# items/ — the full item pool after step 0

`pool_items_base.parquet` (717,199 rows, one per question of 139 video benchmarks) columns:
benchmark, item_id (pipeline qid; unique), video_key (source reference), video_id (content hash of the
normalized video, or "k:<sha1-16>" when the video was not normalized at snapshot time), video_path
(cluster path, not usable off-cluster; frames are read from the frame cache instead), question,
question_raw, options (list of option strings, letter prefixes stripped), options_source,
option_prefix_stripped, n_options, options_raw, answer (gold letter), answer_idx, duration (s),
declared_task, declared_scene, license, format (mcq / open / ...), format_reason,
bench_question_format, in_sample (item belongs to the 300-item E1 sample).

`step0_kept.parquet` (717,199 rows): item_id, format, kept (False = removed in step 0),
removed_reason, dup_of, dup_cos. Steps 1-3 run on rows with kept == True and format == "mcq"
(299,366 items); step 2 and 3 additionally need a frame cache entry for the item's video_id.

`step0_removed.jsonl`: the removed rows with reasons; `screen_counts_full.csv`: per-benchmark counts;
`pool_schema_manifest.csv`: how each benchmark was enumerated.