Alignmenter 0.3.0

Alignmenter 0.3.0 adds a durable application evaluation workflow: capture answers, evaluate application commitments, compare saved baselines, review failures, promote regressions, and gate CI with preserved evidence.

• Durable SQLite capture, explicit recovery, shared judge reservations, and target call caps.
• Typed grounding and faithfulness assessments, strict evidence validation, and custom deterministic evaluator factories with registered metrics.
• Matched case comparisons, explicit missing populations, versioned absolute/regression gates, and offline HTML/JSON/Markdown/JUnit reports.
• Append-only human review, adjudication, evaluator/reference agreement reports, and regression promotion with source lineage.
• Public Python SDK, noninteractive suite CLI, read-only run archives, and an installed offline example.

Install and try the offline example:

pip install --upgrade 'alignmenter==0.3.0'
alignmenter init-suite --out evals/resource-task
alignmenter run-suite evals/resource-task/suite.yaml --out reports

The core supports Python 3.10–3.14 on macOS/Linux for durable execution, without torch or scikit-learn. Legacy persona/scorer APIs remain available; their older semantics and scorer-local budgets are documented in the migration guide. See the release workflow and SDK reference.

Validation includes 352 tests per supported Python version, calibration-extra tests, separate wheel/source installation rehearsals, strict documentation builds, and the static site build. Atlas fixtures preserve real failures and remain draft: actual judge runs and independent owner labels are still needed for product qualification. Device replay, distributed budgets, and hosted review remain future work.

Linux release CI passed all ten jobs before merge. The publishing workflow independently tests and rehearses both distribution formats from tag v0.3.0.