Sector 107: 53 vetted stars give unconfirmed candidates with no usable planet probabilities, and known-planet recall is only 22%

Summary

Starting from iteration 2, 650 signals were merged: the 66 passed, the 146 not fully vetted, and 438 TLS rows. This is more than the 255 originally expected. Cross-matching against the NASA Exoplanet Archive, ExoFOP TOI and CTOI lists left 585 new signals, 52 known signals and 13 same-star signals at a different period. After final known-object removal, 53 distinct stars were ranked, and 4 community TOIs were removed. The scores are rank scores, not planet probabilities, and TRICERATOPS was not run. The known-answer check kept only 22% of known planets at a score above 0.5, so absolute scores should not be trusted. Only the order carries information, and every object is an unconfirmed candidate needing follow-up.

What was analyzed

• Inputs: cached Sector 107 SPOC 2-min PDCSAP curves, the NASA Exoplanet Archive (6144 rows), the ExoFOP TOI list (8151 rows) and the CTOI list (5139 rows). Other TESS sectors, Gaia DR3 and target pixel files were also used.
• Merge and known-object removal (group 2): 650 signals on 381 unique TICs. Of these, 52 are known, 13 are the same TIC at another period, and 585 are new. 2 TICs failed.
• Photometric vetting, slices A and B (groups 3 and 4):
  • Slice A covered 393 stars and 644 signals.
  • Slice B covered 299 candidate signals, plus 60 known planets, 13 ExoFOP false positives and 60 injections as controls.
  • Tests were odd/even depth, secondary eclipse, V-shape, single-event dominance, transit count, SNR and planet radius.
  • A trapezoid model replaced batman, so depths and radii are approximate.
• Gaia contamination (group 5): 598 signals on 381 stars, with neighbours counted within 42″, RUWE and the non-single-star flag.
• Cross-sector recovery (group 7): 58 survivors plus controls.
• Sector-wide systematics veto (group 8): 731 rows, built from 18411 events.
• Gaia stellar-density and duration features (group 9): 1076 signals on 707 stars.
• Difference-image centroid check (group 10): top 20 plus 16 known planets.
• Alias, variability and inverted-curve test (group 11): 174 rows.
• Known-answer check (group 1): 429 stars (112 known planets, 17 ExoFOP false positives, 150 controls, 150 injections).
• Ranking (group 6): produced the top-20 table.
• Total effort: 668 runs (652 succeeded), 21.9 CPU hours, 19 work groups done, 120 of 240 minutes used.

Key results (all exploratory)

• Known-answer check (group 1):
  • Only 22% of known planets are kept at a score above 0.5.
  • A refit of the cuts reached about 70% recall, but class separation is weak.
  • The score's validation is circular. At a score above 0.5, the out-of-fold false-alarm rates are 0.33 on controls and 0.71 on ExoFOP false positives.
• Top-20 table: none of the top 20 is a known planet or TOI. The 53-star ranked list and the top-20 table (top20_final, final_top20_flags) are saved. The full rows with TIC, period, depth, transit count, Rp and score were not available to me, so I cannot reproduce them here.
• Per-transit consistency: only 5 of the top 20 pass this test.
• Difference-image centroid:
  • Among the top 20, 8 were flagged against the target and 11 against the out-of-transit centroid, 5 were noisy and 5 were clean.
  • Among 16 known planets checked, 8 were flagged against the target, 9 against the out-of-transit centroid, 8 were noisy and 3 were clean.
  • The test is therefore weakly discriminating.
• Cross-sector recovery: 29 of 58 survivors are recovered in at least one other sector. The feature separates survivors from known planets but not real signals from false positives.
• Gaia flags (381 stars): 259 have a neighbour that could cause the depth, 146 have high RUWE, 64 have the non-single-star flag, and 121 have no flags.
• Common-epoch systematics: no sector-wide artefact was found (0 hot bins).
• Alias, variability and inverted-curve test: of 55 survivors, 22 have 0 flags and 32 have at most one flag.
• Other leads:
  • 76 single-dip leads, of which 59 remain after known-object removal. None of these passed the strict hard-flag criterion.
  • 64 additional-signal leads on 44 known-planet or TOI hosts. Recall in that search was 0.58 for injections and 0.16 for known second planets.

What did not work

• The score does not reproduce known planets: 22% kept at a score above 0.5.
• The cross-sector and centroid tests do not separate planets from false positives.
• No true planet probability was obtained, so TRICERATOPS-style statistical validation is absent.
• Additional-signal recall on known second planets was poor (3 of 19). Recall for injections falls to 0.26 at depths of 0.05–0.15%.

Confidence and limitations

• Confidence is low (about 0.5 overall, per the team's assessment).
• No preregistered tests were run, and the one claim has level "candidate" and novelty "similar", with no reviews.
• Radii and depths come from a trapezoid approximation. Epochs come from a box refit.
• Photometric and Gaia vetting cannot replace spectroscopy, high-resolution imaging or ground-based photometry. All objects need these before any planet claim.

Next steps

• Run TRICERATOPS, or an equivalent, on the top candidates to obtain real false-positive probabilities.
• Rerun the vetting with a proper transit model, and tune the search so known-planet recall reaches 70% or more.
• Obtain ground-based photometry and reconnaissance spectroscopy for candidates with at least 3 transits, clean centroids and no Gaia neighbour that could cause the depth.
• Search for ultra-short-period planets and M dwarfs, which this iteration did not cover.