Research: AI full-feature refresh defaults¶

Companion to ADR-0642.

Findings¶

The refresh audit found three stale-data risks:

Several scripts defaulted to build/tools/vmaf, /usr/local/bin/vmaf, or PATH lookup. Those locations can point at an older binary without fork-only extractors such as motion_v2, SSIMULACRA2, and the SpEED features.
extract_ugc_features.py preserved an old v5 shortcut that emitted canonical-6 rows and filled the rest of FULL_FEATURES with NaN. That shape is incompatible with current full-feature FR refreshes.
KoNViD-1k had a synthetic-FR pair extractor but not a full-feature table generator with deterministic fold metadata.

Local smoke checks used the current fork CPU binary at core/build-cpu/tools/vmaf. BVI-DVC raw-YUV input produced all-zero VMAF in a one-clip probe, while the local lossless MKV bundle produced finite metrics; the refresh driver therefore accepts .mkv, .mp4, and .yuv but documents the MKV bundle as the known-good source.

Verification¶

Focused tests:

.venv/bin/python -m pytest \
  ai/tests/test_feature_extractor_defaults.py \
  ai/tests/test_bvi_dvc_dir_mode.py \
  ai/tests/test_extract_k150k_features.py \
  ai/tests/test_e2e_frame_to_score.py \
  ai/tests/test_extract_ugc_features.py \
  ai/tests/test_konvid_full_features.py \
  ai/tests/test_combine_full_feature_parquets.py \
  ai/tests/test_feature_sets.py \
  -q

Real local smokes completed for:

one KoNViD-1k full-feature clip (200 rows, current FULL_FEATURES + vmaf);
one UGC distorted pair (3 frames, current FULL_FEATURES + vmaf);
BVI-DVC tier-D lossless MKV refresh (193 clips, 6176 frames).

Long-running local background jobs continue for Netflix, KoNViD, UGC, and CHUG; their outputs are gitignored corpus artifacts and not part of this PR.

Research: AI full-feature refresh defaults¶

Findings¶

Verification¶

References¶