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.
- Gaia limits on companions (11 October 2026): 546 candidates. We reran TRICERATOPS on every counted candidate with limits from Gaia DR3: MOLUSC removes bound companions that Gaia would have resolved or that would raise the star's RUWE, and for bright stars the Gaia epoch RVs rule out an eclipsing binary on the target. The median false-positive probability falls from 0.042 to 0.013, and 236 candidates meet the validation thresholds apart from high-resolution imaging (62 before). With the same bar applied to the constrained values in both directions, 5 candidates left and 10 joined, and two stars that Gaia shows to be spectroscopic binaries moved to likely eclipsing binaries: 488 full-sky candidates and 58 from the earlier searches (details).
- Full-sky search, second release (9 October 2026): 483 candidates (tiers A, B, N and L1 below) from 4.04 million FGKM dwarfs brighter than T = 13.5, combining all of their TESS sectors (1–104). Every one passes TRICERATOPS (FPP < 0.5, NFPP < 0.1) and a false-alarm test calibrated on the inverted light curves of 800,000 stars. New in this release: 169 tier N candidates recovered from crowded fields with TRICERATOPS and a centroid test calibrated on 1,000 known planets; 51 tier L1 candidates, mostly small planets, picked out of the noisy SNR 7.5–9 band by a false-alarm classifier; and TRICERATOPS for every tier A and B candidate. The stricter checks moved 137 first-release candidates to tier C. With the 60 earlier candidates below, the running total is 543 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.
| TIC | Period (d) | Radius (R⊕) | Host (Teff, distance) | FPP | Status |
|---|---|---|---|---|---|
| 354944123 (TOI-4342) | 18.90820 | 2.21 ± 0.11 | 3880 K, 62 pc | 0.14 | no prior report found; third transiting signal in a system with two confirmed planets |
| 61816225 | 12.77703 | 1.99 ± 0.06 | 4380 K, 57 pc | 0.03 | no prior report found |
| 458191673 | 2.570687 | 2.31 ± 0.07 | 4470 K, 66 pc | 0.03 | no prior report found |
| 422351625 | 8.14641 | 2.35 ± 0.16 | 4410 K, 93 pc | – | no prior report found; second signal on the same star ↓ |
| 422351625 | 1.969879 | 1.46 ± 0.09 | 4410 K, 93 pc | – | no prior report found |
| 166321785 | 2.480128 | 2.08 ± 0.09 | 4380 K, 96 pc | – | no prior report found; two T≈15–16 stars at 9″ |
| 147226218 | 3.87159 | 1.57 ± 0.08 | 4840 K, 77 pc | – | no prior report found; T=16.6 star at 3.4″ |
| 451641911 | 3.90974 | 1.36 ± 0.05 | 3650 K, 41 pc | – | no prior report found; crowded field |
| 30470520 | 18.75190 | 1.45 ± 0.04 | 5150 K, 208 pc | – | no prior report found; crowded field, depth varies by sector |
| 135425484 | 8.03350 | 3.12 ± 0.15 | 4230 K, 79 pc | – | no prior report found; impact parameter 0.91 (near-grazing) |
| 286763141 (TOI-6284) | 5.244548 | 0.85 ± 0.04 | 3550 K, 21 pc | 0.03 | independent recovery; reported as "TOI-6284.02" by Tschudi 2026 |
| 231725883 | 9.098778 | 2.47 ± 0.07 | 5110 K, 155 pc | – | independent recovery; in the RAVEN first-stage detection list (Lafarga et al. 2026), not in their vetted list; the update finds two more signals on this star |
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 | Period (d) | Radius (R⊕) | Host (Teff, T mag, distance) | Known on star | FPP | Notes |
|---|---|---|---|---|---|---|
| 32636748 | 0.732324 | 1.43 | 4094 K, 10.4, 55 pc | – | unstable | K/M dwarfs; ultra-short period; TRICERATOPS gives 0.00–0.98 over 10 runs for this 0.7 h transit |
| 117799904 | 3.291093 | 1.39 | 4045 K, 10.9, 61 pc | TOI-2457.01 (PC, 14.20 d) | 0.053 | K/M dwarfs; second signal on TOI-2457 (found by both searches) |
| 124094573 | 13.721761 | 2.18 | 4590 K, 11.4, 110 pc | – | 0.068 | K/M dwarfs |
| 242080609 | 16.955247 | 1.89 | 4500 K, 9.7, 54 pc | – | 0.026 | K/M dwarfs; T=9.7, 5 transits |
