ExoSAGE

ExoSAGE is a point-in-time, risk-controlled research system for exoplanet transit screening and rocky/temperate follow-up allocation.

It treats discovery like a quantitative research problem: remove shared systematics, search residual signals, control a large hypothesis universe, calibrate uncertainty through time, and allocate limited telescope hours under correlated risk.

The project does not claim to detect life or output a ground-truth probability of habitability.

What Works Now

• Immutable NASA Exoplanet Archive TOI snapshots with query, retrieval time, schema, SHA-256 checksums, and overwrite protection
• A tracked 160-row NASA fallback slice for reproducible UI and CI execution
• A real 12,000-point TESS/MAST light curve for Pi Mensae c
• Robust low-rank factor residualization for target-by-time flux panels
• Box Least Squares search with robust cadence-noise estimation
• Benjamini-Hochberg false-discovery control and calibration diagnostics
• Monte Carlo rocky-composition and habitable-zone characterization
• Point-in-time walk-forward backtesting with future outcomes used only for evaluation
• Mixed-integer telescope allocation with budget, target-count, idiosyncratic risk, and shared factor-exposure risk
• A FastAPI research service and responsive vanilla JavaScript/D3 workstation
• CI tests, remote-only training guardrails, and a Linux container

The current candidate ranking is a transparent physics-and-observability screen, not a trained classifier. The workstation says this explicitly.

Architecture

flowchart LR
    A[NASA TOI snapshots] --> B[Point-in-time universe]
    C[TESS and Kepler light curves] --> D[Common-factor residualizer]
    D --> E[BLS signal search]
    E --> F[Vetting and FDR control]
    B --> G[Physical uncertainty model]
    F --> H[Calibrated discovery score]
    G --> I[Candidate evidence]
    H --> J[Walk-forward evaluation]
    I --> K[Covariance-aware allocator]
    H --> K
    K --> L[Telescope follow-up portfolio]

Loading

The system separates three questions:

1. Does the photometric signal look planetary rather than instrumental or astrophysical contamination?
2. Given measurement uncertainty, how plausible are a rocky composition and a conservative or optimistic habitable-zone orbit?
3. Which candidate set maximizes expected scientific value under limited telescope time and correlated failure risk?

Quant Research Design

Quant concept	ExoSAGE implementation
Point-in-time universe	Dated, checksummed NASA snapshots
Factor neutralization	Robust low-rank common-mode removal
Alpha signal	Residual transit-search statistic and vetting evidence
Multiple testing	Benjamini-Hochberg q-values
Calibration	Brier score, ECE, and MCE
Walk-forward research	Grouped temporal cutoffs with an embargo-ready config
Risk model	Candidate uncertainty plus shared sector/stellar/magnitude/radius exposures
Portfolio construction	Binary mixed-integer allocation under telescope-hour constraints

Verified Result

The API analyzes the tracked real TESS light curve and recovers Pi Mensae c at approximately 6.2660 days, close to its known 6.2679-day period. This is a signal-processing validation case, not a trained-model benchmark.

The Python/API suite contains 18 tests and currently maintains more than 80% statement coverage. The browser code is dependency-free apart from pinned D3 v7.9.0 loaded from a CDN, and CI performs a JavaScript syntax check.

Quick Start

Requirements: Python 3.12 and uv. Node is only needed for the optional browser-code syntax check.

make setup
make api

In a second terminal:

make web

Open:

• Workstation: http://localhost:5173
• API: http://localhost:8000
• OpenAPI docs: http://localhost:8000/docs

The committed demo assets are enough to run the complete vertical slice. To create a new full NASA snapshot:

make data

A snapshot for the current UTC date is immutable. The command refuses to overwrite it.

API Surface

Method	Route	Purpose
GET	/health	Data provenance and model-state disclosure
GET	/api/candidates	Ranked point-in-time candidate universe
GET	/api/candidates/{id}	Physical posterior and evidence decomposition
GET	/api/signal/demo	Real TESS signal views and BLS result
POST	/api/portfolio	Risk-aware follow-up allocation

Repository Map

src/exosage/
  data/          immutable ingestion, manifests, repository
  signals/       factor residualization and transit search
  risk/          FDR and calibration
  science/       physical uncertainty characterization
  backtest/      point-in-time walk-forward evaluation
  allocation/    covariance-aware telescope portfolio
  api/           FastAPI research service
  training/      remote-compute policy guard
web/             vanilla JavaScript and hand-written D3 workstation
scripts/         reproducible static-data build and remote training entrypoint
configs/         versioned experiment definitions
data/demo/       tracked public runtime assets
tests/           numerical, policy, and end-to-end API tests

The workstation can also run without the API. make web-data rebuilds the tracked browser payload from the public demo assets, and make web serves it with Python's standard-library HTTP server. See web/README.md for the visual encoding, interaction, and accessibility decisions.

Compute Policy

Training is disabled on macOS. Local execution is limited to data contracts, small public demo assets, unit tests, signal-processing validation, and model forward passes. Full preprocessing and training must run in a versioned Linux container with EXOSAGE_REMOTE_TRAINING=1.

make remote-train

The remote entrypoint remains intentionally locked until a labeled, point-in-time dataset and leakage audit are approved.

Research Standard

Public metrics must come from grouped, point-in-time, walk-forward experiments. Random row splits, future dispositions in features, proxy labels presented as ground truth, synthetic calibration metrics, and unversioned datasets are not accepted.

Rocky and habitable-zone posteriors are physical screening quantities. They are not evidence for an atmosphere, surface water, biological activity, or life.