Honest accounting of what is proven vs aspirational at each phase boundary. These are
evidence documents — they feed ../../IMPLEMENTATION_STATUS.md.
| Retrospective | Canonical | Date |
|---|---|---|
| Phase 1 Retrospective | phase-1-retrospective.md |
2026-06-04 |
- The central distinction: architecture-validated (the platform can represent and orchestrate a concept — proven by unit/integration tests) vs provider-validated (a real model installs, loads, builds a variant, resolves, and generates audio).
- Result: broad architecture validation, narrow provider validation. ~237 backend tests across 55 files at writing prove the contracts; none load real weights, use a GPU, or synthesize audio.
- Only OmniVoice has a real engine. Fish is integrated at the contract level only
(
NotImplementedErrorinload/generate); Kokoro is research-only. - The thesis (model diversity absorbed by the adapter seam) holds in architecture but is unproven empirically — the real stress test is a non-cloning provider (Kokoro) which challenges ADR-0008's build-from-reference-audio assumption.
- Recommendation / readiness gate: do not begin SaaS/billing/auth/marketplace work until one non-OmniVoice provider generates real audio end-to-end.
Related: ../PROVIDER_VALIDATIONS/README.md ·
../../ROADMAP/MILESTONES.md · ../../OPEN_DECISIONS.md