Same sensor channels sliced with 3 different hardcoded .iloc column ranges in one notebook; the mapping dict that would fix it is never used
In a wide-CSV IMU/ biomechanics notebook, sensor features are read with positional slices like imu_data.iloc[:, 54:57] instead of by name. The SAME logical channels are sliced differently in different cells, all off by one from each other:
- left/right foot acc: iloc[:, 54:57] and iloc[:, 103:106] in one cell (gait_enhanced.ipynb:958,960)
- the same foot acc later read as iloc[:, 55:58] / iloc[:, 104:107], plus shank quat iloc[:, 35:39] and shank acc iloc[:, 39:42] (gait_enhanced.ipynb:1277-1283)
- a documented dict maps LT_shank_acc:[38,39,40], LT_foot_acc:[54,55,56], RT_foot_acc:[103,104,105] (gait_enhanced.ipynb:1374-1383)
So the windows actually fed to the model are shifted one column from the named features, and the dict that would prevent this is defined but never referenced (grep shows one hit: its own definition).
Lesson: never duplicate positional column ranges across notebook cells. Define one channel->index map, index only through it, and assert arr.shape[1] == len(expected_channels) right after extraction so a drift fails loudly instead of silently training on the wrong channel.