Evaluation-only OXTS motion attribution

Pre-registered2026-10-30. All36 archived camera-only metric motion estimates, including two failed internal screens. No motion refits or new selection masks. Build an evaluation proxy, explicitly NOT independent motion ground truth: T_camera2_imu = shift(P_rect_02) * R_rect_00 * T_cam0_velo * T_velo_imu. Require camera2_velo to match the existing calibration reference within1e-8. This consumes the target extrinsic only after estimation, never in selection.

Use all available pinned OXTS packets within±2 indices of each18 centers, absolute IMU timestamps and local Mercator world poses. Interpolate translation linearly and orientation by SLERP at original camera timestamps; no extrapolation. All relative world origins use the same first available packet per drive. Relative proxy target-from-source = C_I * inverse(W_I_second) * W_I_first * inverse(C_I). Report rotation and translation disagreement and both translation magnitudes.

Fixed diagnostic variants: common camera clock shift -10,0,+10ms, same-index OXTS instead of time interpolation, and omitted camera/IMU lever arm (rotation retained). Record proxy-change magnitudes and estimate disagreements. These are sensitivity scenarios, NOT measured uncertainty bounds or a search for a correcting timestamp. Do not pick a favorable variant or modify any output based on reference agreement.

All36 pairs remain denominator. Test interpolation at integer nanosecond stamps, no-extrapolation, transform direction and rotational lever-arm effect. Pin packet, time, calibration and archive/source hashes before running. OXTS/target calibration errors and world-pose interpolation remain limitations. No Gaussian covariance, independent calibration accuracy, ICP adoption or SOTA claim follows. No downloads.