Protocols (identical to the runs of the paper; code: https://github.com/Espere-1119-Song/Video-Index)
VIDEO
Frames: one frame per second of the timeline, at most 512 frames per item (uniform thinning beyond; the videos are stored at no more than 2 frames per second and 1,024 frames), each frame resized to short side 224. When a model rejects the request for its size the frame count is halved and the request repeated; the number of frames sent is recorded per item.
Prompt (frames first, then the text):
You are given {n} frame(s) sampled from a video. Answer the question based on these frames.
Question: {question}
A. ...
B. ...
Reply with ONLY the option letter (or the exact short answer if no options).
BLIND
Question and options only, no frames; four option permutations per item, fixed by random.Random("42|").sample(range(k), k); the blind score is the mean over the four permutations.
You are given NO frames from the video. Answer the question from the text alone.
Question: {question}
A. ...
B. ...
Reply with ONLY the option letter (or the exact short answer if no options).
Scoring
Rule scorer (video_index/scoring.py): the leading option letter of the reply, a stated answer ("the answer is B"), or the option whose text the reply repeats; a reply that names no option is wrong. Gain = VIDEO - BLIND on the same items.