Hobbyist's AI agents flag a super-Earth candidate among 126,000 TESS stars

Rabtsevich's search covered 126,000 stars ^[2]. The account does not say how many candidates it raised. Rabtsevich said "the research itself was mine: the question, the plan, the rules, the computing, the decisions on every candidate and what to publish" ^[11]. The phrase "every candidate" suggests TIC 4206066 was not the only one the search raised. Scientific American does not report how many others there were, how many survived vetting, or what the agents cost to run, so neither a false-positive rate nor the price of the tooling can be worked out from the account.

The method is built around distrust of the agents. Rabtsevich used several, and by his account they "wrote the code, ran it and checked each other's work" ^[10]. He is explicit about what their agreement is worth. "Models agreeing with each other isn't evidence," he said ^[8]. What he relies on is "a signal anyone can measure, calculations anyone can rerun, control samples," along with rules published before the next data sets arrive ^[9].

That design rules out two agents vouching for the same bug, but it cannot rule out an instrumental cause. In the 2018 data, the dips tended to align with moments when TESS fired its thrusters, and the article reports that this may have skewed the results ^[6]. He found 22 further dips in the archive ^[4]. His own summary is a "repeating transit-like signal in open data, with a plausible planetary explanation and alternatives that aren't yet ruled out" ^[7].

Pablo Mercader Perez, a PhD student at the Massachusetts Institute of Technology and a TESS team member, attributes the clearer signal to additional observation of that patch of sky. "Thanks to this, we got much more data than we had before for this particular star system, and the signal became a lot clearer," he said ^[15]. In statistical significance, he said, the candidate "became effectively detectable only in January of this year" ^[16]. His team's deep learning model has also flagged it ^[17]. Prof. Michelle Kunimoto's LEO-Vetter software selected it in early 2026, a result Perez described as not public yet ^[18]. Counting the 2020 filter, that makes three searches besides Rabtsevich's that picked up the signal ^[21]. In Perez's account, the change between the 2020 rejection and detection was mostly the volume of data.

Greg Laughlin, who once hunted planets as an astronomer at Yale University and has since moved into the AI industry, said the potential for at least one exoplanet around the star seems "solid" ^[12]. He called the analysis "a great example of human-AI teamwork" and added that large language models such as Claude and Codex have proven effective at combing astronomical literature for overlooked discoveries ^[13]. Laughlin did not say whether that record extends to light-curve searches. The literature work Laughlin cites is a different task from the one Rabtsevich ran.