Knew nothing about astronomy, dug through NASA data with AI, and found a new planet candidate

The star TIC 4206066, 116 light-years away, dims by about 0.05% for roughly two hours every 3.18 days. If a planet is passing in front of the star, that planet is about 1.4 times Earth's radius. Rabtsevich checked 22 dimming events, and all 22 landed on the times predicted by a 3.18-day period. Rabtsevich also found a weaker signal that repeats every 11.13 days, and wrote: "The evidence for that one is much weaker." X +1

Rabtsevich ran a holdout test three times: hide one year of data, then use the remaining years to predict when the star would dim in the hidden year. Each run held out a different year (2018, 2020, and 2025), and all three predictions were right. Rabtsevich also checked the signal against 36 star catalogs and papers and 340,505 automated TESS detection alerts to make sure it had not already been reported. That cross-check covers only public lists. According to Scientific American, the signal was flagged once by a TESS detection filter in 2020 and then dropped for lack of evidence. A search run with LEO-Vetter, a tool astronomer Michelle Kunimoto built to weed out false signals, also picked out this candidate in early 2026 but did not make it public. Over about two weeks, the work with Claude Code and Codex produced 74 analyses and more than 1,000 scripts. reddit +2

Claude Code (Opus 5.5 and Fable 5.1) handled downloading the data, writing the search code, fitting the shape of the dips, checking nearby stars and false signals, drawing plots, and rerunning failed tests. Rabtsevich wrote: "I chose what questions to ask, what tests mattered, and what counted as pass/fail." Codex and a fresh agent session with read-only access reviewed the results separately. reddit

On Sept. 25, Rabtsevich posted a preprint to Zenodo. It has logged 6,542 views and 900 downloads. The approved list of TESS Director's Discretionary Targets (DDT) includes Rabtsevich's proposal to observe this star at two-minute intervals. The entry reads: "Pavel Rabtsevich | Two-minute cadence for TIC 4206066, a nearby K-dwarf with a 1.4R planet candidate", with approval marked "yes". Because of this approval, TESS will record the star's brightness every two minutes during its next observing period. zenodo +1

TESS photographs the sky one patch at a time, and each of these observing periods is called a sector. Sector 110, which covers this star again, runs from Oct. 31 to Nov. 26. On Oct. 6, Rabtsevich pre-registered the times the star should dim during Sector 110, based on both the 3.18-day and the 11.13-day signals. The 3.18-day signal predicts nine dips and the 11.13-day signal predicts three, with a timing uncertainty of about four minutes. The three earlier holdout tests (2018, 2020, and 2025) used data that already existed. This is the first time a pre-registered prediction will be checked against new observations. At the end of the X thread, Rabtsevich linked a 3D site that lets you fly out to the star: tic4206066.pages.dev. zenodo +1

After reading Rabtsevich's post, Thick_Ad_9875 used Opus 5.5 to take a second look at about 2,900 small stars within 200 light-years. The targets were stars TESS had observed across several years, but only in its wide full-frame images, so no search program had ever analyzed those years of data together. Before starting, Thick_Ad_9875 made sure the search setup worked by using it to recover Rabtsevich's signal. reddit

Before getting the 2026 data, Thick_Ad_9875 locked in a record of when each candidate star should dim and the criteria for judging the result, so neither could be changed later. One candidate failed this prediction test. Two dimmed at the predicted times and stayed on the list. reddit

Of the survivors, TIC 451641911 is 134 light-years away and dims by about 0.06% every 3.91 days. TIC 26432942 is 112 light-years away and dims by about 0.07% every 10.39 days. If they are planets, both are about 1.3 times the size of Earth. Neither appeared in any planet or candidate list Thick_Ad_9875 could find, or among roughly 312,000 automated TESS alerts. reddit

Both passed all 17 checks in the vetting tool LEO-Vetter. Another vetting tool, TRICERATOPS, put the probability of a planet orbiting the star at 96.5% for TIC 26432942 and 76.7% for TIC 451641911. The test results and predicted times are public on Zenodo. reddit +1

After seeing Rabtsevich's Reddit post about finding a planet candidate with Claude Code, Chopra spent a morning combing through data on nearby stars that TESS had only captured at long intervals, using an analysis program written with Claude Code. X

TIC 370612437, 158 light-years away in the direction of Cygnus, dims by 0.02% every 3.56 days. If the signal comes from a planet, it is almost exactly Earth's size. HIP 114417, 134 light-years away toward Andromeda, dims by 0.03% every 4.93 days, which points to about 1.1 times Earth's size. Across four observing years between 2019 and 2024, the dips repeated 55 and 21 times, respectively. X

In the holdout test, where one year of data is hidden and the rest is used to predict that year's dips, the predictions were right in three of the four years. Codex and a freshly launched Claude session reviewed the method and results separately. Chopra posted the results on X with a 50-second video and announced plans to report them as community candidates (CTOIs) on ExoFOP, NASA's site for sharing exoplanet candidates. As of Oct. 19, though, the ExoFOP page for TIC 370612437 shows no community candidate entry yet. X +1

Krone had Codex and Claude coding agents analyze archival data from TESS and from ZTF, another sky survey. The result was a candidate outburst: LP 119-10, a cooling white dwarf whose surface pulsates, suddenly brightened in December 2024. Its brightness rose by about 6.5% in TESS data, which covers red to near-infrared light, and by more than 31% in ZTF data taken through a green filter (g band). The two numbers measure brightness in different colors of light. A pulsation with a 575.9-second period also showed up. Fewer than 10 pulsating cooling white dwarfs are known to have outbursts. zenodo

The preprint says "a second AI system repeated the analysis independently from the raw archive files." zenodo