Skip to content

Research-0115: fr_regressor_v2 model-card status closeout

  • Status: Active
  • Workstream: docs-only closeout for fr_regressor_v2 model cards
  • Last updated: 2026-05-14

Question

Do the fr_regressor_v2 model cards still describe the model as scaffolded or planned after the registry has promoted the shipped ONNX to a production smoke: false entry?

Sources

Findings

fr_regressor_v2 is no longer scaffold-only. The registry marks fr_regressor_v2.onnx as a production checkpoint (smoke: false) trained on the vmaf-tune Phase-A JSONL corpus, with an in-sample PLCC of 0.9794 and a stable SHA-256 pin. The main v2 card still described the smoke-mode training path as the shipped artefact, and the older fr_regressor_v2_codec_aware card still implied that a separate fr_regressor_v2_codec_aware.onnx was the next planned output.

The docs should instead distinguish the shipped vmaf-tune v2 graph from the historical ADR-0235 design. The older card remains useful as design history, but its status needs to be "superseded" so operators do not look for a model file that is intentionally not in the registry.

Alternatives explored

Deleting the older codec-aware card was rejected because ADR-0235 and older research digests still link to it. Keeping it but marking it superseded keeps that audit trail intact while making the live production path unambiguous.

Adding a new ADR was rejected because this PR does not change model policy, model bytes, registry schema, or runtime behaviour; it only aligns public docs with already-shipped registry state.

Open questions

  • None for this closeout.