Clip-Length Effects on Human Evaluation — calibrates the length of → Trajectory Segment

The trajectory segment's headline spec, 1-2 seconds, is not an arbitrary round number: it is the output of the clip-length-effects finding that shorter clips take a rater disproportionately longer to understand relative to the information they convey, so the authors picked the shortest length at which rater evaluation time still scaled roughly linearly with duration (deep-rl-human-prefs, §"3.3 Ablation Studies", p. 10). That calibration produced two different concrete lengths rather than one: 1.5 seconds for the MuJoCo robotics tasks and roughly 1.7 seconds (25 timesteps at 15 fps) for Atari, each tuned separately to its own domain's frame rate and information density (deep-rl-human-prefs, §"A.1 Simulated Robotics Tasks", p. 14; §"A.2 Atari", p. 15). The 'no segments' ablation, which shrinks the segment to a single length-1 frame, is the finding's direct stress test, and it was run only on the robotics tasks because the Atari reward model already depends on a stacked window of consecutive frames rather than one frame (deep-rl-human-prefs, §"3.3 Ablation Studies", p. 9). Collapsing the segment to a single state produced large performance drops, showing that the segment's duration is doing real informational work rather than merely amortizing a fixed per-query overhead.