Short noisy native protocol checks

This gate connects the generated inputs, actual nonlinear fitter, continuous forecast and earlier-only physical-position error calculation. It does not establish long-duration prediction accuracy or statistical coverage.

Use twelve cases: three generating models, two processing assumptions and two input laws. Each uses the independently generated two-second, 300 Hz turning and climbing trajectory, nine queries starting at 0.1 seconds every 0.2 seconds, six earlier receiver rows, one declared midpoint start and eight fitting evaluations at most. Preserve all 26/27 calibration columns. Use the existing reported-short receiver policy and the explicit frozen noise settings in the shared input constructor. Each model predicts its own generating truth; these are software controls, not votes between Earth models.

Both gravity and covariance anchors use earlier noisy observations only. Keep weighting fixed at the declared normalized midpoint. Save the fitted point, raw forecast and error-conditioned forecast before generating later truth. Future physical-position errors omit future receiver-error loadings; the same earlier covariance prefix must be retained. No future receiver coordinates or quality values may choose an error transform. Hold the last earlier transform.

Check complete accounting, unchanged measured arrays, before-cast residual binding, native compilation provenance, Gaussian normalization, retained columns, explicit calibration rank/bounds abstention and both raw and conditional scores. Keep numerical convergence and the existing 10 m / 30 m accuracy screen as reported outcomes. Eight evaluations and six training rows cannot guarantee a unique calibration; a truthful unresolved or rank-deficient outcome is valid for this interface gate and cannot release observed inference.

Separate finite allowance

Freeze exact sources, environment and test identities before execution. Per case prepay at most 1,000 native calls, 10,000 conservative packet steps and 30,000 packet-query visits. Reserve two builds separately. Totals: twelve starts, 12,000 native calls, 120,000 packet steps, 360,000 packet-query visits and a cooperative five-minute ceiling on one numerical thread. No retries, budget transfers, original reads or protected reads occur.

Generated construction uses 600 intervals and at most 616 generated samples per case, including truth used only for later scoring. The increment law uses 10,848 normal variates; the point law uses 7,254. All are reserved before draws. The twelve cases therefore permit 108,612 variates. Pure interface regressions have their own finite fixtures and make no native calls.

Only a verified interface result permits freezing the next long-duration noisy fitting study. Its physical prediction and uncertainty criteria remain separate.

Nonlinear qualification · Prediction targets · Shared input generator.