| 257484419 | 18.056783 | 2.58 | 4184 K, 11.2, 81 pc | – | 0.016 | K/M dwarfs |
| 179985715 | 19.092785 | 2.56 | 4047 K, 11.3, 71 pc | TOI-249.01 (PC, 6.62 d) | 0.132 | K/M dwarfs; second signal on TOI-249; grazing (b=0.9); Gaia RUWE 2.0 |
| 317134140 | 0.571863 | 1.31 | 5563 K, 10.7, 174 pc | TOI-2514.01 (PC, 6.39 d) | 0.066 | TOI hosts; ultra-short period |
| 72668830 | 2.000544 | 1.59 | 4985 K, 10.4, 116 pc | TOI-4296.01 (FA, 19.97 d); TOI-4296.02 (PC, 5.10 d); TOI-4296.03 (PC, 12.49 d) | 0.105 | TOI hosts |
| 165827520 | 2.004481 | 1.15 | 3230 K, 12.3, 34 pc | TOI-3494.01 (PC, 7.75 d) | 0.088 | TOI hosts; a star 0.59″ away (ΔI = 2.9) is known from imaging (Matson et al. 2025), so the radius is a lower limit |
| 383313006 | 4.251808 | 2.05 | 3719 K, 11.8, 73 pc | TOI-6758.01 (PC, 12.61 d) | 0.165 | TOI hosts |
| 380671216 | 5.575799 | 2.29 | 5729 K, 10.6, 191 pc | TOI-5495.01 (PC, 9.66 d) | 0.047 | TOI hosts |
| 423445163 | 6.212551 | 2.75 | 4593 K, 12.5, 187 pc | TOI-6489.01 (PC, 2.92 d) | 0.016 | TOI hosts |
| 441590110 | 6.462490 | 4.13 | 6263 K, 11.6, 436 pc | TOI-5187.01 (PC, 17.86 d) | 0.020 | TOI hosts |
| 410449566 | 7.109099 | 2.70 | 5949 K, 11.5, 239 pc | TOI-7546.01 (PC, 9.64 d) | 0.224 | TOI hosts |
| 402992891 | 8.005848 | 2.48 | 5589 K, 11.0, 164 pc | TOI-7356.01 (PC, 17.40 d) | 0.027 | TOI hosts |
| 468983280 | 8.966677 | 1.41 | 3547 K, 11.7, 45 pc | TOI-5489.01 (PC, 3.15 d); TOI-5489.02 (PC, 4.92 d) | 0.121 | TOI hosts; TOI-5489 b and c are confirmed planets, so this would be a third |
| 416739490 | 10.565600 | 2.03 | 5554 K, 10.8, 150 pc | TOI-5697.01 (PC, 6.22 d) | 0.061 | TOI hosts |
| 154437244 | 11.279747 | 2.84 | 6173 K, 11.1, 273 pc | TOI-5989.01 (PC, 6.87 d) | 0.015 | TOI hosts |
| 445837596 | 15.000796 | 2.42 | 5127 K, 11.7, 182 pc | TOI-3896.01 (PC, 4.38 d) | 0.020 | TOI hosts; TOI-3896 b is a confirmed planet |
| 160585062 | 17.392551 | 1.78 | 4568 K, 12.2, 157 pc | TOI-7333.01 (PC, 6.26 d) | 0.040 | TOI hosts; TOI-7333 b is a confirmed planet |
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 | Period (d) | Radius (R⊕) | Host (Teff, T mag, distance) | Known on star | Notes |
|---|---|---|---|---|---|
| 105506140 | 0.706375 | 1.42 | 5216 K, 8.0, 44 pc | – | TCE re-check; HD 85706; weakest: centroid 13.8" in the best difference image but 17.1" on average, grazing (b = 0.95), and ExoMiner++ scores its SPOC TCEs 0.34 and 0.45 (below its 0.5 threshold) |
| 300381700 | 3.145257 | 4.87 | 5999 K, 12.4, 598 pc | TOI-7387.01 (PC, 9.09 d) | TCE re-check; second signal on TOI-7387; reported on 4 October by Tovar Contreras (Zenodo 23134068) |
| 373017346 | 3.521493 | 1.79 | 5844 K, 10.6, 211 pc | TOI-4450.01 (PC, 10.69 d) | TCE re-check; second signal on TOI-4450 |
| 351339274 | 4.225136 | 0.75 | 3529 K, 9.2, 13 pc | – | TCE re-check; GJ 774 A; its companion GJ 774 B, 17.6" away and 1.2 mag fainter, could be the host instead (0.75 R⊕ on A, about 0.9 R⊕ on B) |
| 257484419 | 4.353306 | 2.45 | 4184 K, 11.2, 81 pc | – | TCE re-check; second signal on the star of our 18.06 d update candidate; reported on 3 October by Ozturk (Zenodo 23118499) |
| 124573851 | 9.529248 | 2.48 | 5625 K, 10.2, 143 pc | TOI-669.01 (CP, 3.94 d) | TCE re-check; second signal on TOI-669; matches a weak radial-velocity signal at 9.61 ± 0.52 d (Akana Murphy et al. 2023); confirmed: TOI-669 b |
| 39516274 | 14.216035 | 1.36 | 4674 K, 9.3, 46 pc | TOI-5997.01 (CP, 5.66 d) | TCE re-check; HIP 85850; second signal on TOI-5997, whose planet b the NASA Exoplanet Archive added on 8 October 2026, so this would be TOI-5997 c. Samson Fraser posted the same signal on Zenodo on 9 October (10.5281/zenodo.23249265), the day after this update |
| 269728501 | 50.319649 | 2.98 | 4443 K, 11.6, 119 pc | TOI-4566.01 (PC, 2.08 d) | 20-100 d; second signal on TOI-4566 (see below) |
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.
