How much prediction error should the declared noise produce?

The completed two-minute controls predict later positions accurately, while the earlier twenty-minute noisy study misses its unchanged accuracy screen. The analytic example already shows how gyro noise can produce substantial position error. Do not repeat that example or assume more fitting can remove unpredictable later noise.

The next benchmark uses the declared motion and error laws with known initial state, calibration and nominal trajectory. It computes conditional covariance after the earlier receiver block, then summarizes expected later squared position error. No fit, random draw or observed recording is needed. This gives the error calculation more information than an actual estimator has; it is a benchmark under the stated linear Gaussian law, not a universal bound for every physical estimator or a measurement of actual instrument noise.

Preselect four generated settings: recorded-increment versus sampled-rate errors, each over 120 and 1,200 seconds. Use the existing northern, fast-turning, climbing specified-field trajectory and the existing short-persistence receiver policy. This is not an Earth-model comparison. The known nominal future geometry may set covariance transforms and metric scales in this oracle benchmark; it must never enter the training-only prediction protocol or its fit selection.

Each record is 300 Hz with one query per second, starting at 0.5 seconds: 119 or 1,200 queries, matching the earlier short and long grids. The first two thirds condition the error state; future receiver error loadings are removed for the physical-position target. Keep shared rate endpoints, changing sensor bias and receiver-state dependence.

Report the square root of the expected squared RMS, per-axis endpoint standard deviations and full marginal-error curves. This is not the expected value of a realized RMS. Compare descriptively with the fixed 10 m horizontal / 30 m height RMS screen; do not change that screen or count epochs independently.

First qualify the summary against independent dense calculations. Then qualify the native benchmark construction on short generated cases before a separate finite four-setting launch. That launch must freeze sources, environment, exact construction/native limits and its own worker allocation. No long native work is released merely by this plan.

The benchmark does not establish nonlinear calibration bounds or calibrated coverage. Its purpose is to separate expected noise growth from optimizer error before choosing another fitting study. Earlier failures and the 53-stretch findings remain unchanged; protected ESA groups remain untouched.