# Science write-up: validation pass, 11 October 2026

This folder adds the next round of checks on the two Planethunt leads, run end to end on the
original data in the `evidence-20261010` release archives (Kepler and TESS archive SHA-256
hashes match `analysis/full-archive-review.md`). **Neither lead is validated by this pass.**
One lead gets materially stronger, and the other turns out to be weaker than it looked.

Short version:

| | KIC 5689121 / TIC 120686465 | TIC 309780779 |
|---|---|---|
| Signal real and periodic | Yes | Yes, recovered blind in each half of the data |
| Signal on the target star | Strong: three independent pixel/photometry tests | Not re-tested (MAST unreachable); prior Gaussian-PSF result only |
| Known-object contamination | Kepler-169 excluded | No catalog match within 5′ or 1° harmonic |
| Formal FPP (TRICERATOPS) | **≈ 0.07** (threshold 0.015) | **≈ 0.57** |
| Dominant false-positive scenario | Planet around an unresolved bound companion (STP) | Eclipsing binary bound to the target (SEB) |
| With hypothetical imaging | ≈ 0.03 | ≈ 0.45 |
| What would move it | AO imaging + RUWE-based companion limits | Reconnaissance spectroscopy / RVs; colour-dependent depth |

## KIC 5689121 / TIC 120686465 (Kepler, P = 13.769 d)

**Ephemeris (full baseline, all 13 quarters, PSF flux).** P = 13.769448 ± 0.000065 d,
T0 = −1369.7546 ± 0.0020 BTJD, ρ* = 4.4 (+2.3 −2.8) ρ☉ from the free-density fit, consistent
with the Gaia/TIC density of the 0.51–0.55 R☉ M dwarf. `scripts/kepler/kfit.py`,
`results/kepler/kfit_psf.json`.

**Kepler-169 is not the source.** No candidate transit overlaps a Kepler-169 e transit in the
whole baseline; the closest pair is 67 h apart, and the offset drifts from 2.80 to 3.04 d. The
period differs from Kepler-169 e (13.767101 d) by 32σ. Box depths in the target's PSF light
curve at all five Kepler-169 ephemerides are null (−65 to +143 ppm, all |d| < 1.1σ). Chance
overlaps with Kepler-169 b and d (5 and 1 events) are at the expected rate for short periods.
`k169_overlap.py`, `kdepth.py`.

**The PSF/SAP depth gap is dilution, and its pattern points to the target.** Per quarter, the
SAP depth tracks the target's share of the SAP-aperture flux (FLFRCSAP × PSF flux / SAP flux):
the dilution model fits (χ² 15.4/12, p = 0.22, scale 0.91); a constant depth does not
(p = 0.005); a signal riding on the extra aperture flux is rejected (p ≈ 1e−36). The seasons
with 7-pixel apertures (Q2, Q6, Q14) have the most contaminating flux and the shallowest SAP
depths, the opposite of what a contaminant-hosted signal would do. `kcrowd.py`,
`kdilution_pooled.py`, figure `kic5689121_dilution_centroid.png` (left).

**The bright neighbour shows no dip.** KIC 5689116's own KBonus PSF light curve at the
candidate ephemeris: 24 ± 41 ppm. If it hosted the signal it would need ≈ 1100 ppm in its own
photometry (it is 2.1× brighter in G). Its SAP light curve shows 137 ± 42 ppm, the expected
leakage of the target into its aperture.

**New, independent raw-pixel centroid test.** Flux-weighted centroids computed directly from the
13 calibrated KIC 5689116 target pixel files (which contain both stars), using only stamp
geometry and Gaia positions, no PRF model. In-transit shifts are detected (zero shift rejected,
χ² 504/24) and match the first-order prediction for a dip on the target (χ² 43.9/24, p = 0.008,
mild tension from the linear approximation and unmodelled faint stars). The bright neighbour is
rejected at Δχ² = 934, and every other Gaia source within 30″ at Δχ² ≥ 4,300.
`kcentroid_tpf.py`, `kcentroid_all.py`, figure (right). This shares photons with the Codex PRF
localization but no modelling choices.

