Avril-Exoplanet/Avril

What it does

Avril is a local planet-candidate discovery system for public TESS data. It combines two parts: a frozen stellar prior that ranks the 1,000-target bundled queue by how incomplete the star's planetary census may be, and a transit search engine that runs multi-scale detrending, BLS, optional TLS, cross-sector comparison, harmonic checks, physical vetting and optional pixel-level source localization.

Instead of keeping only one detection peak, the engine keeps every base peak and locked ephemeris in an audit ledger, so candidates can be reviewed later with their periods, events, diagnostics and rejection reasons intact. It targets students, researchers, educators and enthusiasts who want to re-search public light curves on their own Mac rather than rely solely on large survey pipelines.

Quick start on macOS

Requirements are macOS 15 or newer, Xcode command-line tools with Swift 6.1, and Python 3.11–3.13; internet access is needed only to download public TESS products. From the repository root:

git clone https://github.com/Avril-Exoplanet/Avril.git
cd avril
./scripts/bootstrap_engine.sh scout
AVRIL_NOVA_SKIP_RUNTIME=1 ./scripts/build_app.sh
./scripts/run_dev_app.sh

The same engine is usable from the command line after activating .venv:

avril doctor --backend cpu
avril run "TOI-786" --source spoc --pixels none --backend cpu

avril run prints a result and run record for the target; --pixels followup adds optional source localization. Large data can be redirected with AVRIL_DATA_DIR.

Constraints and status

Avril V3 is a research preview. The documented checks—24 public-interface pytest cases, 93 deterministic scientific-core regressions, Ruff, a Swift release build and a source-build signature audit—verify packaging and declared invariants, not population-level completeness or false-positive calibration. The Planetary Census Gap AI scores are out-of-fold Kepler measurements (for example, 0.7862 ROC AUC and 4.82× top-decile lift for the confirmed-clean prior); they are catalog-host enrichment, not calibrated TESS planet probabilities. Before survey-level claims, the project asks for a TIC-disjoint sealed split and injection/recovery in real TESS noise.

The default source build is not notarized. The macOS app is native SwiftUI/AppKit with no WebView, and FITS data, run records and reports stay on the local machine. Search-path changes require deterministic tests and evidence that candidate counts and ephemerides are preserved.

HotRepo's take

Avril appeared on HotRepo's new-project Python list on 2026-11-12 with 103 stars and 3 stars gained over roughly 90 minutes of tracked growth, with no recorded daily all-language rank yet. What stands out for a project this young is the emphasis on auditability: a frozen target prior validated on star-level folds, a CPU/CUDA backend split that the README says changes only execution speed, and an explicit research-preview framing that separates discovery rank from calibrated planet probabilities. The combination of a native macOS workstation and a scriptable engine gives readers two entry points without changing the underlying science pipeline.

Project source and documentation