Account for errors that accumulate over time

The bounded native interface now passes all 24 controls, following 58 adapter and 69 protocol checks. It retains parameter limits and unidentified directions. An exact-arithmetic check identified an unstable dense reference; its original failure is preserved. Nonlinear approximation-error bounds and coverage remain separate requirements.

The shared calculation passes 54 source-bound software checks. Its eight-state storage test handles 64,080 generated observation times with 59.2 MiB of stored operator arrays. Small problems agree with independent dense calculations, including all 27 parameter columns, changing time steps and correlations between axes.

A small gyro error can tilt the reconstructed direction of gravity. The resulting position error grows and remains correlated across later times. This component provides a practical way to calculate with a supplied error law while preserving that correlation. It also supports continuous blocks, so processing a recording in sections does not reset its error history.

Small constant-level examples with a perfect initial state and declared white gyro/accelerometer noise match the analytic covariance. The long case checks storage and execution without constructing a dense reference. The aircraft error law and uncertainty from fitted calibration parameters still need establishing. Existing flight fits, accuracy limits and protected recordings are unchanged.

The current qualification receipt binds exact source versions, environment and passing tests. The earlier 26-check and 32-check and 46-check snapshots remain preserved. Shared archive storage retains the versions without copying a library tree. The existing receiver-error operator is unchanged.

Conditional future means and per-epoch uncertainty now match independent dense Gaussian conditioning at four training cutoffs. The calculation uses only an earlier-error checkpoint. Unknown future errors remain random; their effects are retained rather than treating earlier withheld positions as known.

A thirteen-state example now also includes slowly changing gyro and accelerometer bias. It agrees with independent covariance integrals and the noise generator's separate endpoint/integral formulas, including the declared stress amplitudes at 1/300, 1 and 400 seconds. These are generated software checks with fixed parameters.

The joint-noise mode also preserves a query and endpoint driven by the same increment. Its covariance and earlier-only conditional predictions match dense controls; replacing the shared noise with independent errors gives a different answer. The original after-transition mode retains its meaning.

The receiver-error extension passes another 15 source-bound checks. It carries offsets, drift and up to two persistent error components across the training/prediction boundary, preserving changing cross-axis transformations. Small generated examples agree with independent dense calculations and earlier-only predictions. The largest checked state has 27 coordinates. All 24 publication artifact hashes are verified; this adds no flight fitting or instrument-noise estimate.

The native calculation now also passes 38 complete-history checks, following its 18 single-step checks. Their combined path passes twelve more native checks, including every active calibration column and earlier-only conditional prediction.

The linear calibration extension passes 15 checks. It includes uncertainty from learning calibration on earlier observations. It returns no finite prediction covariance when an unknown calibration direction could change the forecast. Dense reference calculations agree for all 27 columns. Its 23 artifact hashes are verified. This assumes a linear mean and unrestricted calibration; native nonlinear fits and parameters at bounds still require qualification.

The ESA rate-point law passes 15 checks against an independent calculation from sample errors. Neighboring intervals share an endpoint sample, so their errors remain correlated after conversion to increments. Point variance and continuous noise PSD remain distinct inputs. All 23 artifact hashes are verified. The native adapter now passes 65 checks within a 182-check source-bound qualification. It retains the larger rate-point state and all 33 states in the full receiver combination. Nominal positions, covariance and calibration columns agree with independent controls. The training-only native fitting adapter now passes 56 checks across all three models and both input formats. It preserves every calibration column and the earlier-only boundary without a dense whole-record covariance. Active calibration bounds and nonlinear coverage remain to qualify.

The prepaid-work extension passes 112 bookkeeping checks. Long calculations can reserve a bounded block of native calls once, retaining the full charge after interruption and enforcing each attempt locally. This removes per-packet durable transactions without changing the numerical calculations. No measured speedup or prediction claim is made.

The sampled-rate error generator passes 54 checks, including an independent covariance calculation and exact earlier-prefix reproducibility. This supplies realistic error sharing for the next nonlinear fitting controls; it does not estimate the actual instrument's noise.

The complete short native protocol now passes 128 software checks. All twelve two-second examples return forecasts within the physical screen, while retaining numerical nonconvergence and unidentified calibration. This qualifies the connected interface, including separate receiver-reading and physical-position targets. Longer noisy fitting and uncertainty coverage remain to test.

The two-minute study now returns all 24 forecasts within the physical limits, with 20 numerical passes. All fits reach calibration bounds, so the uncertainty gate remains open. This is a different duration from the earlier twenty-minute failed controls; it cannot establish that the revised weighting fixes that longer prediction problem.

Frozen software plan · Conditional forecast plan · Changing-bias plan · Shared query/endpoint noise plan · Persistent receiver errors · Estimated calibration uncertainty · Shared rate-point errors · Why gyro error accumulates · Prediction error assumptions · Roadmap.