GJ 237 (TIC 293689266), 29.29 d: not counted. The TCE re-check also found this signal, and it passes every LEO-Vetter test (MES 12), but the star is half of a pair of early M dwarfs 3.7″ apart that the difference images cannot separate. TRICERATOPS makes it almost certainly a planet (eclipsing-binary scenarios < 0.1%) but cannot tell which star it orbits: 50% the primary, 36% the companion (FPP 0.48, NFPP 0.54), so it fails the NFPP < 0.1 bar we apply to signals with a close neighbour. Samson Fraser listed it as a candidate on Zenodo on 9 October (10.5281/zenodo.23249265), with the same host ambiguity.
TIC 235005571, 1.33 d: rejected. It passed every automated check, but an RNAAS note shows the star is an Algol-type eclipsing binary.
| Stage | TCE re-check | 20–100 d, K/M dwarfs | 20–100 d, TOI hosts |
|---|---|---|---|
| Stars searched | 8,698 | 3,661 | 3,271 |
| Signals | 14,112 | 9,179 | 6,421 |
| Pass pre-filters | 4,393 | 333 | 167 |
| Match a known planet, TOI or CTOI | 14 | 15 | 8 |
| Reported elsewhere (literature check, including ExoMiner++) | 251 | 2 | 3 |
| Match a neighbour's known signal | 30 | 4 | 0 |
| Unmatched and passing every LEO-Vetter test | 62 | 2 | 3 |
| … centroid more than 15″ from the target | 35 | 1 | 2 |
| … a TIC star within 15″ could produce the dip | 15 | 1 | 0 |
| … no FFI difference image, and SPOC's centroid is more than 15″ off | 3 | 0 | 0 |
| … none of these, MES ≥ 7.1, per-sector χ²/dof ≤ 5 | 8 | 0 | 1 |
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).
| TIC | Period (d) | Radius (R⊕) | Host (Teff, T mag, distance) | Notes |
|---|---|---|---|---|
| 100624547 | 1.922279 | 8.58 | 3541 K, 12.3, 77 pc | grazing (b ≈ 1), so the radius is poorly constrained (5.7–13.5 R⊕); crowded field (TIC contamination 0.32) |
| 308023474 | 2.074868 | 1.24 | 3687 K, 12.1, 62 pc | G 203-64 |
| 38017042 | 2.695824 | 1.44 | 3496 K, 12.4, 54 pc | two signals on this star (2.70 d and 4.14 d) |
| 355815567 | 3.519453 | 0.89 | 3149 K, 12.3, 18 pc | GJ 3514, an M4.5 dwarf 17.6 pc away |
| 38017042 | 4.141260 | 1.64 | 3496 K, 12.4, 54 pc | two signals on this star (2.70 d and 4.14 d) |
| 59214833 | 4.145226 | 2.00 | 3587 K, 12.2, 58 pc | crowded field (TIC contamination 0.25) |
| 137609439 | 4.294043 | 2.25 | 3339 K, 12.4, 61 pc | LSPM J1004+8023 |
| 102034645 | 4.760759 | 1.67 | 3671 K, 11.6, 56 pc | the radio galaxy ESO 243-29, 25" away and not in the TIC, adds ~39% of the flux, so the radius is underestimated |
| 333114636 | 7.510269 | 2.43 | 3894 K, 11.9, 87 pc | LSPM J0044+1748 |
| 386518973 | 9.159667 | 1.94 | 3467 K, 12.5, 55 pc | – |
| 380095420 | 9.574513 | 1.97 | 3745 K, 11.9, 77 pc | chromospherically active |
| 459222850 | 9.810046 | 2.03 | 3752 K, 12.3, 99 pc | – |
| 118244692 | 10.249236 | 2.16 | 3509 K, 12.2, 56 pc | H-alpha emission |
| 435924571 | 11.422839 | 2.60 | 3759 K, 12.4, 94 pc | HG 7-176 |
| 435255286 | 13.726825 | 2.10 | 3477 K, 12.2, 57 pc | G 5-38 |
| 117798466 | 14.286705 | 1.80 | 3653 K, 11.9, 67 pc | possibly young (space motion matches the AB Doradus moving group) |
| 32154846 | 16.107661 | 1.62 | 3816 K, 12.4, 102 pc | in the southern continuous viewing zone |
| 193718989 | 24.532923 | 2.58 | 3745 K, 12.2, 86 pc | – |
| Stage | Faint M dwarfs, 11.5 < T ≤ 12.5 |
|---|---|
| Stars searched | 22,448 |
| Signals found | 38,545 |
| Pass pre-filters | 2,281 |
| Match a known planet, TOI, CTOI or TCE | 346 (96 pass every LEO-Vetter test) |
| Reported elsewhere (literature check) | 8 |
| Match a neighbour's known signal | 18 |
| Unmatched and passing every LEO-Vetter test | 47 |
| … centroid more than 15″ from the target | 14 |
| … a TIC star within 15″ could produce the dip | 9 |
| … no FFI difference image | 2 |
| … none of these, MES ≥ 7.1, per-sector χ²/dof ≤ 5 | 21 (18 new, 2 ours, 1 on GitHub) |
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.
