Independent searches of NASA TESS light curves for small transiting planets. The first combines the newest TESS sectors (97–105) with every earlier sector of 9,979 nearby K and M dwarfs. The second, added on 8 October 2026, looks for additional planets around the hosts of 6,430 TESS Objects of Interest (TOIs) after masking the known TOIs.

• Full-sky search, second release (9 October 2026): 432 candidates (tiers A, B and N below) from 4.04 million FGKM dwarfs brighter than T = 13.5, combining all of their TESS sectors (1–104). Every one now passes TRICERATOPS (FPP < 0.5, NFPP < 0.1) and has MES ≥ 10. New in this release: 169 tier N candidates recovered from crowded fields with TRICERATOPS and a centroid test calibrated on 1,000 known planets; false-alarm rates measured on the inverted light curves of 400,000 stars; TRICERATOPS for every tier A and B candidate; and a search of the SNR 7.5–9 band, which turned out to be mostly noise (92 signals listed as tier L, not counted). The stricter checks moved 137 first-release candidates to tier C. With the 60 earlier candidates below, the running total is 492 candidates.
• First release (7 October 2026): 10 transit-candidate signals on 9 stars for which we found no prior report, including a possible third transiting planet in the TOI-4342 system, and two signals that others had reported earlier in 2026.
• Update (8 October 2026): 20 more candidates for which we found no prior report: 6 from completing the fully coherent search of the K/M dwarfs and 14 from the TOI-host search, plus two further signals on TIC 231725883. Every TOI-host candidate is on a star that already has a TOI, which makes it less likely to be a false positive.
• Update 2 (8 October 2026): 6 more candidates for which we found no prior report, from re-checking SPOC detections that have not become TOIs and from a search for periods of 20–100 days, plus independent recoveries of 2 signals that other independent researchers posted on Zenodo days earlier. They include possible second planets around the confirmed-planet hosts TOI-669 (matching a weak radial-velocity signal) and TOI-5997, a 50.3 d signal on TOI-4566, and a sub-Earth-sized signal on GJ 774, an M-dwarf pair 13 pc away.
• Update 3 (8 October 2026): 18 more from a search of 22,448 fainter M dwarfs (11.5 < T ≤ 12.5), including GJ 3514, an M4.5 dwarf 17.6 pc away, and a star with two signals. The same search found two of our earlier candidates again, and one signal that another independent project had posted on GitHub three days earlier.
• Update 4 (8 October 2026): 12 more from the faintest M dwarfs searched so far (12.5 < T ≤ 13.5, 84,863 stars), including LHS 1083 (36 pc) and a possible member of the young Octans association.

These are planet candidates, not confirmed planets. Each passes automated and pixel-level vetting, but TESS alone cannot rule out every false-positive scenario. Ground-based photometry, high-resolution imaging, or radial velocities are needed to confirm or reject them.

Radii use TIC 8.2 stellar radii. FPP is the TRICERATOPS false-positive probability from TESS data alone (no follow-up imaging, no multiplicity boost), averaged over 10 runs from the update onwards; the statistical-validation threshold is FPP < 0.015.

Each candidate has a folder in candidates/ with the vetting plot and metrics, the transit fit (P, T0, Rp/R*, b, ρ*), PRF centroid offsets, nearby stars bright enough to mimic the signal, and the TRICERATOPS result where computed. All parameters are in results_public/candidates.csv.

TOI-4342 b and c are confirmed sub-Neptunes at 5.54 and 10.69 d (Tey et al. 2023; masses from ESPRESSO, 2026, which also reports a probably non-transiting 47.5 d RV candidate). We find a third transit signal at 18.9082 d with the same radius as b and c (2.2 R⊕), with transits in 8 sectors (11σ in Sectors 13–93, 7σ in 101–103). The difference-image centroid is within 7–16″ of the star in each 2-min sector, and no star within 30″ is bright enough to produce the dip. Its predicted RV semi-amplitude is about 2 m/s. The TOI-host search, which masks b and c before searching, finds it again as the strongest remaining signal.

