An autonomous exoplanet-candidate discovery and vetting system for TESS's blind spots — classical algorithms detect transit signals, an adversarial AI tribunal tries to kill every candidate, and survivors are published as fully-provenanced dossiers on a static web experience.
Everything this system outputs is a vetted candidate, never a "confirmed planet." Confirmation requires radial-velocity follow-up, which is out of scope. Candidates are framed ExoFOP-style: for community follow-up.
TESS observes millions of stars; NASA/MIT pipelines vet a curated subset. Citizen scientists keep finding real planets in the leftovers by eye — proof the bottleneck is attention, not data. Long-period single-transit events are the starkest gap: periodic searches need ≥ 2 transits, so a planet that transits once per sector is invisible to them by construction. This system supplies that attention at scale.
| What | Result |
|---|---|
| Known-planet recovery (calibration gate) | 8/8 (100%) — WASP-18 b, WASP-121 b, WASP-126 b, WASP-100 b, WASP-62 b, pi Men c, LHS 3844 b, HD 219666 b all re-found blind; pi Men c at 221 ppm is the depth floor demonstrated |
| Injection-recovery completeness | 73% overall — 192 trapezoid injections into RAW flux, searched through the full production detrend+search chain (≥93% at ≥3000 ppm, ~20–40% at 300 ppm: the honest small-planet limit at 10-min binning). An earlier 81% figure that bypassed detrending was caught in adversarial review and corrected. |
| Reliability | 95% (recovered injections / all SDE ≥ 9 detections, incl. block-shuffle null-trial false alarms and wrong-period recoveries — full accounting in data/calibration.json) |
| Blind backlog batch | 40 sector-1 SPOC targets not in any TOI/confirmed/EB catalog: 37 honest no-detections, 1 periodic candidate, 3 single-transit events (period lower bounds 18.1–22.2 d, event-to-edge) |
| Tribunal verdicts | all 8 calibration planets correctly killed by the catalog skeptic ("not novel" — dedup proof); all 4 novel candidates flagged for human review with low-confidence judge syntheses and concrete follow-up plans; 0 survivors — the honest outcome for single-sector evidence. Notably, the blend skeptic surfaced a real 7.3σ in-transit centroid shift on the periodic candidate. Every skeptic's full reasoning is in data/candidates.json. |
Caveats, stated plainly:
- The false-positive number is heuristic-fpp-v1 — a transparent, documented ranking
score (formula in
pipeline/validate.py, published verbatim on the site's Method page). It is not a Bayesian FPP; TRICERATOPS integration is future work. - Hot Jupiters in the calibration set score "suspicious" on that heuristic because they really do have secondary eclipses — a known, visible limitation, not a hidden one.
- The periodic blind candidate sits at the period-search grid edge (13.94 d), a classic alias signature; the tribunal weighs exactly that.
- Injection-recovery used hundreds of trials (single-session compute), not thousands;
scope is recorded in
data/calibration.json. - The Villanova TESS-EB catalog endpoint was unreachable during this run; EB dedup relies
on TFOPWG FP dispositions + the confirmed/TOI lists. Recorded in every dossier's
catalog_status. - This repo was built and adversarially reviewed in a single session by Claude (Fable 5) on real MAST data; the review found and fixed two genuine science defects (injection path bypassing detrending; single-transit period bound overstated) before publication.
Two decoupled layers joined by a JSON contract (data/):
pipeline/ Layer A — deterministic Python. ingest (lightkurve/MAST, SPOC 2-min
PDCSAP) → detrend (transit-safe Savitzky-Golay) → search (TLS + BLS
cross-check) → diagnostics (odd-even, secondary w/ bootstrap, trapezoid
shape, TPF centroid shift, TIC-cone neighbors, systematics flags,
rotation aliases) → single-transit matched filter → heuristic FPP →
catalog crossmatch (TOI / NASA confirmed / EB best-effort)
panel/ Layer B — adversarial AI. Six skeptics (binary, blend, instrument,
stellar, alias, catalog) + a judge, driven through headless `claude -p`
with strict JSON schemas. Skeptics reason ONLY over computed diagnostics;
physical skeptics are blind to catalog matches. Survival is enforced by
a mechanical rule in code — the judge cannot override a fatal kill.
calibration/ The trust gate: known-planet recovery + injection-recovery (completeness,
reliability, 1-day block-shuffle null trials). Runs BEFORE any claim.
data/ The contract: candidates.json, calibration.json, run-meta.json.
web/ Next.js 16 static export. Renders ENTIRELY from data/*.json — no number
on the site exists outside the pipeline output.
tests/ pytest: synthetic recovery, EB-tell diagnostics, single-transit hunter,
contract integrity (16 tests).
The prime directive: the LLM never invents a number. Every quantity is computed in
pipeline/; the AI layer interprets, argues, and writes prose. Diagnostics that could
not be computed are reported as computed: false with a reason — missing evidence is
never papered over.
# environment (Python 3.12)
python3.12 -m venv .venv
.venv/bin/pip install -r requirements.txt
# 1. calibration gate (downloads real TESS data from MAST; ~1 h)
.venv/bin/python -m calibration.run
# 2. science pipeline: calibration hosts + blind backlog batch
.venv/bin/python -m pipeline.run --calibration --blind
# 3. adversarial tribunal (needs the Claude Code CLI authenticated)
.venv/bin/python -m panel.run
# 4. the site
cd web && npm install && npm run build # static export in web/out/
# tests
.venv/bin/pytest tests/Every dossier records TIC ID, sector, data product, pipeline + contract versions, generation time, stellar-parameter source, every diagnostic, every skeptic verdict, and the folded light-curve series — exportable as JSON from each dossier page for independent verification or refutation.
MIT © 2026 Dennis Limbu.
Built on public data and open science tooling: TESS light curves via
MAST (lightkurve / astroquery), period search with
transitleastsquares, and catalog crossmatch against
the NASA Exoplanet Archive / TOI lists. This project outputs vetted candidates for
community follow-up, never confirmed planets.