Skip to content

Research-0109: tiny-AI real-weight limitation cleanup

  • Status: documentation consistency digest
  • Date: 2026-05-15
  • Relevant ADRs: ADR-0215, ADR-0223, ADR-0255, ADR-0261

Question

Which user-facing tiny-AI docs still describe FastDVDnet or TransNet V2 as placeholder / smoke-only surfaces after their real upstream weights landed?

Findings

  • model/tiny/registry.json and the model cards already describe fastdvdnet_pre and transnet_v2 as real-weight entries with smoke: false.
  • docs/metrics/features.md and docs/ai/roadmap.md still carried the older placeholder / real-weights-follow-up wording, which makes the user-facing feature matrix contradict the shipped registry state.
  • The package AGENTS.md notes for DNN / feature code also kept the old placeholder framing or stale ADR numbers, which raises rebase risk when the tiny-AI extractor contracts are touched.

Decision Support

Option Trade-off Decision
Leave the docs stale until the FFmpeg consumers land Avoids churn, but tells users the shipped real ONNX files are still placeholders. Rejected
Change code / registry to match the stale docs Would regress already-shipped real-weight model cards and registry entries. Rejected
Refresh user docs and invariant notes only Correctly separates shipped real-weight extractors from remaining consumer follow-ups. Accepted

Non-Goals

  • No model bytes, registry schema, or extractor runtime behaviour change.
  • No claim that the pending FFmpeg vmaf_pre_temporal or per-shot CRF consumers have shipped.