| TIC | Period (d) | Radius (R⊕) | Host (Teff, T mag, distance) | Notes |
|---|---|---|---|---|
| 44453747 | 1.106646 | 1.29 | 3088 K, 13.4, 36 pc | LHS 1083, an M4.5 dwarf 36 pc away |
| 239585318 | 1.983922 | 1.37 | 3121 K, 13.4, 37 pc | an equally bright star is 32" away |
| 1441057 | 2.093531 | 2.74 | 3602 K, 12.5, 109 pc | possible member of the young Octans association (~30 Myr); spot rotation 3.05 d |
| 377097159 | 2.604427 | 1.74 | 3315 K, 12.9, 50 pc | slow rotator (57-60 d) |
| 26986661 | 3.996223 | 2.03 | 3395 K, 13.3, 67 pc | LP 371-15 |
| 165502832 | 5.085657 | 2.24 | 3616 K, 13.5, 132 pc | – |
| 257433294 | 5.345440 | 1.75 | 3276 K, 12.8, 57 pc | flaring M dwarf |
| 211405734 | 7.297477 | 7.75 | 3488 K, 13.5, 127 pc | listed without a period in a giant-planet (GEMS) team's observing plan, so it may be their unpublished candidate |
| 357945537 | 8.887083 | 2.60 | 3451 K, 12.9, 77 pc | LP 335-28 |
| 233409999 | 9.578851 | 2.30 | 3475 K, 13.1, 90 pc | – |
| 142472039 | 10.086808 | 2.39 | 3848 K, 12.7, 114 pc | – |
| 394657305 | 13.941080 | 3.69 | 3606 K, 13.4, 113 pc | a brighter star (T = 11.0) 18.5" away could be the source (TIC contamination ratio 1.23) |
| Stage | Faint M dwarfs, 12.5 < T ≤ 13.5 |
|---|---|
| Stars searched | 84,863 |
| Signals found | 139,231 |
| Pass pre-filters | 8,521 |
| Match a known planet, TOI, CTOI or TCE | 625 (83 pass every LEO-Vetter test) |
| Reported elsewhere (literature check) | 27 |
| Match a neighbour's known signal | 130 |
| Unmatched and passing every LEO-Vetter test | 104 |
| … centroid more than 15″ from the target | 73 |
| … a TIC star within 15″ could produce the dip | 18 |
| … none of these, MES ≥ 7.1, per-sector χ²/dof ≤ 5 | 13 (12 candidates; TIC 9994636 rejected) |
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).
| Stage | All eight parts |
|---|---|
| Stars searched | 4,039,203 |
| Signals found | 7,459,491 |
| Pass pre-filters (SNR ≥ 9, ≥ 3 transits, duration and shape cuts) | 609,021 |
| Match a known planet, TOI, CTOI or TCE | 3,829 |
| Reported elsewhere (literature check, including ExoMiner++, RAVEN and T16) | 2,477 |
| Unmatched and passing every LEO-Vetter test | 13,986 |
| … and MES ≥ 7.1, per-sector χ²/dof ≤ 5 | 9,962 |
| … and no centroid offset or capable TIC neighbour within 15″ | 726 |
| … likely binaries: in a binary catalog at the target or a matching eclipsing binary nearby, or fitted radius > 20 R⊕ | 140 |
| … reported in a paper found by NASA ADS (NGTS-19b, a brown dwarf; one DTARPS candidate) | 2 |
| … tier C: grazing (b ≥ 0.9) or Gaia RUWE > 1.4 | 184 |
| … tier A: more than one signal on the star | 19 |
| … tier B | 381 |
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).
On 11 October the earlier candidates were also held to the full-sky TRICERATOPS bar (with Gaia limits and the
centroid Bayes factor); TIC 351339274 and 394657305 fail it on NFPP and move to tier C, leaving 24 in tier A
and 34 in tier B.
Reliability and completeness, measured on a 1% pilot (40,797 random stars; superseded by the 798,769-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 798,769 random full-sky stars, in
two independent samples of about 400,000, 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_798769_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:
| Clean | Neighbour within 15″ only | Centroid offset | |
|---|---|---|---|
| SNR ≥ 9, real | 726 | 1,294 | 7,918 |
| SNR ≥ 9, expected false alarms | 96 (19 found) | 202 (40) | 2,164 (428) |
| SNR 7.5–9, real | 382 | 448 | 2,652 |
| SNR 7.5–9, expected false alarms | 314 (62) | 273 (54) | 2,169 (429) |
False alarms concentrate at low multiple-event statistic (MES): 46% of the clean SNR ≥ 9 signals with
MES < 10 are expected to be false alarms, 16% of those with MES 10–12, and 2% above. 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 in the first sample 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 314 of them
are expected to be false alarms.
A false-alarm classifier (tier L1). A gradient-boosted classifier trained to tell the real search's
signals from the inverted ones, on search, LEO-Vetter and stellar features (fa_classifier.py,
validation/false_alarm_classifier.md), selects 69 of the 382
SNR 7.5–9 signals. Trained on the first inverted sample only, it lets through about 6 of the false
alarms expected among them, judged on the second sample, which it never saw: about 92% are
astrophysical, close to tier B. The 51 that also pass the binary, grazing and literature checks and
TRICERATOPS (FPP < 0.5, NFPP < 0.1) form tier L1: median radius 2.6 R⊕ (45 smaller than 4 R⊕), median
period 11 d. The remaining 56 with MES ≥ 9 are listed as tier L and not counted. The classifier's score
for every signal is in the catalog column fa_score; 36 tier A, B and N candidates score below 0.8,
where most false alarms lie, and are flagged (fa_flag) but not moved. Separately, 23 counted
candidates lie in the core of the Neptunian desert (P < 3.2 d, 4–10 R⊕; Castro-González et al. 2024),
where real planets are rare, and are flagged (desert_flag).
