Codec Score Runtime Denominators

Codec score sources can be measured under three different runtime policies. Keep them separate in experiment summaries and paper text.

live_pyav

The runner decodes video and extracts H.264 metadata through the current PyAV helper inside each model arm. This is the conservative local wall-clock denominator. It measures what this repo pays today, but it repeats extraction for each score source unless a script explicitly shares sidecars.

Use this denominator for implementation honesty. Do not describe it as free.

sidecar

A CPU prepass writes runner-ready score grids to .npz sidecars, and the model arm loads those grids during inference. This separates one-time metadata extraction from model-side sparse-vision timing.

Use this denominator for session or corpus serving analysis where a stable video can be preprocessed once and queried many times. Report sidecar build time, sidecar load time, and the first-query drift policy separately.

Claim-bearing sidecars must be bound to the current manifest item IDs, current git commit, current score-projection version, geometry details, and score-grid hash. A clean sidecar from an older commit is stale evidence unless the preregistration explicitly permits historical reuse.

decoder_integrated

The video decoder exposes macroblock metadata while producing pixels, avoiding the current separate PyAV pass. This is a systems hypothesis until measured in a decoder-integrated implementation.

Use this only as future-work or as a modeled ceiling. Do not report it as a reproduced wall-clock result.

Reporting Rule

For every codec-scored Track B or OV-8 row, report:

• codec_score_runtime_source: live_pyav, sidecar, or decoder_integrated hypothesis.
• model-side end-to-end time excluding score runtime.
• end-to-end time including the measured score runtime for the current path.
• sidecar build/load time or live extraction time, whichever applies.
• for sidecars: unique-item extraction total, projection/write total, total prepass wall time, and per-item load time. Do not sum per-source copies of the same unique-item H.264 extraction time.

Never multiply Track A active-refresh budget, Track B sparse-vision speed, and C-PERSIST follow-up speed. They are different denominators.