These passed every check: no match in the ExoFOP TOI and CTOI lists, the NASA Exoplanet Archive, the SPOC TCE tables or the literature check (below); all 17 LEO-Vetter tests; MES ≥ 7.1; per-sector depths consistent (χ²/dof ≤ 5); a difference-image centroid within 15″ of the target; no TIC star within 15″ bright enough to produce the dip; and no report found by a web search on each TIC ID and host TOI (novelty_check.md). "Known on star" lists the TOIs (with TFOPWG disposition) on the same star.

TIC 231725883: three signals. The 9.0988 d signal from the first release (also in the RAVEN first-stage list) is joined by signals at 4.6593 d (1.43 R⊕) and 18.9653 d (1.95 R⊕), each found by the coherent search after masking the previous one, each passing all LEO-Vetter tests and each with consistent depths across sectors. The period ratios are 1.95 and 2.08. A TIC star 13″ away could produce the 4.66 d and 18.97 d dips (but not the deeper 9.10 d one) if it were an eclipsing binary, so these two are listed separately in results_public/candidates_2026-10-08.csv, together with the other signals that pass everything except the 15″ neighbour check.

TIC 143168991 (TOI-7580), 26.56 d: rejected. The K/M dwarf search found it with 4 transits, but two of them fall within 0.2 h of TOI-7580.01 transits; with the TOIs masked only two remain.

Two more searches found 8 signals that pass every check: 6 for which we found no prior report, and 2 that other independent researchers posted on Zenodo a few days earlier (TRICERATOPS not yet run):

• SPOC detections that have not become TOIs. SPOC reports every signal above its threshold as a threshold-crossing event (TCE); TOIs are selected from these, and many TCEs are never followed up. We re-measured the 14,112 TCEs on 8,698 FGKM dwarfs that are not TOIs or CTOIs, adding all later sectors, and vetted them like our own signals (build_tce_targets.py, tce_refit.py, triage_tce.sh). ExoMiner++ (Valizadegan et al. 2025, 2026), a neural network that classifies every SPOC TCE, lists 17 of the 79 signals that passed every LEO-Vetter test as planet candidates, so those count as reported. Most of the signals it scores as false positives also fail our centroid test. TCEs from SPOC's latest multi-sector search (Sectors 1–96) have not yet been through TOI vetting, so some of these may still become TOIs.
• Periods of 20–100 days. A stack-slide search of the 3,661 K/M dwarfs and 3,271 TOI hosts with at least six sectors, with the TOIs and the signals from the earlier searches masked.

TIC 269728501 (TOI-4566), 50.32 d. Six transits in 17 sectors. SPOC's Sectors 1–96 search listed two of them as a 754.8 d TCE, 15 times this period.

TIC 235005571, 1.33 d: rejected. It passed every automated check, but an RNAAS note shows the star is an Algol-type eclipsing binary.

The TCE re-check's last row includes TIC 235005571 and the two signals reported on Zenodo.

A stack-slide search of 22,448 M dwarfs (Teff ≤ 3,900 K, R ≤ 0.6 R☉) with 11.5 < T ≤ 12.5, fainter than the first K/M sample and mostly observed only in full-frame images (TESS-SPOC and QLP light curves, Sectors 1–104), vetted and checked at the pixel level in the same way. It recovered 346 signals already known as TOIs, planets or TCEs. 18 signals pass every check and have no prior report; TIC 165827520 (2.00 d) and TIC 383313006 (4.25 d), from our update 2 list, were found again; and TIC 417732194 (G 249-11, 5.307 d) had been posted on 5 October by another independent project (tess-transit-hunter).

The same search of 84,863 M dwarfs with 12.5 < T ≤ 13.5. It recovered 625 known TOIs, planets or TCEs. 104 new signals pass every LEO-Vetter test; the pixel checks reject 91 of them (in these crowded fields mostly for a centroid offset or a capable neighbour), leaving 13. TIC 9994636 (0.894 d) is rejected: it is a double-lined spectroscopic binary (Kounkel et al. 2021), so its 5%-deep dip is probably an eclipse.

