Skip to content

Research-0100: vmaf-tune / Tiny-AI Scaffold-State Audit

  • Date: 2026-05-14
  • Scope: tools/vmaf-tune/, model/tiny/, docs/state.md
  • Trigger: user request to continue unblocked AI and vmaf-tune scaffold cleanup before the expected upstream HDR-model release.

Findings

  1. tools/vmaf-tune/src/vmaftune/auto.py still emitted a single hard-coded PQ hdr_args tuple for every HDR cell. That bypassed vmaftune.hdr.hdr_codec_args(), so x265/SVT-AV1/NVENC/VVenC cells could not surface their codec-specific HDR signalling.
  2. auto.py computed recipe-adjusted effective_thresholds and then re-created thresholds = confidence_thresholds or ConfidenceThresholds() before the cell loop. The JSON metadata and confidence decisions therefore ignored the recipe-tightened F.3 gate.
  3. model/tiny/registry.json still marked fr_regressor_v2_ensemble_v1_seed{0..4} as smoke: true. ADR-0321, ai/AGENTS.md, the model card, PROMOTE.json, the per-seed sidecars, and the on-disk SHA-256s all point to the production full-corpus flip.
  4. docs/state.md still listed T-HDR-ITER-ROWS and Tiny-AI C1 in Deferred even though both have Recently closed rows and landed implementation evidence.
  5. Fetched upstream/master still has no model/vmaf_hdr_*.json entry as of this audit. origin/master only carries the local model/vmaf_hdr_model_card.md fallback.

Action

  • Wire auto HDR cells through hdr_codec_args(codec, info).
  • Preserve recipe-adjusted F.3 thresholds through the cell loop and emitted metadata.
  • Flip the five ensemble seed registry rows to smoke: false, backed by their existing sidecars and PROMOTE gate evidence.
  • Remove stale Deferred rows from docs/state.md.

No New ADR

No new architectural decision is introduced. These changes implement already-accepted contracts in ADR-0300, ADR-0321, and ADR-0325 and repair stale scaffold leftovers.