meeting-assistant/docs/alpha-validation.md

Alpha paired validation

Run the complete Assistant suite with PYTHONPATH=src:<Lab candidate root>. tests/test_alpha_pair.py exercises the real adapter, context, orchestration, transcript selection and generation persistence with mocked audio preparation, Whisper, diarization and model responses. It covers de/en, all performance profiles, anonymous generation, mapped regeneration, glossary provenance and immutable history. The test skips when Lab is unavailable; a release validation must run it with Lab available and report no skips.

Run Lab's full python -m unittest discover -s tests at its candidate revision. History tests inject write/publication failures and a process interruption. UI tests use the real executor and reject worker access to session state.

These tests do not measure recognition quality or model language compliance. Before tagging, perform the final local GUI/model smoke with the exact paired revisions and configured model/runtime. Reuse copied historical transcripts where possible; do not rewrite original regression evidence. Preserve the known GTM-Hub human-reference whitespace. Record both candidate SHAs in release notes.

Historical Lab fixture dependency

Four older experiment tests require six ignored files, absent from a fresh Git worktree. Do not add them to the alpha source boundary or silently skip the tests:

• artifacts/experiments/evidence_observations_v3/20260819_v3_single_run/h_resulting_action/parsed_observations.json
• artifacts/experiments/negative_act_form_v0/20260820_qwen35_9b_single_run/na-01/v3_style_input_observations.json
• The same negative-act filename under na-03, na-04, na-05, and na-06.

For the alpha audit, these files were copied from the original Lab worktree to a separate temporary test directory. That directory exposed the candidate's tests, prompts, samples, src, and scripts as resource links. The full suite ran there with the candidate Lab root on PYTHONPATH and the candidate's absolute tests directory supplied to unittest discovery. All 413 tests passed. The bare candidate worktree run instead reports four missing-fixture errors. This is a historical test portability limitation, not a runtime dependency.

The paired Assistant suite passed all 107 tests with the candidate Lab backend; no Whisper, diarization or Ollama inference was executed. These counts describe the preparation validation and do not replace the final smoke-test record.