Skip to content

Release notes

These notes highlight user-visible changes in the current release candidate. The repository changelog is the authoritative history for every release.

0.1.0, the first public release

0.1.0 is the first version published to PyPI. It builds on the earlier 0.1.0rc1 through 0.1.0rc5 candidates, which were not published. The release:

  • separates the dataset-agnostic implementation into kuration.core and ships pixel shift, action-state consistency, and SLAM together under the maintained kuration.gameplay domain;
  • adds direct operator methods, lifecycle statistics and evidence, synchronized video review, smart dataset import, artifact export, checkpoint maintenance, and the unified kur command;
  • makes analyzers first-class pipeline terminals, including bounded tick-stream execution and artifact-backed checks;
  • integrates native SLAM reconstruction, portable exports, trajectory quality/diversity curation, point-cloud reduction, preflight, and kur-slam; and
  • strengthens fail-closed publication and reuse, exact cross-executor selection semantics, pinned artifact reads, source-level error handling, and boundary validation.

Dataset writes are no-clobber by default, persisted reuse requires a strong source identity, and unsupported future artifact or job versions fail loudly. See the 0.1.0 changelog for the complete evolving list.

Prior 0.1.0rc4 candidate (not published)

The 0.1.0rc4 candidate focused on correctness, artifact integrity, and explicit extension contracts. It was not tagged or published to PyPI from this repository.

New analyzer capabilities

  • A discoverable analyzer registry kind now supports user-owned capabilities that aggregate evidence across multiple segments. Register and resolve analyzers with register_analyzer() and get_analyzer(), or inspect them with kur-list analyzers and kur-describe analyzer ....
  • The gameplay domain's action-state consistency capability now exposes its clip-level checker as an analyzer.
  • The dependency-free kuration/split-leakage analyzer audits existing source and group assignments across dataset splits.

Correctness and validation

  • Pipeline and bounded execution now reject duplicate sources, feature-name collisions, invalid vector widths, scorer-parameter mismatches, and test-split errors consistently at public execution boundaries.
  • Adapter and media handling is stricter for integer timestamps, missing categorical values, Parquet inputs, Arrow batches, joined media, multi-video metadata, and timestamp alignment.
  • Plugin loading is transactional: malformed targets and true no-op entry points remain visible as failures, while valid registrations from pre-imported packages continue to work.
  • JSON Schema enum, const, and uniqueItems validation now distinguishes JSON types, so booleans no longer alias integers.

Artifact integrity and identity

  • iter_shards(), read_shard_arrays(), ShardDataset, and curated-dataset adapters verify recorded shard sizes and SHA-256 digests by default. Set verify=False only for artifacts you explicitly trust.
  • Manifest paths, commit sidecars, split names, and rebased legacy paths are validated before reads, preventing path traversal and reuse of stale or corrupted shards.
  • Shard filenames and manifest entity IDs preserve value types. Values such as integer 1 and string "1" no longer collide, and unsupported extension objects fail before writing.

Built-in SLAM

  • Exported timestamp maps must cover every trajectory exactly; each timestamp array must match the trajectory length, contain finite values, and increase strictly.
  • Point-cloud manifests and NPZ files now require supported Kuration format and version markers.

Upgrade notes

  • Artifact verification is now enabled by default. Callers that intentionally skip integrity checks must opt out explicitly with verify=False.
  • Custom entity-ID objects without a canonical value representation are no longer writable. Use None, booleans, integers, finite floats, strings, bytes, tuples, frozensets, or compatible NumPy scalars.
  • A plugin entry point must register at least one owned component; a successful import with no registration is reported as a plugin failure.
  • Action-state consistency now ships as a built-in analyzer using the analyzer registry API.

See the compatibility policy for beta-version guarantees and the releasing guide for the maintainer workflow.