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.
Gaia limits on companions (11 October 2026). We reran TRICERATOPS on every counted candidate and on every signal that had failed only the TRICERATOPS bar: ten runs with the default priors and ten with limits from Gaia DR3, on the same light curves. MOLUSC (Wood et al. 2021) removes bound companions that Gaia would have resolved or that would raise the star's RUWE above its measured value. For 141 bright stars, the Gaia epoch-RV scatter rules out an eclipsing binary on the target. The median FPP of the counted candidates falls from 0.042 to 0.013, and 236 meet the validation thresholds apart from imaging (62 before). The bar now uses the constrained values, in both directions:
- 5 candidates leave, all within the run-to-run scatter of the threshold.
- 10 join: 5 from tier C and 5 neighbour-recovered signals, which were fitted for this update.
- TIC 94673748 and 258641782, whose Gaia RVs vary by 26 and 57 km/s, become likely eclipsing binaries.
Columns FPP_g, NFPP_g, FPP_gc and NFPP_gc hold the constrained values, and rv_flag the Gaia
RV result. The method and per-signal results are in
validation/gaia_companion_limits.md.
| Tier | Full-sky first release | Second release | With Gaia limits (11 October) |
|---|---|---|---|
| A (several candidates on the star) | 19 | 11 | 11 |
| B (single) | 381 | 252 | 254 |
| N (neighbour within 15″; TRICERATOPS and centroid) | – | 169 | 173 |
| L1 (SNR 7.5–9 selected by the false-alarm classifier; TRICERATOPS) | – | 51 | 50 |
| C (grazing, RUWE > 1.4, MES < 10, or TRICERATOPS FPP ≥ 0.5 or NFPP ≥ 0.1; not counted) | 184 | 456 | 455 |
| L (SNR 7.5–9, MES ≥ 9; not counted) | – | 56 | 56 |
| likely binary, likely nearby false positive, likely false alarm, reported elsewhere | 142 | 1,431 | 1,427 |
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, L1 and L have folders in
candidates/allsky/.
candidates/ephemerides.csv gives the ephemeris of each registered
candidate, with coordinates, the next transit, and the TESS sectors planned for the star. There are 553:
the 543 counted on 10 October and 10 that joined on 11 October (column registered). The files are
append-only. Candidates that leave the counted tiers keep their rows, with their current tier.
candidates/predictions.csv
lists all 36,423 predicted transits from the registration date through June 2027, with the 1σ uncertainty,
the UTC time at Earth's centre, and the TESS sector when the transit falls inside a published TESS
orbit window and the star is on a detector. The median 1σ uncertainty of the next transit is 12 min.
Of the 543 registered on 10 October, 62 have 656 transits in the published windows of Sectors 109–111 (through 3 December
2026), all with 1σ uncertainties shorter than the transit duration, and 343 candidates are on
detectors in sectors planned through Sector 134 (tess-point 0.9.5.post1). These files were
committed before those data exist: for a planet, each transit should appear at the predicted time
with the fitted depth.
The fitted reference epoch is usually the first transit in the data, where the epoch and period
errors are strongly correlated, so predict.py propagates the errors from the information-weighted
centre of the observed transits. For the fitted epoch itself, this reproduces the MCMC uncertainty
(median ratio 0.99; 5–95%: 0.82–1.29). The predictions assume a constant period, so transit-timing
variations would move transits away from these times. For TIC 343942488 (0.65 d) the catalog lists
the search period because the fitted period differs from it by 0.13%; its ephemeris uses the fit.
These files replace the first-release predictions of 7 October (commit 9d5f501), which stay in the
git history.
The approach follows Pavel Rabtsevich's search that found TIC 4206066 (write-up, data): detect a signal in recent TESS data and confirm it in archival sectors. SPOC's latest multi-sector search covers Sectors 1–96, and stars observed only in full-frame images never get a SPOC multi-sector search at all. A shallow transit that is 5–8σ per sector can stay below every single-sector threshold while reaching 10–20σ when all sectors are combined.
-
Targets (
fetch_targets.py,build_targets.py): TIC 8.2 dwarfs with T ≤ 11.5, 0.1 < R* ≤ 0.75 R☉, Teff ≤ 5300 K, observed by QLP in Sectors 97–105: 9,979 stars. TOI hosts (build_toi_hosts.py): every host of a TOI dispositioned CP, KP, PC or APC: 6,430 stars. -
Light curves (
index_products.py,download.py): one product per star and sector, preferring SPOC 2-min PDCSAP, then TESS-SPOC FFI PDCSAP, then QLP. 73,226 light curves for the K/M dwarfs and 54,189 for the TOI hosts. Each worker process fetches one star's sectors concurrently and stores them as one compressed array file. -
Search (
search.py,fastbls.py): per-sector biweight detrending (wotan, fitted to 10-minute means and interpolated; see Performance), Lomb–Scargle prewhitening of fast rotators (P_rot < 3 d), and box least squares over 0.5–30 d on a period grid set by each star's density. Three variants:- coherent: astropy BLS on the full-baseline grid (~17 s per star);
- semi-coherent: per-season BLS statistics summed, then the top peaks refined coherently (~4 s per star), which misses ~15% of weak signals (see Validation);
- stack-slide (
--method stack, added in the update): each season is folded once per coarse frequency; for each fine frequency the folded seasons are shifted by the phase they accumulate and summed, which is phase-coherent across seasons (~9 s per star).
