Skip to content

Research 0684: CHUG extraction report provenance

Question

Which CHUG feature-extraction artifacts still lacked ADR-0661 provenance before HDR MOS training starts?

Findings

  • chug_extract_features.py --split-manifest writes the local content-to-train/validation/test map that defines the CHUG holdout boundary.
  • chug_extract_features.py --audit-output writes the HDR metadata preflight used to catch malformed PQ/HLG rows before training.
  • Both files are durable local evidence, but neither recorded the exact CHUG JSONL, clip/cache directory, VMAF binary, argv, or extraction arguments that produced them.
  • The feature JSONL rows already carry model-facing metadata; the missing piece is report-level provenance for the split and audit JSON files.

Decision

Add optional run_provenance plumbing to write_split_manifest() and audit_chug_hdr_metadata(), and have the CLI pass an ADR-0661 block covering the input JSONL, clips/cache directories, VMAF binary, split/audit targets, feature output, argv, and parsed arguments. Keep feature-row JSONL output unchanged.

Commands

rg -n "split-manifest|audit-output|run_provenance" ai/scripts/chug_extract_features.py ai/tests/test_chug.py docs/ai/chug-ingestion.md
.venv/bin/python -m pytest ai/tests/test_chug.py -q