| # Protocols (identical to the paper's meta-benchmark runs; exp/analysis/meta/meta_runner.py) |
| |
| ## VIDEO (video1fps) |
| Frames: one frame per second of the timeline, at most 512 frames per item (uniform subsample beyond; the normalized |
| store holds at most 1,024 frames per video), each frame resized to short side 224. When a model's context or memory |
| rejects the request the frame count is halved and the request retried; the frames actually sent are recorded per item. |
| |
| Prompt (frames first, then the text): |
| These are {n} frames sampled from a video, in order. Answer the multiple-choice question about the video. |
| Question: {question} |
| Options: |
| A. ... |
| B. ... |
| Answer with the option's letter from the given choices directly. |
| |
| ## BLIND |
| Question and options only, no frames; four option permutations per item (seed 42, exp/analysis/meta/meta_runner.py |
| permutations_for); the blind score is the mean over the four permutations. |
| |
| ## Scoring |
| The rule scorer of the paper (exp/pipeline/scoring.py, mode "list"): the first option letter in the reply, or the option |
| whose text the reply repeats; a reply that names no option is wrong. Gain = VIDEO - BLIND on the same items. |
| |