The K/M dwarfs were searched with the coherent and semi-coherent variants, the TOI hosts with stack-slide. Up to 3 signals per star (4 for TOI hosts) by iterative masking.
-
Masking known planets (
search.py --mask-tois): for the TOI hosts, every TOI is masked before the search, using its catalogue ephemeris widened by 3σ of its propagated timing error and our own refit where the signal is detected. Vetting removes the same transits. -
Known-signal matching (
known.py,fetch_known.py): ExoFOP TOIs and CTOIs, NASA Exoplanet Archive planets, all public SPOC TCE tables, including period harmonics and TIC neighbours within 2.5′. -
Literature check (
literature.py, added in the update): the RAVEN and T16 candidate tables; the SPOC TCEs that ExoMiner++ classifies as planet candidates (Valizadegan et al. 2025, 2026); the LEO-Vetter M-dwarf candidates (Kunimoto et al. 2025, VizieR J/AJ/170/280); the Eschen et al. (2024) M-dwarf candidates; and TIC IDs, TOI designations and nearby periods in the tables and text of 2,037 arXiv astro-ph.EP papers whose abstracts mention TESS, TOIs or TICs (read from arXiv's HTML versions, plus the e-print for papers whose candidate tables are machine-readable only). A matching signal is taggedreported. Run on the first release, it flags exactly the two signals that our manual check had found already published. -
Vetting (
vet.py): transit-masked re-detrending, the 13 false-alarm and 4 false-positive tests of LEO-Vetter (Kunimoto et al. 2025), per-sector depth consistency, and old-vs-new sector detection. LEO-Vetter's per-cadence loop is replaced by a compiled equivalent (same results, 2–5× faster vetting). The diagnostic plot is drawn only for signals with at most one LEO-Vetter failure (--plot-max-fails). -
Pixel and field checks (
followup.py, automated in the update): TIC stars within 2.5′ that could produce the depth as eclipsing binaries (neighbors.py), and FFI difference-image PRF centroids with transit-diffImage in the two sectors with the strongest transits and smallest cutouts, with LEO-Vetter's 15″ limit. For the first release these were run by hand (pixel_vet.py,tpf_pixels.py). -
Characterisation: batman + emcee transit fits with a TIC stellar-density prior (
fit_transit.py) and TRICERATOPS FPP (export_fold.py,fpp.py, now vectorised, with the Gaia query cached and 10 runs averaged). -
Novelty: for the first release, ExoFOP pages, TOI/CTOI lists, NASA Exoplanet Archive, SPOC TCEs, the T16 catalogue (Roth et al. 2026), the RAVEN tables, the Tschudi 2026 survey source, and web searches on every TIC ID, run on 2026-10-07 (see
results_public/novelty_check.md); for the update, steps 5 and 6 plus web searches on every candidate TIC ID and host TOI.
| Stage | K/M dwarfs | TOI hosts |
|---|---|---|
| Stars searched | 9,972 | 6,430 |
| Signals found | 41,758 | 15,222 |
| Pass pre-filters (SNR ≥ 9, ≥ 3 transits, duration and shape cuts; TOI hosts: FGKM dwarfs only) | 1,549 | 660 |
| Match a known planet, TOI, CTOI or TCE | 236 (of which 109 pass every LEO-Vetter test) | 229 (of which 94 pass every LEO-Vetter test) |
| Match a published list outside ExoFOP (literature check) | 14 | 10 |
| Match a neighbour's known signal | 23 | 1 |
| Unmatched | 1,276 | 420 |
| Unmatched and passing every LEO-Vetter test | 38 | 40 |
| … centroid more than 15″ from the target | 11 | 11 |
| … a TIC star within 15″ could produce the dip | 14 | 11 |
| … neither | 13 | 18 |
| … and MES ≥ 7.1, per-sector χ²/dof ≤ 5 | 13 | 16 |
The last row includes six first-release candidates (K/M dwarfs) and TOI-4342's 18.9 d signal (TOI hosts); TIC 117799904 is counted in both columns; TIC 143168991's 26.56 d signal (above) was rejected by hand.
-
Blind recovery of TIC 4206066. Both of Rabtsevich's signals were found without prior knowledge: 3.18278 d (his 3.182785 d) and 11.13274 d (his 11.13274 d), with Rp = 1.41 R⊕ (his 1.4) and the nearest depth-capable neighbour at 44.6″ (his ≳ 43″). See
figures/validation_tic4206066.png. -
Known planets. TOI-700 b and c, and L 98-59 b, c and d are recovered (
validation/). With the TOIs masked, the TOI-host search recovers TOI-4342's 18.9 d signal and the community candidate L 98-59 .04 (CTOI, 1.05 d). -
TOIs in our light curves. Before masking a host's TOIs, the TOI-host search refits each one at its catalogue period: 6,258 of the 6,321 TOIs with periods up to 30 d and at least two cycles in the data (99.0%; 99.7% of the confirmed planets) are detected at SNR ≥ 7. That includes 84 of the 87 TOIs on the hosts of the 70 TESS planets the NASA Exoplanet Archive added on 8 October 2026; the other three have periods of 754 and 1,637 d or a single transit. This is a check at known periods; blind recovery is measured by the control systems above and the injections below.
