TIC 38699429: an unconfirmed transit candidate around a faint nearby M dwarf, from a hash-registered search of public TESS-SPOC light curves

Version 4 (1 November 2026): corrects the pixel position test (zero-valued pixel columns in the Sector 3 cut-out had inflated its calibration: on-target p = 0.027 instead of 0.087, the failure criterion is still not met); results that do not use these cut-outs are unchanged (Version 3: 10.5281/zenodo.23661624).

We report an unconfirmed transit candidate on TIC 38699429 (Gaia DR3 4676612253064169344; Tmag 15.27, Teff about 2914 K, d about 57 pc), a faint nearby M dwarf with 31 TESS-SPOC full-frame light curves (Sectors 1-69) and no 2-min data. The signal has P = 11.718179 d, a depth of 24621 +/- 893 ppm in the frozen template (about 2.5 %) and a measured duration of about 1.4 h. It was the only survivor of a hash-registered coherent multi-sector search of 400 M dwarfs without 2-min data; the threshold comes from 110 control stars and 331 scrambled runs (at most 0.05 expected false positives for the noise in the scrambled data). The ephemeris was written to file before Sectors >= 86 were looked at (its hash was recorded in a hash file only afterwards), and the signal is recovered there (z = 10.8, 17 events). A reproduction (blind only for the periodogram) and a counter-evidence analysis, both by AI instances of the same model family using the same photons, find the signal real, strictly periodic and located on or near the target within a few arcseconds. The candidate is not validated: TRICERATOPS gives FPP 0.19 and NFPP 0.16, the pixel position test gives p = 0.027 for the on-target hypothesis (corrected in Version 4; the failure criterion p < 0.01 is not met) and shows the best position 5.7 arcsec from the Gaia position, and a bound faint companion, an unresolved background eclipsing binary and a twin eclipsing binary at twice the period are not excluded. The radius of about 2.8 Earth radii holds only if the signal is on a single star. The signal is also not new to the SPOC pipeline: a multi-sector SPOC data-validation report with a consistent period has been public since 14 September 2024, although the star is not a TOI or CTOI as of 25 October 2026; no claim of first detection is made. The analysis was AI-assisted (Claude, Anthropic); the author checked all figures, data sources and analyses, ran the (AI-written) scripts in his own Python environment and verified the results (statement of the author, 25 October 2026); the reviewers are AI instances, not independent human replications.

What changed in this version (Version 4, 1 November 2026):

Correction of the pixel position test. The TESScut cut-outs of Sectors 3 and 87 contain two pixel columns with values near zero instead of sky flux; in Sector 3 they lie in the background ring of the pixel position test (gate G7, Sect. 7) and inflated its injection calibration (95th percentile of the calibration Δχ²: 5501 instead of about 1). Recomputed with the same scripts and random numbers (300 injections), with these columns excluded from the ring and the weights, or with Sector 3 omitted, the p value for the on-target hypothesis is 0.027 (8 of 300) instead of 0.087, and the Δχ² of the target position against the best cell is 35.1 or 28.6 instead of 34.5. The statements "p = 0.087" and "moderate power" of Version 3 are withdrawn. The best cell remains (-4″, +4″), 5.66″ from the Gaia position, and all seven Gaia neighbours within 60″ remain excluded (smallest Δχ² 470 or 444 instead of 463). The pre-set failure criterion (p < 0.01) is still not met; the corrected calibration speaks somewhat more clearly against the on-target hypothesis, and the offset of the dimming from the Gaia position remains unexplained. In the held-out pixel photometry of the reproduction, masking the two columns improves the failed target-position fit of Sector 87 (χ²/dof 18.4 to 3.7) but gives a target flux of 297 e/s instead of 17 e/s, still outside the pre-set plausibility range of 80-220 e/s, so Sector 87 remains excluded and the seven-sector result (3.39 ± 0.18 e/s, z = 15.4) is unchanged; the eight-sector value (3.06 ± 0.17 e/s), which includes Sector 87, was not recomputed. The SPOC light-curve detection, the held-out TARS test (z = 10.79), TRICERATOPS and all light-curve results do not use these cut-outs and are unchanged; the signal remains an unconfirmed candidate. The PDF is corrected accordingly (abstract, Table 3, Sects. 6 and 7, Limitations, new Sect. 12 "Corrections in Version 4"). The recomputation was run by an AI instance and has not yet been re-run by the author; its scripts are not part of the supplementary archive, which is unchanged from Version 3.