*Caveat:* the KBonus `CENTROID_COLUMN/ROW` columns show no in-transit response at all
(`kcentroid.py`, χ²(zero) = 25.5/26); they appear to be model-derived and are not usable for
this test. Only the raw-TPF version is quoted.

**Search statistics alone do not single it out (new negative result).** Ranked against all
1,736 matched inverted Kepler Bonus searches (pilot + both extension stages), 2 inverted stars
pass the first screen at P > 1.5 d with BLS SNR ≥ 26.7 and holdout SNR ≥ 15.5 (KIC 8701308 at
4.25 d, KIC 5283912 at 8.78 d). The empirical per-star false-alarm rate at this detection
strength is therefore ≈ 1e−3, i.e. ≈ 2 expected null-level signals in a survey this size. The
case for KIC 5689121 rests on the vetting (shape, density, localization, dilution pattern), not
on detection strength. `null_fap2.py`.

**TRICERATOPS FPP ≈ 0.07, not validated.** Run offline on the target alone (on-target and
unresolved-blend scenarios), Kepler mission settings, 29.4-min exposure, the folded PSF light
curve, and a field population built with TRICERATOPS' own Gaia recipe from the 120″ Gaia DR3
cone already in the evidence pack (resampled to the 0.1 deg² its priors assume).

| Background population | Runs (N = 1e6 each) | FPP |
|---|---|---|
| Gaia G < 21 | 4 | 0.059, 0.089, 0.058, 0.069 (mean 0.069) |
| Gaia G < 21 + extrapolated to G = 22.5 | 2 | 0.058, 0.088 |
| G < 21 + *hypothetical* Keck K-band AO curve | 2 | 0.022, 0.032 (mean 0.027) |

The PSF depth itself depends on the measurement choice: 1830–2430 ppm across box widths and
baseline windows in the independent check. It is 13–22σ against the event-to-event and
random-phase scatter, which is far more conservative than the formal errors (≈ 43σ).

Mean scenario probabilities (G < 21): TP 0.75, PTP 0.17, **STP 0.068**, DTP 0.009, BTP+BEB
< 0.001, every eclipsing-binary scenario ≈ 0. The FPP is almost entirely "a planet transiting an
unresolved bound companion". Even a typical AO contrast curve (hypothetical, not an
observation) only brings it to ≈ 0.02, because companions inside ~0.1″ (~30 au at 298 pc) stay
allowed. Gaia RUWE 0.98 already disfavours some of these, but TRICERATOPS 1.1 does not use it;
a MOLUSC companion file would.

NFPP is reported as 0 only because resolved neighbours were not passed to TRICERATOPS (that
requires a Kepler PRF/aperture model). The resolved-neighbour scenarios are instead excluded by
the neighbour light curve and the centroid test above.

## TIC 309780779 (TESS, P = 8.7236 d)

Host: TIC Teff 5013 K, R 0.685 R☉, M 0.83 M☉, d 74.9 pc, T = 10.03. Light curves are TESS-SPOC
FFI products at 1800/600/200 s plus two QLP sectors (89, 90), outside the documented QLP timing
problem (99–104).

**Blind recovery in each half of the data (new).** An independent BLS (astropy, own
detrending) on the full data returns 8.72363 d (SDE 19.6); on the inverted light curve the best
peak is SDE 7.8. Searching sectors 1–38 alone returns 8.72374 d as the top peak (SDE 12.6;
inverted best 6.0). Searching sectors 61–90 alone returns 8.72348 d as the top peak (SDE 12.9).
`bls.py`.

**Frozen-ephemeris holdout (new).** The ephemeris fitted to sectors 1–38 predicts the
sectors 61–90 mid-transit 89 cycles later, and the late transit lands on the prediction within
its uncertainty: the MCMC fits give −0.95 min (prediction ±7.5 min, 0.12σ). An independent
check with separate code finds +4 min (−12 to +11 min). Depth in the late sectors at the frozen
ephemeris is 380–450 ppm depending on box width (9–10σ), with 22/22 events positive.
Discovery-half and late-half periods agree at 1.2σ. `holdout.py`, figure
`tic309780779_holdout_phase_scan.png`.

