Claude Code helped find a planet candidate: what TESS will actually verify

What is true in the post and what needs clarification

Such news should be verified in two steps: first, find a description of the author's work itself, then — independent confirmation of the part of the story that can be verified by another organization. The number of reposts does not replace either of these steps.

• "A vibe coder discovered a new planet." More accurately: Pavel Rabtsevich reported an exoplanet candidate. The author himself does not call it a confirmed planet.
• "Found it in old images from 2018." According to his description, the initial signal was in data from late 2025; then he checked observations from 2020 and 2018. The subject of analysis is changes in brightness, not a photograph of the surface of a new world.
• "A year on the planet lasts three days." The detected period of the main signal is approximately 3.18 days. This will be the planet's orbital period if the planetary explanation is confirmed.
• "NASA took the calculations to the bank." More specifically, it has been officially confirmed: the TESS list includes active DDT program No. 100 for this star. This is an approval for observation, not a confirmation of discovery.

Sources for these clarifications: the author's post and a the official TESS mission DDT list. Next, we separate the author's results from the mission's confirmations.

Star TIC 4206066 and a barely noticeable drop in brightness

In the metadata of the preprint from September 25 the author describes two transit-like signals near star TIC 4206066. The main one repeats approximately every 3.18 days, has a depth of about 500 ppm — i.e., 0.05% — and is traced in 23 transits. If the object orbits the target star, its estimated radius is 1.4 ± 0.1 Earth radii. The second, weaker signal has a period of about 11.13 days. These are the author's results, which we have not independently recalculated.

A small dip in brightness is not yet an image of a planet. The transit method looks for changes in light that may occur when an object passes in front of a star from our perspective. NASA specifically emphasizes: not every dip is caused by an exoplanet; other astronomical reasons and glitches are possible. NASA's explanation for citizen science participants.

Imagine you are looking at a lamp across a room, and someone regularly covers a small part of the light. You can measure the timing of the dimming, but based on these measurements alone, you won't know exactly who or what is passing in front of the lamp. The analogy is simplified, but it explains the difference between a signal and its physical interpretation.

AI wrote the code. A human determined what to consider as evidence

According to Rabtsevich, Claude Code helped upload and process data, write search code, build graphs, and check alternative explanations. For individual revisions, he used Codex and new agent sessions. The author notes that he himself chose the questions and verification criteria; approximately 74 analyses and over a thousand scripts accumulated in about two weeks. This is his description of the process, not our audit of all files. The original account.

The figure "a thousand scripts" might sound like proof of seriousness, but the amount of code does not measure the correctness of a conclusion. One incorrect processing method, repeated a thousand times, remains incorrect. In this case, it is much more important which checks could have failed and how the author handled such results.

For vibe coding this is a useful lesson: before launching an agent, it is worth defining what it should do, which data can be trusted, and what result would make you abandon a hypothesis. An agent can quickly perform technical work. It should not inconspicuously change the rules for evaluating its own result.

Why "candidate" is not a cautious word used merely for formality

The preprint's abstract explicitly states that statistical validation is not claimed. Among the alternatives remains a scenario in which the signal is associated with an unresolved stellar companion. The TRICERATOPS calculation gives a probability of a false-positive scenario of 0.03–0.04 within the applied model. Author's preprint.

These numbers should not be relayed as "it is 96–97% proven to be a new planet." The model estimate depends on assumptions, data, and considered alternatives. It also does not automatically establish which star the object orbits or whether its discovery is new to scientific catalogs.

There are at least three different questions: is there a repeatable signal; what physically causes it; and is the light source correctly identified. If these are combined into a single "planet found," the reader loses the main part of scientific work—distinguishing between alternatives.

We are not adding claims about surface composition, habitability, or guaranteed temperature to the story. For this article, there is no need to make them stronger than the available data. A potential exoplanet is interesting even without the promise of a "second Earth."

What could be confirmed regardless of the author

There is indeed an active entry on the MIT TESS page DDT 100: Pavel Rabtsevich, TIC 4206066, two-minute cadence, one target. The description contains the phrasing "planet candidate." The DDT procedure involves review of the application by the mission manager, a scientific expert, and the operations team. Official DDT program.

