Ten-second motion traces gave 32.45% balanced film accuracy in 68 NNDb adults from nine films when each test person was excluded from training

A fixed template rule used 29 minutes of first-run motion estimates to predict the viewed film in 68 NNDb adults from nine films. It used ten-second bins of log framewise displacement. It removed the trace mean and a linear time trend, then set each trace to unit norm. Each test person was excluded from all training templates. Balanced accuracy was 32.45%, compared with an 11.11% equal-film reference. The rule predicted 24 of 68 labels correctly. Both 9,999-draw reference tests gave p = 0.0003. These tests shuffled film labels or independently shifted traces in time. The approximate person bootstrap interval was 12.64% to 41.71%. Three films had zero recall. Thirty-second bins gave 11.90% balanced accuracy. An earlier correlation study used the same people. Film assignment was not fully random. The result does not establish a causal film effect or neural decoding bias. The package preserves 175 fixed source files, code, exact random draws, all planned results, figures, and reproduction steps. Independent direct calculations passed. A fresh source download and full repeat gave identical bytes in all 19 result files.

head motion · framewise displacement · film fMRI · leave-one-person-out prediction · balanced accuracy · naturalistic neuroscience