**Fixed-ephemeris diagnostics.** Depth 401 ± 30 ppm; odd/even 365 ± 41 vs 444 ± 42 (1.3σ);
depth flat across cadence eras (367 ± 67, 423 ± 55, 408 ± 44 ppm). No secondary at any phase:
phase 0.5 is −20 ± 26 ppm (−23 to −48 ppm in the independent check, consistent with zero), the
largest off-transit dip in 377 phase trials is 3.2σ, and a phase-shift null never approaches
the transit (13.5σ). `scan.py`.

**Shape.** With a 15% Gaussian prior on the TIC density: b = 0.953 ± 0.015, Rp/R* = 0.0274
(+0.0027 −0.0019), Rp ≈ 2.05 (+0.20 −0.14) R⊕, T14 = 1.07 h, 14% of posterior grazing.
Without the prior the fit runs into the grazing branch (85% grazing, large radius ratio): the
data cannot tell a ~2 R⊕ planet at high impact parameter from a grazing larger body. `tfit.py`.

**TRICERATOPS FPP ≈ 0.57, not validated.** Same offline setup, TESS settings, 600-s-equivalent
series (native 600-s cadences + 200-s cadences averaged in 600-s bins, 34 of 46 events), field
population from the 90″ Gaia neighbour list in the pixel output. Three runs: 0.556, 0.579,
0.568. Dominant scenario: **SEB ≈ 0.42**, an eclipsing binary bound to the target (best fit: a
0.18 M☉ pair, grazing at b ≈ 1.0, diluted by the K dwarf), with TP 0.35. This is the textbook
false positive for a short, high-impact, shallow dip on a bright star. With a *hypothetical*
832-nm speckle contrast curve the FPP only falls to 0.44–0.47 (SEB still 0.38), because imaging
cannot resolve a close hierarchical companion.

## What was not possible here, and why

- **TESS pixel re-analysis with the official PRF, Gaia RUWE/NSS for TIC 309780779, live
  ExoFOP/TOI status for either star:** MAST, the Gaia archive, VizieR and ExoFOP are unreachable
  from this environment (network allowlist / robots). These are hard obstructions, not skipped
  steps.
- **High-resolution imaging, reconnaissance spectra, RVs, multi-band transits:** need telescopes.
- **NFPP for KIC 5689121 inside TRICERATOPS:** needs a Kepler PRF + aperture flux model for every
  neighbour; handled with direct tests instead (above).

## What would actually move each lead

**KIC 5689121.** AO imaging (Keck/NIRC2 or similar; the target is Kp 16.3, K 12.97) plus a
companion constraint that uses RUWE 0.98 (MOLUSC). A multi-band transit (STP predicts
colour-dependent depth) would also help. A radial-velocity mass is not realistic for this faint
M dwarf. Realistic endpoint: *statistically validated*, if imaging and RUWE squeeze STP below
~1%.

**TIC 309780779.** The bound-EB scenario dominates, and imaging cannot reach a close
hierarchical companion. Two or three reconnaissance spectra (the star is T = 10) would show
an SB2 or a large RV trend. For a 2 R⊕ planet the expected semi-amplitude is ~2 m/s, so a
true RV mass needs a precision spectrograph. A multi-colour or space-based (e.g. CHEOPS)
transit would test chromaticity and pin down the shape. Until then it is a candidate with an
FPP near 50%.

## Reproducing

Scripts use absolute paths from the working session (`/root/ph/...`). Point them at the merged
release archives: Kepler PSF/SAP light curves and FITS under
`planet-search/database_expansion/kbonus*`, TESS light curve `evidence/tess/lightcurves`. Python
3.13; TRICERATOPS 1.1.0, lightkurve 2.6.0 and pytransit 2.2.0 in a virtualenv; batman 2.5,
emcee 3.1, astropy 8.0. TRICERATOPS is driven without its `target.__init__` (no MAST): see
`scripts/triceratops/run_*.py` for the exact star tables and inputs. Contrast curves in
`scripts/triceratops/hypo_cc_*.txt` are **hypothetical**, for what-if runs only.

No candidate was submitted, named or announced.