Important detail: this is not about the telescope flying to a planet. TESS observes patches of the sky. A two-minute cadence means the time step of data for a selected target, not that the "telescope will approach" the object.

Official Sector 110 page indicates the observation period October 31–November 26, 2026. As of the date of this publication, it is still ahead. Therefore, reporting the results of this specific future test would be premature right now.

An approved application is a significant step: the hypothesis has gained the opportunity for verification through new measurements. The decision to conduct an observation is logically different from the decision that an object is confirmed. If we call the former the latter, we erase the very reason for additional observation.

The strongest part of the story: the forecast before the data appears

On October 6, the author published the pre-registered test protocol for Sector 110. Its description records the expected transit times, data processing, and outcome rules. For the primary signal, nine events are predicted; for the second, three. Four predicted conclusions: the signal is recovered, it does not match the forecast, it is excluded within the test, or there is insufficient data.

The protocol itself clarifies: archival data has already been reviewed during signal selection, so this is not an independent new test. New measurements provide a different verification. At the same time, a repeated signal does not automatically prove a planetary nature, and an inconclusive result does not prove its absence.

This is why the previous forecast is so useful. Upon seeing a new graph, it is easy to find a convincing explanation after the fact. If time windows, processing, and criteria are set in advance, it is much harder to covertly tailor the story to the result. This is our assessment of the methodological utility, not an independent validation of this candidate.

Moreover, observation and the publication of ready data are different stages. The DDT page explains that data enters the archive along with other sector materials. Therefore, one should not expect a final conclusion simply on the day observations end. DDT data delivery order.

Practical benefit: learning how to test your own hypothesis

To try citizen science, you don't necessarily have to start with an automated search of a large archive. NASA suggests Planet Hunters TESS: participants learn to read light curves, review data, and discuss results with other participants and scientists.

For your own small study, we suggest the following process. This is an author's recommendation, not a reproduction of Rabcevich's entire analysis:

1. Start with a known example. Take training data with an explained result to check if you understand the method.
2. Preserve the data provenance. Record the source, date, units of measurement, and all filters.
3. Ask the AI to show intermediate steps. A graph without a description of processing is hard to verify.
4. Create a control example. Conditions where the sought signal is absent should also yield the expected answer.
5. Set aside a portion of the data. Write down the forecast before using them for evaluation.
6. Allow for an "I don't know" result. Insufficient data is a separate conclusion, not necessarily a victory or a defeat.

Help verify this hypothesis, rather than confirm it.
First, name alternative explanations and assumptions.
For each test, indicate what should happen if I am wrong.
Do not change the input data. Show units and intermediate results.
Separate completed verifications from proposed ones.
If there is insufficient data, state it directly.

This prompt is an unverified educational template. It does not replace astronomical knowledge or code verification. To work with convincing but unconfirmed answers, use our practical AI fact-checking guide.

What this means for the future of science and Vibe Coding

Our conclusion: such a case shows a promising way to use agents — helping humans ask more verifiable questions of open data. The most valuable skill here is not getting a beautiful answer from AI. It is building a verification process after which you are ready to change your mind.

At the same time, rapid coding can also accelerate the emergence of false conclusions. If every interesting curve immediately results in a published "discovery," information noise will increase. Therefore, the human role shifts to task setting, data control, choosing alternatives, and verifying interpretations. This is a forecast of possible practice development, not a claim that AI has already replaced scientists.

If this candidate is confirmed, the story will become an example of a significant result of working with AI assistants. If not, the open protocol and honestly disclosed negative result will also be beneficial. Science moves forward when it is possible to find out exactly what withstood verification.

Sources and limits of our verification

Verified October 9, 2026. We read the official mission pages, the author's post, and Zenodo record descriptions via API. This is an editorial fact-check, not a repeat scientific analysis: we did not run all the author's scripts, did not recalculate radii, and did not verify every point of the full PDF protocol. The preprint is not presented as a published peer-reviewed article.

• TESS: active DDT program 100 and application review order.
• TESS: sector 110 calendar.
• Rabtsevich: preprint on two transit-like signals, September 25.
• Rabtsevich: pre-registered test, October 6.
• Author's description of using Claude Code and Codex.
• NASA: transits and false signals.
• NASA: Planet Hunters TESS.