An AI-ranked transit signal is a research candidate, not a finished discovery

AI ranking can improve a follow-up queue while scientific status still depends on diagnostic evidence, review and the meaning of catalogue labels.

An AI-ranked transit signal is a research candidate, not a finished discovery

A telescope records a small, repeating dip in a star's brightness. Software ranks the signal highly, and the research team has more promising targets than it can examine immediately. AI can make that queue more useful. It can also create a tempting shortcut: a confident classification begins to sound like a newly discovered planet. The useful scientific workflow keeps the score, the evidence and the eventual status of the object separate.

Prepared and source-checked on 17 October 2026, this analysis draws on NASA's January 2026 ExoMiner++ account, an explicitly versioned May 2025 author paper and NASA Exoplanet Archive definitions. These are related parts of a research ecosystem, not three independent replications. The proposed review queue below is hypothetical. No new planet, model test or telescope observation is claimed here.

Read the word “candidate” carefully

NASA reported on 22 January 2026 that an initial ExoMiner++ run identified 7,000 TESS targets as exoplanet candidates. Its account explicitly says candidates require follow-up observations to confirm, and describes distinguishing planet-like transit signals from other phenomena such as eclipsing binary stars. The reported candidate count should not be relabelled as 7,000 confirmed new planets. [NASA ExoMiner++ account](https://science.nasa.gov/open-science/deep-learning-exoplanets-tess/)

The ExoMiner++ paper's arXiv version four, recorded on 19 May 2025, describes diagnostic inputs, classification and ranking, including ambiguous TESS labels and multi-source training that combines Kepler and TESS examples. It discusses ranking as a way to prioritise follow-up work. That version's experiment is not automatically identical to the initial run described in NASA's later news account. [Author paper, version four](https://arxiv.org/html/2502.09790v4)

NASA Exoplanet Archive documentation separately defines confirmed planets, project candidates and candidates not yet dispositioned as confirmed or false positive. Its TESS project-candidate category can include events later identified as false positives. Reading these definitions is necessary before combining catalogue totals or treating a candidate list as a discovery count. [Archive category definitions](https://exoplanetarchive.ipac.caltech.edu/docs/counts_detail.html)

The distinction does not diminish the value of an effective classifier. A better queue can help researchers spend scarce attention on useful targets. Its contribution is clearer when the team reports which part of the discovery process became more reliable or efficient, rather than claiming the entire process was automated.

Build a review packet around the signal

In a fictional observing programme, each ranked entry could carry a stable target identifier, the data release, the model version and the diagnostic material used to produce the score. Preserve the original measurements and the processing choices that turned them into inputs. An attractive plot should not become detached from the observations and assumptions behind it.

The reviewer needs more than a single probability. A packet could show the transit-like pattern, relevant diagnostic views, missing information and the reason the target reached the queue. Keep a way to inspect the underlying data when the summary is misleading. A score can help arrange work without explaining every ambiguity in an individual case.

Use separate fields for model classification and researcher disposition. A reviewer might retain a candidate, request additional work or identify a known false-positive explanation. Record the evidence and reason without overwriting the original model output. This allows the team to examine failures later and avoids a dataset in which every human correction looks like something the model predicted initially.

Evaluate the queue the team can actually use

Decide whether the main goal is recovering as many plausible candidates as possible or putting the most useful candidates near the top of a limited follow-up list. Those goals can favour different thresholds. Overall classification accuracy alone may not tell the team how many worthwhile objects it will review during the available observing time.

Keep a held-out evaluation set and examine errors by relevant signal and data conditions. A model trained on one mission's examples may face different noise, sampling or label quality in another dataset. Compare the proposed ranking with the existing prioritisation process on the same eligible targets. Record when target selection itself depends on other scientific constraints, rather than pretending the model score is the only reason an object was observed.

Review a sample below the chosen threshold as well as the highest-ranked objects. Otherwise the programme may become good at confirming what it already prefers while learning little about what it misses. Include unresolved cases in the accounting. A target with no follow-up result should not be counted as a correct rejection or an unsuccessful discovery merely because its status is inconvenient for a summary chart.

Preserve the path from candidate to published result

The review queue should connect to the actual follow-up and publication process, not replace it. When new evidence changes a target's status, retain the link between the earlier classification and the later record. Make the date and basis of the change visible. A public explanation should distinguish a model identifying a promising signal from a research team establishing the interpretation.

That workflow also creates responsibilities for people maintaining data products. Someone must reconcile identifiers across releases, document processing changes and make correction paths usable. Junior researchers can learn from disagreements between diagnostics and classifications, provided the system makes those disagreements visible. Treating the model as an unquestionable gatekeeper would remove that learning opportunity as well as narrowing scientific scrutiny.

Open code and data can support inspection, but availability alone does not prove that another team reproduced a result. A useful reproduction record would state the exact version, inputs, processing and evaluation choices. If those differ, describe the exercise as an adaptation rather than reporting the original performance as if it had been independently recovered.

AI-assisted transit vetting can make large archives more approachable and help researchers choose what to inspect next. It cannot remove the need to interpret a signal, maintain careful labels and communicate scientific status accurately. The practical starting point is a traceable candidate packet and a review process that counts unresolved evidence honestly. A stronger queue is a meaningful result; it does not need to be advertised as a finished discovery.