Discovering an Exoplanet Candidate with Claude Code: How AI Can Do Real Science

A Reddit user shared a genuinely pioneering experience: using Claude Code to help analyze observational data from NASA's TESS space telescope, he discovered an exoplanet candidate that had never been recorded before. This isn't just another "AI writes code" story — it's a real-world case study in how to use AI agents as rigorous scientific tools.

An Unusual Star 116 Light-Years Away

The discovery started with a star that "just didn't look right." Located about 116 light-years from Earth, it showed a roughly 0.05% dip in brightness every 3.18 days, lasting approximately two hours — the classic signature of a planet transiting across the face of its star.

The author first noticed this dip in more recent TESS data. He then took a critical step: tracing back through historical observations. The same transit signal appeared consistently in two completely independent sets of TESS data from years earlier. Three separate, independent datasets from different time periods all produced the same periodic signal — strong evidence that this was a real celestial object rather than noise.

Based on the transit depth, if this is indeed a planet, its radius would be approximately 1.4 times that of Earth — putting it in the "super-Earth" category. The author repeatedly emphasized one point: this remains a planet candidate, not a confirmed planet. That restraint is itself a mark of scientific integrity.

TESS (Transiting Exoplanet Survey Satellite) is a full-sky survey telescope launched by NASA in 2018. Its core detection method is the "transit method": when a planet passes in front of its star, it blocks a small fraction of the star's light, causing a brief, periodic, and tiny dip in observed brightness. A 0.05% brightness change is nearly imperceptible in everyday life, but falls within the detectable range for a precision photometer. The transit method is currently humanity's most productive means of discovering exoplanets, having helped confirm thousands of them. TESS divides the sky into sectors and photographs each one in turn, observing each sector continuously for about 27 days before moving on. This means the same star may be revisited multiple times across different years during the mission — which is precisely what allowed the author to find consistent transit signals across three independent datasets.

What the AI Agent Actually Did

What makes this case genuinely valuable to the tech community is how the author used AI. He relied on Claude Code (running the Opus 5.5 and Fable 5.1 models) to handle nearly all of the concrete execution work:

• Downloading and parsing raw TESS data
• Writing signal-search code
• Fitting transit light curves
• Checking for interference from nearby stars
• Searching for secondary eclipses and other false-positive indicators
• Generating plots and automatically re-running scripts when tests failed

The human researcher handled the higher-level judgment: what questions to ask, which tests actually mattered, and what criteria defined "pass" or "fail." Over roughly two weeks, this work grew into 74 independent analyses and more than 1,000 scripts.

The author distilled one principle worth remembering for anyone using AI: when doing science with agents, the right question isn't "is this a planet?" — it's constantly giving the agent opportunities to prove "it isn't." This falsification-oriented mindset is the core defense against being misled by AI outputs that seem plausible but are wrong.

Using Cross-Audits to Combat the AI Persuasion Trap

The danger with large language models is that they can produce an analysis that is internally consistent, apparently airtight — and completely wrong. To counter this, the author built a multi-layered audit system.

Independent verification came from Codex, as well as from brand-new agent sessions started from scratch with read-only access. This "have a completely independent mind re-check the work" approach mimics peer review in scientific research.

His most satisfying test was actually quite simple: he withheld an entire year of observational data, used the remaining two years to calculate the orbit, predicted exactly when transits should have occurred during the hidden year, and then went back to check only those predicted timestamps. He did this three times — for 2018, 2020, and 2025 — and all three predictions hit (3/3).

The audit process genuinely caught errors. At one point, a statistical conclusion appeared "stronger than it actually was." The audit flagged the problem, so he deleted that conclusion and redid that portion of the analysis. The core 3.18-day signal held up through all scrutiny.

The LLM "persuasion trap" is fundamentally a miscalibration problem: there is no positive correlation between a model's output fluency and its factual accuracy. Confidently worded answers and uncertain ones are nearly indistinguishable on the surface. Researchers call this "hallucination," but in data analysis contexts the harm is more insidious — the model doesn't fabricate numbers from thin air; instead, it selects real data, applies real statistical methods, then introduces a subtle logical error or overfits at some intermediate step, ultimately delivering a conclusion that looks impeccable but is inflated. The exaggerated statistical conclusion the author discovered in this case is a textbook example of this failure mode. Introducing an independent agent session with read-only access is equivalent to running a "blind review" on the same data — the new session shares no context with the original analysis chain and therefore cannot inherit whatever biases may have accumulated.

Exhaustive Literature Search to Confirm "Nobody Found It First"

To confirm the signal hadn't been previously recorded, the author searched through 36 star catalogs and literature sources, as well as 340,505 TESS automated planet-search alerts. Neither signal had ever been reported.

Additionally, there may be a second candidate around the same star — with a radius roughly 2.2 times Earth's and an orbital period of 11.13 days. But the author acknowledged that the evidence for this one is considerably weaker, so he treated it separately rather than conflating it with the primary signal. This tiered treatment of evidence at different confidence levels is another hallmark of scientific rigor.

Pre-Registration: Making the Conclusion Truly Falsifiable

The most impressive aspect of this case is how far it pushed the concept of falsifiability.

TESS will re-observe this region of sky from approximately October 31 to November 26. The author submitted an observation proposal requesting two-minute cadence monitoring of the star — the proposal was approved as Program #100.

More critically, before any new data was collected, he publicly published the exact predicted timestamps for when transits should occur, along with the explicit rules for what would constitute a "pass" or "fail." This means the conclusions cannot be quietly revised after the fact: if the dip appears as predicted, the evidence will be substantially strengthened; if it doesn't appear, that counts as evidence against the candidate.

He also built a free interactive 3D system model for anyone to "fly through," and publicly released the DOI links for the preprint paper and prediction data. This open, pre-registered, verifiable workflow is exactly what AI-assisted science should look like.

Pre-registration is a crucial practice in modern scientific research for combating outcome bias: researchers publicly document their hypotheses, methods, and decision criteria before collecting or analyzing new data, eliminating the possibility of cherry-picking favorable results or reinterpreting data after the fact. This practice gained widespread adoption following the replication crisis in psychology and has since spread to astronomy, medicine, and other fields. In the context of AI-assisted research, pre-registration takes on even greater importance: large language models are naturally adept at generating plausible-sounding narratives. If researchers are allowed to decide "what story to tell" after seeing AI outputs, systematic bias becomes nearly impossible to detect externally. Publicly committing to specific transit timestamps and pass/fail rules before any new data arrives places the entire analysis chain under public scrutiny that cannot be rewritten — and that is the most methodologically instructive design choice in this entire case.

What This Means for AI Research Applications

This story matters far beyond a single astronomical discovery. It demonstrates a model for human-AI collaboration: AI handles the heavy lifting of data processing and code execution; humans control the scientific judgment and verification standards; and multiple independent agent cross-audits guard against "convincing but wrong" results.

The author closed his post with a question worth the entire community's reflection: when an AI-generated result looks completely convincing and yet could still be wrong, how do you audit it? As more and more people use AI for actual research rather than just writing software, the answer to that question may matter more than discovering any single planet.