-
Injection–recovery. Transits injected at a nominal SNR of 10 into 300 random targets (
validation/injection_recovery_snr10_n300.csv):Search Recovered at the injected period Detected with SNR ≥ 9 Median time per star coherent 60% 92 (31%) 17.5 s stack-slide 62% 88 (29%) 8.9 s semi-coherent 52% 78 (26%) 4.2 s Of the SNR ≥ 9 detections, 6 were found only by the coherent search and 2 only by stack-slide; 18 were found only by the coherent search and 4 only by the semi-coherent one.
-
False alarms. Searching the inverted light curves of 1,500 random K/M dwarfs (
EXO_INVERT=1), where any dip is a false alarm: 157 signals pass the pre-filters, 1 passes all LEO-Vetter tests, and none also has consistent per-sector depths. The real light curves of the same stars give 6 new signals passing all LEO-Vetter tests, 4 of them with consistent depths. Inversion measures false alarms from noise and systematics, not astrophysical false positives such as background eclipsing binaries, which steps 8 and 9 address. -
Automated vs manual checks. On the first-release signals,
followup.pypasses 7 of the 12 cleanly, flags the close neighbours we had noted by hand for 4 (TIC 166321785, 147226218, 451641911, 30470520), and puts TOI-6284's 5.24 d centroid at 15.2″, just over the limit (our 2-min pixel analysis gave 2.8″). It rejects TIC 264151965, which we had set aside, with a 106″ offset. -
Compiled LEO-Vetter loop. On 6 test signals the compiled version gives the same pass/fail results and metrics equal to ~10⁻⁹ (
EXO_SLOW_LEO=1runs the original).
Measured on 200 random K/M dwarfs (median 5 sectors), with the stack-slide search:
| Change | Speed-up | Effect on results |
|---|---|---|
| Download in worker processes instead of threads (parsing was serialized by Python's GIL) | 3× on 16 vCPUs, more on larger machines | identical files |
Stack-slide kernel: no division per trial box, branch-free shift-and-add, floor instead of % |
1.42× | same first signal and SNR on all 200 stars |
Biweight trend fitted to 10-minute means (EXO_TREND_BIN=0 restores full cadence) |
1.16× (14× for the trend itself) | residuals differ by ~30 ppm on quiet stars, within the trend's own noise; 100 injections at SNR 10: 30 detections against 28 |
| AWS Graviton4 (c8g) instead of Intel (c6i), same number of vCPUs | 2.6× | same first signal and SNR on all 200 stars; identical vetting results |
Together a star costs about 5× less to search than before (200 stars in 70 s on a 16-core
c8g.4xlarge). On Graviton, batman-package is built from source (requirements.txt). Splitting
long continuous stretches into shorter blocks for stack-slide (EXO_SEASON_MAX) gained only 10%
on this sample and changed some weak peaks, so it is off by default. Before these changes the
stack-slide kernel took 81% of the search time and detrending 14%.
uv venv --python 3.12 .venv && uv pip install --python .venv/bin/python -r requirements.txt
# transit-diffImage is installed from GitHub (see requirements.txt). Its PRF download URL uses
# http://archive.stsci.edu, which now returns 404; change it to https in tessprfmodel.py.
uv venv --python 3.11 .venv-trice && uv pip install --python .venv-trice/bin/python -r requirements-triceratops.txt
cd pipeline
../.venv/bin/python fetch_targets.py # TIC criteria query (~10 min)
../.venv/bin/python build_targets.py # QLP S97-105 lists + Gaia GCNS columns
../.venv/bin/python fetch_known.py # TOI/CTOI/planets/TCE catalogs
../.venv/bin/python literature.py # RAVEN, T16, arXiv mentions; ~2 h the first time (arXiv allows
# one request per 3 s), then only new papers are fetched
../.venv/bin/python index_products.py # MAST product list
../.venv/bin/python download.py # ~130 GB transferred, ~12 GB stored
../.venv/bin/python search.py targets.txt --method semi --out ../results/search_semi
../.venv/bin/python search.py targets.txt --method coherent --out ../results/search
./triage.sh # collect -> vet -> shortlist
../.venv/bin/python followup.py # neighbours + difference-image centroids
# TOI hosts
../.venv/bin/python build_toi_hosts.py
../.venv/bin/python index_products.py --targets ../data/catalogs/toi_hosts.parquet --out ../data/catalogs/manifest_toi.parquet
../.venv/bin/python download.py ../data/catalogs/toi_hosts.txt --manifest ../data/catalogs/manifest_toi.parquet
../.venv/bin/python search.py ../data/catalogs/toi_hosts.txt --stars ../data/catalogs/toi_hosts.parquet \
--out ../results/search_toi --method stack --mask-tois --max-signals 4
../.venv/bin/python collect.py --search ../results/search_toi --stars ../data/catalogs/toi_hosts.parquet \
