Known-calibration noise benchmark: construction checked
27 checks pass, including both native input-error formats. On two-second generated records, the streaming mean matches the qualified wide calculation exactly. Its maximum difference from independent trajectory truth is under one micrometre. This is numerical software agreement, not instrument accuracy.
The calculation conditions on six earlier epochs and summarizes the later physical-position covariance. The independent dense comparison agrees. Receiver observation loadings stop at the earlier boundary, and the expected-squared-RMS summary remains distinct from an expected realized RMS.
Initial state, calibration and nominal future geometry are supplied as an explicit oracle assumption. This benchmark cannot enter the real training-only prediction pipeline. It measures expected error under declared laws, without a fit, random draw or original/protected data read. The longer four-setting benchmark needs its own finite freeze; it has not run.