Narrative Contracts 0.1.0a2 — reproducible open alpha

Narrative Contracts a2 adds reproducible evidence and a second application domain while keeping the core offline and dependency-free.

• Independent customer-support workflow: trusted transitions, branch selection, declared claims and valid paraphrases.
• Expanded synthetic campaign: 21 cases, 8/13 attributed fault detections, 4/5 controls preserved, and 3 explicit exclusions. Survivors and false positives remain public.
• Real-model pilot: all 32 responses from qwen3:1.7b and gemma3:1b, with frozen prompts/settings, raw outputs and deterministic offline replay.
• Community templates for missed obligations, valid text rejected and new contracts.
• Updated technical report. No human labels or LLM-as-judge labels were used; natural-output acceptance is not semantic accuracy.

Validation: 91 tests, Ruff, strict mypy, package builds, isolated wheel/plugin installation, eight selected source mutants killed, and offline replay. The attached provenance identifies the verified commit and CI run.

The pilot is deliberately published without score tuning: Qwen3 supplied 16 valid envelopes with consistent declarations but none passed the strict profile; all 16 Gemma3 responses failed the envelope. These results describe this small workload, not a general model ranking.

Install from the attached wheels:

python -m pip install \
  https://github.com/pablomate4b/narrative-contracts/releases/download/v0.1.0a2/narrative_contracts-0.1.0a2-py3-none-any.whl \
  https://github.com/pablomate4b/narrative-contracts/releases/download/v0.1.0a2/pytest_narrative_contracts-0.1.0a2-py3-none-any.whl

Core and plugin use MIT. Model weights are not distributed. The natural pilot zip includes model metadata, prompts and generated outputs; it is not human-approved advice. PyPI publication, independent annotation, external adoption and research submission remain optional follow-up work.

Technical report · Replay instructions · Submit a counterexample