Skip to content

Evidence pack — v0.0.22 STT observatory (Wave 1)

Dated: 2026-08-02
Evidence version: 0.0.22-observatory-v1
Linear: JOE-2216 (parent JOE-2215)

What shipped in Wave 1

  • Versioned observatory corpus + report schemas in aurum-core
  • Fail-closed budget comparison (Rust + scripts/eval/compare_stt_budget.py)
  • Redistributable core corpus and production pack recipe (external open data)
  • Profile evidence version pin to this pack
  • Model/profile matrix under evals/observatory/budgets/matrix.json

What is not claimed yet

  • A full 60-minute multi-speaker human speech baseline checked into CI — the production pack is operator-prepared from licensed open sources.
  • Automatic change of the global base default.
  • Universal field WER from the synthetic core alone.

How to extend to the full production baseline

  1. ./scripts/eval/prepare_stt_observatory_corpus.sh --slot <name> for each slot
  2. Generate evals/observatory/cache/corpus.production.json
  3. Run the local model matrix; retain reports under evals/reports/stt/
  4. Update committed budgets with report diffs + changelog
  5. Link the reviewed report on JOE-2215 / JOE-2226

Privacy

No private user audio, Plaud exports, API keys, or absolute private paths in retained public reports.