↓ Skip to main content
Pranav
  1. Projects/

Can an AI agent find exoplanets?

2026 - present ·3 mins
Astronomy Evidence Systems Python Local AI
Pranav
Author
Pranav
Coding with Physics, Physics-ing with code
Table of Contents
IncludesTESS signal triageBLS diagnosticsEvidence dossiersLocal AI agent
TL;DRAstral is a bounded local AI agent that triages archival TESS signals, collects evidence, and stops before making a planet claim.

Overview
#

Astral is a local agent for archival TESS analysis. Astral selects the next evidence to collect for signals that may show a transit.

Astral is a triage system. It does not confirm planets. It does not calculate false-positive probabilities. A human must review any candidate.

The workflow uses a frozen TIC manifest, Lightkurve, BLS, archive queries, and catalog context from Gaia, SIMBAD, ADS, and the literature. A local Qwen model can prioritize work. Python validates every action. SQLite stores the workflow state and evidence artifacts.

Current progress
#

The offline workflow has 13 passing tests. The tests cover candidate review, abstention, source gaps, evidence dossiers, and queue exhaustion.

A live check succeeded for the NASA Exoplanet Archive and SIMBAD. ADS needs credentials. MAST, Gaia, and the live manifest remain limited in this environment.

The agent has five roles: Scout, Analyst, Context, Skeptic, and Reporter. A review_candidate is a provisional triage state. It is never a confirmation.

Architecture
#

The workflow has explicit controls. Qwen can suggest a work order. Python validates the TIC IDs, allowed actions, time limits, and evidence state before it runs the action.

Astral’s controlled Scout → Analyst → Context → Skeptic → Reporter loop.

Evidence and diagnostics
#

The current implementation uses a frozen label-blind manifest, bounded multi-sector SPOC analysis, deterministic BLS and diagnostic gates, and content-addressed artifacts. External context is retained as provenance and evidence; it cannot silently change the signal disposition.

Archival DSS2 field centered on TIC 149992242. The image shows the host-star field, not a detected planet.
Example phase-folded signal artifact from the presentation evidence set.
Example Gaia-neighbor context artifact.

Next steps
#

  1. Build a calibrated holdout cohort before starting a five-candidate search.
  2. Complete multi-sector signal handling and test recovery on a frozen cohort.
  3. Calibrate the LEO/PRF pixel adapter before it affects any disposition.
  4. Improve live-service recovery and provenance review.
  5. Add a human audit of the evidence dossiers.
  6. Prepare the current system and evidence package for a possible professor review. This review has not happened yet.

The goal is an auditable evidence trail and a controlled stop condition. The system does not produce an automated astronomical conclusion.

Tools
#

  • Python with typed schemas and deterministic service adapters.
  • Lightkurve/BLS for the frozen signal-analysis path.
  • SQLite WAL for workflow state, idempotency, events, model calls, and case snapshots.
  • Local Ollama models only through constrained JSON plans.
  • Offline fixtures for repeatable tests when archive services are unavailable.

Related