API Reference¶

The API reference is generated from the shipped type annotations and docstrings. It is the authoritative source for constructor signatures, defaults, public methods, properties, return types, and model-specific caveats. Use the user guide for lifecycle and selection guidance.

Public package surface¶

• Base interfaces and protocols
• Built-in catalog and declarative configuration
• Anomaly models
• Transformers
• Drift detectors
• Dataset registry, cache, and streams

Objects exported from the documented package __init__.py files are public. Private names beginning with _ are implementation details. Optional PyTorch objects are public only through their explicitly documented import paths and are excluded from wildcard exports when the dependency is absent.

Compatibility policy¶

Starting with 1.0.0, ABERRANT follows Semantic Versioning for the documented public API. Within 1.x, existing public import paths, constructor parameters, methods, properties, return types, and documented behavior remain compatible. Minor releases may add compatible capabilities; incompatible public API changes require a new major version.

Private modules, names beginning with _, and undocumented internal attributes are excluded from this guarantee. Experimental model descriptions qualify algorithm maturity; they do not exempt documented public interfaces from the compatibility policy.

Bug fixes may correct numeric results, including score values and seeded score sequences, to match documented semantics. Exact floating-point results or random sequences across package, dependency, and platform versions are not guaranteed. Review the changelog, revalidate score thresholds, and run a chronological canary stream when upgrading.

Model object layout and pickle/joblib checkpoints are not a stable interchange format. Cross-version serialization compatibility is not promised; see persistence and upgrades.

Shared model shape¶

Most anomaly models satisfy ModelProtocol:

• learn_one(x) -> None learns one feature mapping;
• score_one(x) -> float scores without learning the candidate event.

That shared method shape does not standardize input fields, warm-up, memory, or numeric scale. In particular:

• isolation-family scores are not universally bounded because RandomCutForest defaults to raw CoDisp and StreamRandomHistogramForest returns a raw log-mass score;
• graph detectors default to raw scores but expose optional normalization;
• moving statistics can return signed changes when abs_diff=False;
• RandomModel is a generator-backed baseline rather than an anomaly model.

Read the model score contracts before calibrating or comparing scores.