Working assumptions for ESA receiver errors

The receiver export reports uncertainty in north, east and height, including the correlations between those axes. These values help weight a fit; they are not an independent measurement of the receiver's true error.

The completed position assessment found residual correlation after fitting local trends. That supports testing persistent errors, but does not determine their exact size or duration. The following two cases are fixed research assumptions. Neither provides calibrated confidence or a detection threshold.

The floor is added in quadrature to the reported covariance. Scaling applies to the full matrix, preserving its cross-axis terms. Five percent of the working variance remains independent between observations. The rest follows the stated exponentially decaying correlation. The time-varying loading uses each working matrix's lower Cholesky factor, followed by the common receiver-coordinate conversion. This fixes the cross-time interpretation as well as the within-epoch variances.

The numbers above are chosen sensitivity cases, not sensor specifications, estimated error bars or guaranteed bounds. They are set before examining any ESA model preference. A common preferred geometry across these cases would show robustness to these choices, not establish statistical significance.

Frames and data boundaries

• Retain the receiver's full north/east/up covariance. The three cross terms are north–east, north–height and east–height; height is upward.
• Reject nonfinite or nonpositive-definite reported covariance. Do not repair it by adding the floor or clipping eigenvalues.
• Use the same receiver-coordinate weight for every physical candidate. The conversion to latitude, longitude and ellipsoidal height uses the receiver's declared WGS84 chart. It is an observation convention, not new curvature evidence.
• Use actual, strictly increasing timestamps within one GPS week. Do not interpolate measurements or fill gaps. Interval selection handles gaps separately.
• Construct training weights from training rows only. Gyro values, fitted residuals, later positions and preferred-model labels are not policy inputs.

Implementation and verification

The shared lll.airquantumgrav_error_policy adapter uses the existing lll.ilvis0_error_model and lll.ilvis0_structured_covariance components. Policy identifiers and numerical values are versioned. Receiver measurements are preserved, with a hash of the fields used by the policy.

The passing software tests check full cross-axis handling, height sign, time ordering, invalid covariance rejection before floors, unchanged input rows and independence from model/outcome metadata. They compare small dense covariance matrices against the structured operator for residuals and all 27 derivative columns. A dense request specifies an epoch cap, at most 2,048, and fails before allocation when exceeded. Long records use the structured operator.

The 76 software checks have saved source, environment and test-report bindings. The policies are now used in the completed first pilot and the completed gravity revision. The wider-precision revision completes three adequate matched comparisons under the same policies. Their use does not establish that either covariance describes the true receiver error or the complete uncertainty of an inertial prediction. No protected confirmation data is released by these software checks or pilots.

The completed known-noise test retains these weights while adding generated receiver and IMU errors. All 24 fits converge, but none passes the physical accuracy screen. Receiver-only weights are not a qualified complete prediction-error model under those stress settings. The weights and failed outcomes remain unchanged; full sensor-error propagation and fitted-calibration uncertainty are the next gates.

Software verification.

Receiver assessment · Training-only controls · Roadmap.