Finite training-only receiver-reference assessment

Assess the first 292 one-second receiver epochs of the previously studied IMU21 recording ILVIS0_applanix_57300_610_LVIS_38614965.013, segment 570b37d23af5e245dd461051. Their 291-second span precedes the withheld positions in the saved prediction comparison. This is retrospective development of a common error source, not a new Earth-model comparison.

Verify the original publication receipt before extracting the prefix. Freeze only those training coordinates, the source hash and the parent-selection hash. The assessment API cannot receive later coordinates, gyros, fused navigation, Earth-model residuals or candidate identities. Test that changing every later coordinate leaves its input unchanged.

Use the fixed affine chart north = 6371000 × latitude_rad, east = 3185500 × longitude_rad, and reported height in metres. The longitude scale is fixed at cos(60°)=1/2, not estimated from noisy latitude. This is a reference chart, not the older receiver-dependent geographic error metric. Exact residuals retain the supplied binary receiver-coordinate values.

Report all six predeclared combinations: time-polynomial degrees 3, 5, 7 and discrete OU correlation families approximately 10 and 300 seconds. Three axes are independent under the declared law. White standard deviations are 0.3/0.3/0.7 chart metres; OU, constant-offset and training-anchored drift standard deviations are each 1/1/3. Their common multiplier is bounded from training data. OU transitions are rounded to the nearest 24-bit dyadic number. Coefficients are ordinary numerical least-squares references; their residual quadratic forms and all tail/endpoint certificates use the qualified exact calculation.

Every bound uses per-assumed-law failure probability 0.01. Report all combinations without selecting a covariance using model ranking. No simultaneous coverage or empirical Gaussian-law verification is claimed. A bad reference trajectory can inflate a conservative upper bound, so it is not a measurement of receiver accuracy or a reason to enlarge a candidate's own error allowance.

Freeze six identities and at most 5,256 scalar innovation epochs, 131,072-bit rational numerator/denominator limits, ten cooperative seconds per axis, thirty seconds per identity and 240 seconds overall. One worker reserves four numerical threads. Charge each identity and its 876-epoch reservation before calculation. Zero native evaluations, Earth-model optimizer starts, original IMU reads or protected ESA evaluations. No retries, extensions or process signals.

Acceptance means every exact calculation completes within those limits with a nonnegative quadratic form and a finite certified endpoint. It does not mean the declared covariance is physically adequate. Preserve failures and stop the sequence if qualification fails. Broad nonlinear prediction ranges and absolute predictive adequacy remain separate gates; no observed Earth-model fit follows.

Qualified mathematics · Saved prediction evidence · Roadmap.