Pavel Rabtsevich reports two transit-like signals in TESS photometry of TIC 4206066, a nearby late-K dwarf about 116 light-years away. The main candidate repeats every 3.18 days, dims the star by roughly 0.04 to 0.05 percent, and would correspond to about 1.4 Earth radii if the signal is on the target star. A second, weaker 11.13-day signal may correspond to about 2.2 Earth radii, but is explicitly treated as tentative.
The signal was first noticed in TESS Sector 98, then traced backward into independent older TESS observations from Sectors 32 and 6. The author says the workflow used Claude Code heavily for data down and parsing, transit-search code, fitting, nearby-star checks, false-positive checks, plotting, and reruns; Codex and separate clean agent sessions were used for independent review. The key framing is not "AI found a planet," but "AI agents helped run a falsification-heavy analysis loop."
Technical checks mentioned:
- Leave-one-observation/year-out prediction: fit on other data, predict hidden transits, then check only those windows.
- Nearby-star and pixel-localization checks.
- Gaia astrometry, photometry, and radial-velocity vetting.
- TRICERATOPS false-positive runs, with reported FPP around 0.03 to 0.04.
- Search across 36 catalogs/literature sources and 340,505 automated TESS alerts.
- Audit caught at least one overstrong statistical claim, which was removed.
Important caveat: this is not statistically validated and not confirmed as a planet. The main remaining alternative is an unresolved bound companion; high-resolution imaging and radial velocities would be needed to strengthen or rule that out.
The strongest part is the preregistration. TESS approved DDT Program 100 for 2-minute cadence observations of TIC 4206066 in Sector 110, from 2026-10-31 to 2026-11-26. Before those observations exist, Rabtsevich published fixed predictions and decision rules: 9 expected transits for signal A, 3 for signal B, expected depths, processing steps, and outcome categories: recovered, inconsistent, excluded, or insufficient data. Frozen code and checksums are included.
Sources:
- Reddit post: https://www.reddit.com/r/ClaudeAI/comments/1wzw8zd/i_think_i_found_a_planet_nobody_knew_existed_i/
- Preprint: [https://doi.org/10.5281/zenodo.22967456](https://doi.org/10.5281/zenodo.22967456)
- Supporting code/data: [https://doi.org/10.5281/zenodo.22967099](https://doi.org/10.5281/zenodo.22967099)
- Preregistered Sector 110 test: [https://doi.org/10.5281/zenodo.23175179](https://doi.org/10.5281/zenodo.23175179)
- Official TESS DDT list showing Program 100: [https://tess.mit.edu/science/ddt/](https://tess.mit.edu/science/ddt/)
Tiny translation: he did not ask Claude "is this a planet?" and blindly post the vibes. He used agents as tireless lab interns with paranoia turned up, then preregistered the future test so he cannot move the goalposts later. That is the good stuff.