First release; the second release below re-tiers it with measured false-alarm rates.

We searched every FGKM dwarf in TIC 8.2 with T ≤ 13.5, Teff ≤ 6,500 K and R ≤ 1.5 R☉ that our earlier searches had not covered: 4,166,638 stars, 4,086,108 of them with TESS-SPOC, QLP or SPOC light curves (22.4 million light curves, Sectors 1–104), indexed from the STScI open-data bucket (build_allsky_targets.py, index_s3.py) and searched with stack-slide on AWS Graviton instances in eight parts (run_shard.py).

The pixel checks reject 93% of the signals that pass LEO-Vetter, mostly for a centroid offset or a neighbour that could produce the dip: in full-frame images most signals at these depths come from blended eclipsing binaries. All candidates, with fits, centroids and binary checks, are in results_public/candidates_allsky.csv; tiers A–C have a folder in candidates/allsky/.

Reliability tiers. Following an outside review, every candidate is now checked against Gaia DR3 (RUWE, eclipsing binaries, non-single-star orbits), the TESS eclipsing-binary catalog (Prša et al. 2022), APOGEE binaries (Kounkel et al. 2021) and VSX (binary_check.py), which flags both binaries that earlier passed every automated check (TIC 9994636 and TIC 235005571). Tier A (several signals on the star) and tier B (single) are counted as candidates; tier C (grazing, or RUWE > 1.4) is listed but not counted. Of the 66 earlier candidates, 24 are tier A, 36 tier B and 6 tier C (results_public/tiers_earlier_candidates.csv).

Reliability and completeness, measured on a 1% pilot (40,797 random stars; superseded by the 399,408-star measurement in the second release below). Inverted copies of 20,430 light curves, in which any dip is a false alarm, give 24 signals that pass LEO-Vetter and the cuts (SNR ≥ 7.5), and none survives the pixel checks; the real light curves give 14 (12 at SNR ≥ 9, 2 at 7.5–9). Of 300 transits injected at a nominal (white-noise) SNR of 10, 57% are recovered at the injected period and 21% at SNR ≥ 9: red noise in full-frame light curves lowers the measured SNR. These are candidates, not validated planets; we expect false-positive rates of tens of percent in tier B, lower in tier A, and the remaining blind spot is companions closer than ~1–2″, which only high-resolution imaging can rule out.

Is the centroid test discarding planets? The pixel checks rejected 93% of the signals that pass LEO-Vetter. We ran the same vetting and pixel checks on 1,000 TOIs on stars like ours (601 confirmed or known planets and 399 planet candidates, T = 10–13.5, using the TOI ephemerides). The centroid test wrongly rejects only 2–3% of them (5.6% at MES 7–20), and 95% of the confirmed and known planets lie within 9.3″ of the target. Of 203 full-sky signals that match a TOI on a neighbouring star, the test flags 98–100% of those whose source is more than 15″ away, and the measured offsets track the separations (medians of 19″, 32″, 51″ and 73″ for sources at 15–21″, 21–42″, 42–63″ and beyond). The 7,918 signals rejected for an offset are therefore mostly blends (validation/centroid_calibration_tois.csv, validation/centroid_calibration_neighbour_tois.csv).

