a blind search.
- Inject before you claim. Completeness comes from synthetic sources recovered at the flux and motion you care about, reported at a stated false-positive rate.
- Write the pass line down first. Thresholds and outcome categories are published before the data arrive. Frozen code, checksums.
- Separate the agent that searches from the agent that audits. A clean session with no memory of the hunt reviews the claim.
- Code, prompts and results are public. Someone should be able to kill your candidate without emailing you.
- Nulls are deliverables. "Nothing found" without a completeness curve is not.
· * . * ·
Apply
. · .
claudestronomy.com