Architecture¶

SXS is organized by scientific responsibility. Command orchestration is thin; domain modules create machine-readable artifacts and reports that can be audited independently.

flowchart TB
    CLI[src.cli] --> BASE[baseline]
    CLI --> SCALE[scaleup]
    CLI --> SEARCH[search]
    CLI --> VALIDATE[validate]
    BASE --> CORE[Shared processing modules<br/>ingest · preprocess · detect · model]
    SCALE --> CORE
    SEARCH --> CORE
    VALIDATE --> AUDIT[Independent evidence<br/>FAP · photometry · crossmatches]
    CORE --> RECORD[(Reports and artifacts)]
    AUDIT --> RECORD

Module boundaries¶

Package	Responsibility
`src.ingest`	Catalog access, MAST acquisition, manifests, positive/negative datasets
`src.preprocess`	FITS extraction, quality filtering, interpolation flags, detrending
`src.detect`	Period grid, BLS evaluation, distinct peaks, recovery metrics
`src.model`	Feature/folded-view construction and baseline RF/CNN evaluation
`src.scaleup`	Scaled target selection, acquisition, training, and frozen model policy
`src.candidate_search`	Unknown pool, deterministic sample, model scoring, shortlist
`src.independent_validation`	FAP, physical vetting, external crossmatches, classification
`src.validate`	Baseline catalog checking and benchmark reporting

Four public workflows¶

Baseline¶

A six-stage resumable orchestration: environment, acquisition, preprocessing, BLS detection, machine learning, and catalog validation.

Scale-up¶

Builds a larger quality-filtered benchmark, trains RF v2/CNN v2, applies the prespecified selection policy, and writes the production model selection metadata. It does not run candidate screening.

Search¶

Builds the eligible unknown pool, selects targets deterministically, processes four products per target, scores five BLS peaks per target, and freezes the top-20 queue after preliminary checks.

Validate¶

Consumes the frozen queue and processed light curves. It runs the independent FAP, photometric, physical, and external-evidence stages before finalizing categories.

Design invariants¶

configurations are versioned;
target identities group cross-validation splits;
raw mission inputs are cached rather than rewritten;
expensive workflows support resumption where safe;
accepted stages have explicit artifact contracts;
ranking is separate from validation;
candidate labels include an unconfirmed warning; and
structured run records preserve status and provenance.