Automo 0.3.0a2 public alpha contract
0.3.0a2 is the current public Automo alpha, extending the initial release with program/model/hypothesis governance and immutable model-composition research artifacts.
Supported use cases
- deterministic fit/validation/OOS research;
- user-defined data sources, features, trainers, evaluators, and external services;
- model registry, calibration lineage, lifecycle history, and benchmark evidence;
- model pools and data-iteration refresh workflows;
- bounded automated research with explicit budgets and OOS gates;
- composable model graphs and leakage-safe first-level stacking/meta-model training;
- logical model candidates with exact upstream candidate composition;
- controlled composition/ablation experiments for incremental meta-model value;
- datasets using time, ordered IDs, groups, stable hashes, or predefined partitions.
Not claimed
The alpha does not promise globally optimal model search, production deployment, distributed training, arbitrary online learning, deep recursively nested cross-fitting, or sandboxing of third-party code.
Persistent artifacts
Public registry, refresh, pool, and research artifacts are written with artifact_type and schema_version. The initial public schema version is 1.
Extension-code trust
Project plugins, trainers, evaluators, codecs, calibrators, injected services, and workflow providers run with normal Python process privileges. See SECURITY.md.
Release gate
The release candidate must pass the canonical full-repository pre-commit gate locally and in CI. That gate includes repository hygiene, Ruff lint/format, and pytest locally plus GitHub Python 3.11, 3.12, and 3.13 compatibility jobs. Source compilation, runnable examples, and wheel/sdist build/install smoke tests remain separate release gates.
Agent research governance
The public alpha includes packaged agent-facing research guidance, .automo/ program/model/hypothesis governance, and an optional bounded GetDone capability handoff. GetDone development state remains separate under .agent/.