TIC 147685784: an unconfirmed transit hint at P = 0.9796 d on a K dwarf, from a short-period search of public QLP light curves, with hash-registered predictions for TESS Sectors 120 and later

We report an unconfirmed transit-like signal on the K dwarf TIC 147685784 (Tmag 11.39, Teff about 5000 K, 151 pc, no sign of binarity in Gaia DR3: RUWE 0.86). It was found in a short-period search (P = 0.25–4 d) of the public QLP light curves of bright K and M dwarfs (sectors up to 96; programme described in the first record of this series). The dip has a depth of about 410 ppm (about 1.6 Earth radii if it is a planet on this star) and a duration of about 1.5 h at P = 0.9796 d, with a positive depth in each of the nine search sectors (198 transits): 419 ± 33 ppm in QLP SAP flux, 409 ± 27 ppm (SAP) and 411 ± 26 ppm (PDCSAP) in TESS-SPOC FFI light curves, and 352 ± 37 ppm in separately reduced TESScut pixel photometry (same pixels, same analysis chain; not independent measurements). The search peak (SNR 8.35) is one of 37 peaks in 31 stars that pass the six filters of the original rule in the 4500 stars searched so far (57 peaks in 47 stars under a later, post-hoc filter rule; status 29 October 2026, Chunk 6 not included); it ranks 28th of the 37 by search SNR, and the star was singled out by a pixel-level stage that 4 of the 27 stars analysed there pass formally. Block-shuffled copies of the star's own light curve reach SNR 8.35 in 0 of 1000 runs (strongest null peak SNR 6.43), but that null describes only white and short-timescale noise, not systematics, red noise or stellar variability, and it is not the false-alarm probability of the signal: in the real search 295 of the 4500 stars (6.6 %) have a peak of at least SNR 8.35, and among 4500 stars about four to five are expected to fall below a per-star false-alarm probability of 0.001 by chance alone. We see no second dip in any phase (phase 0.5: 16 ± 34 ppm), LEO-Vetter 1.2.0 raises no flag in QLP (MES 11.3) but the flag "chi" (inconsistent single-transit SNRs) in the pixel photometry (CHI 7.37 against a threshold of 7.8; QLP 7.82, and leaving out any one of the nine sectors raises the flag for QLP in 6 of 9 cases), the difference image places the dimming within 1.4 sigma of the target, and TRICERATOPS gives FPP 0.07–0.09 with NFPP 0.000 (not a validation: no contrast curve, no imaging). Open problems stated plainly: odd and even transits differ by 2.5–3.2 sigma in TESS-SPOC light curves with the pipelines' own noise estimates (odd deeper in 7 of 9 sectors; 2.0 sigma against the random-period null distribution of the registered rule; the maximum over four pipelines, so a selection-adjusted value is nearer 2.8 sigma) but by about 1 sigma in QLP and pixel photometry, so an eclipsing binary at twice the period with unequal minima is not excluded; the photometry prefers transit times in Sectors 120/121 about 25–29 min earlier than the frozen search ephemeris; the data prefer a transit duration of 2.0 h to the 1.5 h grid value of the search, which would put the transit density 2.4 sigma below the density of the star; the peak was selected among many; the filter rule v2 and the pixel criterion (b) v2 were introduced after seeing data on other stars (this star passes without them). What is registered: an ephemeris (E2), the predicted transit times for Sectors 120, 121 and 127 (windows from Sector 127 on extrapolated, ±10 d), the statistic, a decision rule with the outcomes SUCCESS, FAILURE, UNDECIDED and NOT TESTABLE, the evaluation script with a self-test, and the data products, all fixed and hashed locally (SHA-256 of the rule 2525dbb0…5d4d, 29 October 2026, 05:55:28 local time, no external timestamp) before any data of this star from Sectors ≥ 97 were loaded (the star has no observations in S97–S119); the deposition of this record, its first external time stamp, precedes the observation of Sector 120. We claim no priority: ExoFOP lists no TOI and no CTOI for this star on 29 October 2026 (queried again at 21:20 CET, immediately before the deposition), and the star is bright enough for QLP-based searches by the TESS teams. The analyses were executed by AI instances (Claude, Anthropic) on public data under the author's standing instructions; the author has reviewed a summary of the results but has not independently re-run the analyses, so all conclusions are tentative.