Signals with a neighbour inside 15″ (tier N). The other pixel check, a TIC star within 15″ bright enough to produce the dip, also rejected 31% of the known planets, because crowded fields are common. We ran TRICERATOPS (Giacalone et al. 2021) on the 1,318 signals rejected only by this check or with no usable difference image (fpp_batch.py): FFI cutouts from the TESS cubes on S3 (s3cut.py), Gaia DR3 field stars from VizieR, ten runs per signal, and nearby-star scenarios for neighbours within 30″, since the centroid excludes sources farther out. TRICERATOPS does not use the centroid, so its probabilities are combined with a centroid likelihood ratio: how likely the measured offset is if the transit is on the target (the offsets of the 1,000 TOIs at similar MES) or on a capable neighbour (the same scatter around the neighbour's position). On the calibration samples this ratio favours the neighbour for 83–100% of the neighbour-caused signals whose source is 5–21″ away (75% within 5″), and the target for 88–96% of the TOI planets when tested against a neighbour 8–14″ away (76% at 5″). Tier N requires a combined FPP below 0.5 and NFPP below 0.1, MES ≥ 10, and the same binary, grazing and literature checks as tiers A and B: 169 candidates. 111 of them have NFPP < 0.01, and 13 have FPP < 0.015 and NFPP < 0.001, the TRICERATOPS validation thresholds, though validation also needs high-resolution imaging. The FPP_c, NFPP_c and BF columns of the catalog give the combined values and the centroid ratio.

False alarms, measured. We searched the inverted light curves of 399,408 random full-sky stars, in which every dip the pipeline finds is noise or a systematic, with the full pipeline down to SNR 7.5 (validation/false_alarms_inverted_399408_stars.csv). Signals passing LEO-Vetter and the cuts, by pixel-check outcome, with the inverted counts scaled to the 4,039,203 stars searched:

False alarms concentrate at low multiple-event statistic (MES): about half of the clean SNR ≥ 9 signals with MES < 10 are expected to be false alarms, 16% of those with MES 10–12, and fewer than 6% above (90% upper limit). We therefore moved the 78 first-release candidates with MES < 10 to tier C, and tier N requires MES ≥ 10. TRICERATOPS cannot recognise noise: on the 24 inverted false alarms with a close neighbour it passed 5, but only 1 of them has MES ≥ 10, so we expect about 10 false alarms among the 169 tier N candidates (validation/triceratops_inverted_close_neighbour.csv). These rates apply to the full-sky search; the earlier samples keep their tiers, because their hosts (TOI hosts and M dwarfs) have far more planets per star, so the same false-alarm rate per star is a much smaller fraction of their candidates.

The SNR 7.5–9 band. Injections show why a lower threshold is tempting: of 700 transits injected at a nominal SNR of 10 into random full-sky targets, 22% are recovered at SNR ≥ 9 and 48% at SNR ≥ 7.5 (validation/injection_recovery_allsky_snr10_n700.csv). We re-downloaded the light curves of the 267,000 stars with a signal at SNR 7.5–9 and vetted their 302,000 signals the same way (collect.py --from-signals): 382 pass every check, but about 270 of them are expected to be false alarms. The 92 with MES ≥ 9 that also pass the binary and grazing checks are listed as tier L (about 30% false alarms expected) and are not counted.

TRICERATOPS for tiers A and B. We ran TRICERATOPS on all 322 tier A and B candidates that remained after the false-alarm cut, with the same settings and the centroid ratio over capable neighbours within 30″. The median FPP is 0.05 (0.025 for tier A), and 41 candidates have FPP < 0.015 and NFPP < 0.001. The 59 that fail the tier N thresholds (FPP < 0.5, NFPP < 0.1) look like eclipsing binaries (median radius 8.9 R⊕ against 3.6 R⊕ for the rest, median impact parameter 0.78) and move to tier C, so every counted full-sky candidate meets the same bar. In all, 137 first-release candidates moved to tier C (78 for MES < 10 and 59 for TRICERATOPS), and 4 tier A candidates became tier B because the other signal on their star moved.

All 2,426 signals, with fits, centroids, binary checks and TRICERATOPS results where computed, are in results_public/candidates_allsky.csv (column set: SNR ≥ 9, SNR 7.5–9, or neighbour within 15″); tiers A, B, C, N and L have folders in candidates/allsky/.

candidates/predictions.csv lists every predicted transit of the first-release candidates through June 2027 with 1σ uncertainties (5–15 min). TESS observes TIC 61816225, 458191673 and 451641911 again in Sectors 111–114, and TIC 422351625 and 135425484 in Sector 114. This file was committed before those data exist.

The approach follows Pavel Rabtsevich's search that found TIC 4206066 (write-up, data): detect