--out ../results/cands_toi.csv --signals ../results/signals_toi.parquet --dwarfs
../.venv/bin/python vet.py ../results/cands_toi.csv --out ../results/vet_toi --stars ../data/catalogs/toi_hosts.parquet
../.venv/bin/python shortlist.py --cands ../results/cands_toi.csv --vet ../results/vet_toi --out ../results/vetted_toi.csv
../.venv/bin/python followup.py --vetted ../results/vetted_toi.csv --vet-dir ../results/vet_toi \
--stars ../data/catalogs/toi_hosts.parquet --out ../results/followup_toi
# validation
../.venv/bin/python inject_test.py 300 --methods coherent,semi,stack
EXO_INVERT=1 ../.venv/bin/python search.py sample.txt --method stack --out ../results/search_inv
# update 2: unpromoted SPOC TCEs, and periods of 20-100 d
../.venv/bin/python build_tce_targets.py # SPOC TCEs on dwarfs that never became TOIs or CTOIs
./triage_tce.sh # re-measure them with later sectors, vet, follow up
../.venv/bin/python search.py stars.txt --method stack --pmin 20 --pmax 100 \
--premask ../results/search --out ../results/search_long # 20-100 d, earlier signals masked
# updates 3 and 4: faint M dwarfs
../.venv/bin/python fetch_targets.py --tmin 11.5 --tmax 13.5 --rmax 0.6 --step 0.02 \
--out ../data/catalogs/tic_faint_mdwarfs.parquet && ../.venv/bin/python build_faint_targets.py
../.venv/bin/python ads_check.py candidates.csv ads.csv # NASA ADS + Zenodo; ADS token
EXO_DATA_SOURCE=s3 ../.venv/bin/python download.py ... # read from the AWS mirror
# full-sky second release (inside AWS): TRICERATOPS for signals with a neighbour inside 15", fits,
# the SNR 7.5-9 band from the signals tables, and inverted light curves for false alarms
EXO_DATA_SOURCE=s3 ../.venv/bin/python fpp_batch.py list.csv --stars stars.parquet --out ../results/fpp_recover
../.venv/bin/python fit_batch.py list.csv --stars stars.parquet
../.venv/bin/python collect.py --from-signals ../results/signals_allsky.parquet --snr 7.5 --max-snr 9 \
--stars shard.parquet --out ../results/cands_low.csv
EXO_INVERT=1 ../.venv/bin/python run_shard.py sample.parquet manifest.parquet --shard 0 --nshards 1 \
--out ../results/search_inv_allsky --keep-snr 7.5
../.venv/bin/python fa_classifier.py ../validation/false_alarm_classifier_table.csv.gz \
--stars-real 4039203 --stars-inv 399408 399361 # false-alarm classifier (tier L1)
# Gaia limits on companions (11 October): MOLUSC per star, then default and constrained TRICERATOPS runs
../.venv/bin/python molusc_batch.py stars.parquet --gaia gaia_dr3.csv --molusc MOLUSC/code --out ../results/molusc
EXO_DATA_SOURCE=s3 ../.venv/bin/python fpp_batch.py list.csv --stars stars.parquet --out ../results/fpp_m \
--molusc ../results/moluscEXO_DATA_SOURCE=s3 reads light curves from the STScI open-data bucket and falls back to MAST.
Inside AWS us-east-1 this is about 20 times faster than downloading from MAST.
targets.txt is one TIC ID per line. Per-candidate follow-up: fit_transit.py, export_fold.py,
fpp.py (in .venv-trice), molusc_batch.py, candidate_table.py, predict.py, make_figures.py,
package_candidates.py. Light curves and catalogs are downloaded into data/ and are not part
of the repository.
pipeline/ code
candidates/ one folder per candidate, ephemerides.csv and predictions.csv
figures/ figures used here
results_public/ all signals, vetted signals, candidate tables, novelty check
validation/ control-star vetting outputs and the injection-recovery tables
This work uses data from the TESS mission, obtained from MAST at STScI: SPOC light curves (Jenkins et al. 2016), TESS-SPOC FFI light curves (Caldwell et al. 2020), and QLP light curves (Huang et al. 2020). It uses data from the ESA mission Gaia, processed by the Gaia DPAC, and the Gaia Catalogue of Nearby Stars (Gaia Collaboration, Smart et al. 2021). This research has made use of the Exoplanet Follow-up Observation Program (ExoFOP; DOI: 10.26134/ExoFOP5) website, which is operated by the California Institute of Technology, under contract with the National Aeronautics and Space Administration under the Exoplanet Exploration Program, and of the NASA Exoplanet Archive, and of the VizieR catalogue access tool and the CDS XMatch service (CDS, Strasbourg). The literature check uses the RAVEN (Lafarga et al. 2026) and T16 (Roth et al. 2026) catalogues and arXiv. Software: astropy, numpy, scipy, numba, lightkurve, wotan, batman, emcee, LEO-Vetter, transit-diffImage, tess-point, astrocut, TRICERATOPS and MOLUSC.
Method inspired by Pavel Rabtsevich's TIC 4206066 study. Built with Claude Code (Anthropic).
Code: MIT (see LICENSE). Results, tables and figures: CC BY 4.0.