Five periodic variables absent from VSX, found in TESS light curves of 517 high-latitude stars

Archive behind the note "Periods for Five Variable Stars Missing from VSX" (Antoine Olivier, in preparation for the Research Notes of the AAS). A search of TESS-SPOC light curves of 517 high-latitude stars found five periodic variables absent from the AAVSO International Variable Star Index (VSX). Three of them have no period or type in SIMBAD or VizieR either, apart from an automated fit at twice the period for TIC 243167066 (Green et al. 2023, not selected as a candidate). The other two were reported before: Thiemann et al. (2021) list TIC 436556517 at 0.44 d, and Green et al. (2023) give automated fits at 2.2007 d for it (score 0) and at 0.5089 d for TIC 402964191, which they scored 0.85 but did not select (amplitude below their cut). Maíz Apellániz et al. (2023) flag all five as variable from Gaia scatter alone. Results Star Type Adopted period TIC 105616985 BY 0.6834 ± 0.0004 d TIC 243167163 BY 0.4930 ± 0.0004 d TIC 243167066 ROT 2.30 ± 0.09 d TIC 436556517 EA 2.2056294 ± 0.0000024 d (eclipse ephemeris) TIC 402964191 VAR 0.25473 ± 0.00005 d, or 0.50945 d if the variation is ellipsoidal Files Scripts, catalogue extracts, intermediate tables and figures behind the note, including the two validation samples and the held-out test. Preregistration excerpt and held-out decisions table, with RFC 3161 and OpenTimestamps stamps of their content (file dates 19 October 2026, after the analysis). ZTF light curves used (IRSA) in data/ztf/. TESS light curves are not included. They are public at MAST: TESS-SPOC (doi:10.17909/t9-wpz1-8s54) and QLP (doi:10.17909/t9-r086-e880). Outside the zip, listed in README.txt: the files of the independent check described below. Reproduction reproduce.py downloads the 18 TESS-SPOC light curves it needs (SHA-256 checked), re-runs the main analyses and compares each output with the archived copy. 15/15 checks passed on 19 October 2026 (125-129 s, Windows 11, Python 3.13.15). Same result on 29 October 2026, on a fresh extraction of the deposited zip (Python 3.13.17). Not covered: the search over 517 stars, the validation samples, the catalogue queries and the ZTF/QLP code. See README.md, section "How to reproduce". Independent period search (29 October 2026) DeepSeek V4.1 Flash received the five TIC numbers, the TESS sectors, the ZTF file names, the file formats and one sample row of each file. It did not receive any period, type or text of the note. It wrote its own generalised Lomb-Scargle search (numpy). This is an independent reimplementation of the period search only. Rule written before the brief was sent: agreement within 1% of the note's period; P/2, 2P and daily aliases counted separately. verify_periods_deepseek_original.py is its script, unchanged. It fails at its first numpy.linalg.solve call under NumPy 2.5.3. verify_periods_deepseek.py differs by one line, changed by the author's AI assistant (Claude): the array shape passed to that call. Run on the 18 TESS-SPOC and 3 ZTF light curves (verify_periods_output.txt): Star Independent period Verdict TIC 105616985 0.6834362 d agrees TIC 243167163 0.4918573 d agrees (within 0.25%) TIC 402964191 0.2547333 d agrees; like the note, it does not decide between P and 2P TIC 243167066 2.3325 d 1.4% away: outside the 1% rule, inside the note's uncertainty TIC 436556517 1.102792 d half the note's period; counted as P/2 TIC 243167066: its five per-sector peaks are 2.199, 2.295, 2.270, 2.429 and 2.322 d. It reports a second peak at 2.402 d. TIC 436556517: the script labels 1.102792 d fundamental. Its test compares one-term (sinusoidal) powers at P, 2P and P/2, so it cannot see unequal minima. A complementary script by Claude, not part of the independent check (oddeven_436556517.py), folds that star at twice the script's per-sector periods (2.2096, 2.2089 and 2.2111 d in sectors 17, 57 and 84). The alternate minima are 78.2 and 12.3, 72.4 and 14.5, 75.7 and 7.6 ppt deep. The period is twice the script's value, consistent with the note's eclipse period. In ZTF, only the r band of TIC 105616985 peaks at the period (0.6834087 d). The other first peaks lie near 1 d (TIC 243167163 and 243167066, both bands) or at 0.2885 d (TIC 105616985, g band). The check does not test the note's ±0.0000024 d precision, which needs cycle counting across sectors. Control copies The script's false-alarm estimate and peak recovery were each checked once, on TIC 105616985 sector 76 (6551 points), with control copies written by Claude (make_controls_periods.py). Flux permuted (seed 20261029): the first peak (0.0748 d) is not significant (false-alarm probability 0.30). Permuted flux multiplied by 1 + 0.005 sin(2πt / 1.2345 d): the first peak is at 1.239 d (false-alarm probability 0). These are single controls, not a calibration of the false-alarm probability. Limits The preregistration has no independent timestamp (git commit clock only). The box-least-squares thresholds were set in uncommitted notes. The independent check covers the period search only. Use of AI The code, the catalogue and literature searches, a separate reanalysis, critical reviews and a first draft of the note were produced with AI assistants (Anthropic Claude; OpenAI Codex for one review) under the author's direction. The independent period search was written by DeepSeek V4.1 Flash. The author takes responsibility for the content. No AI system is an author.