VESPER

Vetting Exoplanet Signals with Pixel-informed Evidence, Rapidly.

VESPER is a fast, learned vetter for transiting-planet candidates around TESS M dwarfs. Give it a star's TESS photometry and a candidate ephemeris (period, epoch, duration); it returns a calibrated probability that the signal is a planet, seven named impostor scores — in the output's aux_* column order: background eclipsing binary (aux_beb), window alias (aux_alias), eclipsing binary on target (aux_eb), stellar rotation (aux_rot), companion dilution (aux_comp), no signal at the ephemeris (aux_vide, printed as "empty"), spacecraft systematics (aux_sys) — and a keep/reject disposition, in about a tenth of a second per star, entirely offline (no external service calls).

The frozen model, its full performance record and its known limitations are documented in the model card, linked from this page: https://gitlab.com/yohanntschudi/exoplanets-project/-/blob/main/documentation/pipeline_ml/VETTER_MODEL_CARD.md

Install

Python ≥ 3.10 required — check with python3 --version. Work in a fresh virtual environment (system Pythons are often older and their site-packages messy):

python3.11 -m venv vesper-env          # any Python >= 3.10
source vesper-env/bin/activate         # Windows: vesper-env\Scripts\activate
pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install "tess-vesper[acquire]"

CPU-only torch is enough. The [acquire] extra adds lightkurve, used by vesper fetch to download and prepare TESS photometry for you. The wheel bundles the frozen model (weights, calibrator, decision threshold, M-dwarf stellar catalog) — nothing is downloaded at vet time — and the artifacts' SHA-256 fingerprints are checked at every load.

How it works

1. You provide the ephemeris — VESPER judges a candidate at a given (period, T0, duration); it complements your transit search and your follow-up, taking the statistical-vetting step in between.
2. VESPER prepares its own photometry. From stitched, median-normalised PDCSAP flux before any detrending, it removes stellar variability with its own Gaussian-process model on a 200 s grid — the exact preparation the network was trained and certified on. vesper fetch builds that input for any TIC directly from the TESS archive. Supplying your own curves instead: one pickle per host, named TIC{tic}_stitched.pkl, holding a dict with keys time (BTJD), flux (PDCSAP, quality flag 0, per-sector median-normalised to ~1.0, before any stellar-variability detrending), and optionally flux_err and sector — exactly what vesper fetch writes. Do not feed already-detrended flux: the certified numbers do not hold for it.
3. It folds and judges. The signal is folded into shape views (global, transit zoom, secondary, odd/even, per-observing-block), normalised to the star's own noise, and weighed together with the stellar parameters and — when you have them — difference-image centroid measurements.
4. It answers in calibrated units. p_cal is a calibrated probability; the shipped threshold (0.29565…, set on the model's sealed calibration split to keep ≥ 95 % of its planets — 96.1 % realized there, 92.2 % of confirmed planets kept on the certified hold-out) turns it into a disposition, and the named impostor heads tell you what the signal most resembles when it is not a planet.

Domain of validity: TESS M dwarfs (Teff ≲ 4,000 K), periods ~0.5–100 d, native cadences 20–300 s (served on a uniform 200 s grid; coarser products — e.g. native 600 s — are not servable). Anything the data contract cannot serve comes back as an explicit blocked:... status, never a silent skip. The contract checks the data, not the star: VESPER does not verify that your target is an M dwarf — a K- or G-dwarf curve will serve and return a number, but an unvalidated one. Check Teff yourself; outside the M-dwarf regime the output is uncertified.

Try it — the bundled example

Twelve signals on six M-dwarf hosts, both sides of the decision: L 98-59 (three small confirmed planets, plus one unconfirmed 1.05 d candidate signal — judge it yourself), TOI-700 (four confirmed planets, including the habitable-zone TOI-700 d), and four TFOPWG-certified false positives spanning the impostor zoo — a 22.6 % deep eclipse, a short-period eclipsing binary, a long-period shallow false alarm, and a 790 ppm signal with a perfectly planet-like depth.

vesper example
vesper fetch --candidates candidates_example.csv --out-dir example_curves
vesper vet-batch --candidates candidates_example.csv --curves-dir example_curves --out example_results.csv

The fetch downloads every SPOC-family sector of all six stars — TOI-700 alone has 30+ — through lightkurve's cache (~/.lightkurve/cache): a few hundred MB of FITS and several network-bound minutes on the first run. Stars whose pickle already exists in --out-dir are skipped, so an interrupted fetch can simply be re-run.

For your own targets, write the same kind of CSV (header exactly as below, no spaces after the commas; extra columns such as toi or note are ignored):

tic_id,period_days,t0_btjd,duration_hours,depth_ppm,planet_num
307210830,7.450726,1355.290002,1.68,1002.2,4

t0_btjd is BTJD = BJD_TDB − 2,457,000 (native TESS time). ExoFOP lists full BJD epochs — subtract 2,457,000, or the row comes back blocked:t0_not_btjd. The four first columns are required on every row; depth_ppm (feeds the deep-eclipse prior) and planet_num (row labelling) are optional. Single candidates:

vesper vet --tic 307210830 --period 7.450726 --t0 1355.29 --duration-h 1.68 \
           --curves-dir example_curves

Reading the output

The results CSV (--out) echoes your input columns, then carries the science columns below, plus diagnostics (p_raw — the pre-calibration score; kept_0.5/kept_0.9 — stricter fixed thresholds; cascade_depth_source; n_sec_measured/n_sec_served — centroid-evidence sector count before/after its cap; log_scatter; cadence_s_native/ cadence_s_served). The console prints the same information as one aligned table per host.

What p_cal is calibrated against. The calibration set is real M-dwarf light curves carrying injected planets and constructed impostors, held out from training — p_cal is well-scaled under that mixture. It is not a TRICERATOPS-style false-positive probability computed against a Galactic stellar-population model, and a high p_cal is not by itself grounds for statistical validation. Use it to rank and triage; validation still requires independent evidence.

License and citation

MIT. The reference article is in preparation; until it is out, cite this package and its model card (linked above), which carries the full provenance of the frozen model.