Focused NASA analysis plan after the October 11 review

The covariance/prediction comparison has finished with 10 of 12 fits converged and an unresolved combined hierarchy. The subsequent numerical recovery is complete: 41 globe preferences, none for flat and 12 unresolved stretches after completed precision recovery. The next priority is bounded continuation of the three new budget failures, followed by separate assessment of the other nine; the roadmap describes that investigation. Additional diagnostics below remain available for later development; preserve their completed evidence.

Keep the 53-stretch first pass unchanged. It found 12 complete globe preferences, no complete flat preference and 41 unfinished comparisons. Unfinished comparisons provide no positive flat-Earth result. More completed fits still require separate adequacy and uncertainty checks before calibrated scientific claims.

An IMU (inertial measurement unit) measures turning and acceleration. These NASA records come from survey-grade airborne equipment. Exact device identification remains parallel, lower-priority work; numerical and physical sensitivity tests can proceed using the decoded measurements and stated assumptions.

Evidence already available

The review identifies relevant weaknesses but does not cover several newer completed tests. Those tests answer narrower questions than full scientific validation.

The verified-stopping trial completed five of six fits and did not qualify. Its saved failure audit identifies a bounded linear residual-step probe 1.55 times the stability limit despite tiny cost change and gradient. The six-start allowance is spent; replication remains deferred. This document grants no additional starts or corpus rerun.

1. Investigate clock and joint calibration

The two-recording clock audit is complete, with no observed fits or native predictions. UTC tags accumulate 9.48 and 9.65 ms more than nominal 200 Hz timing over the guarded stretches, about 22 ppm. The POS elapsed-time tags also differ from nominal and from UTC. Neither establishes the increment integration period by itself. Five independent physical controls passed, including a gyro gain/offset/clock rank check. The actual timestamp grids can produce about 20 m of synthetic stationary height error under an incorrect fixed-period assumption; a tiny gravity adjustment removes most of that control's error. The actual fitted gravity changes are larger and opposite in sign, so the full three-dimensional compensation path remains unexplained. Preserve this distinction when planning profiles.

The coupled three-dimensional controls are now complete. Independently integrated force/angle increments recover known motion, and the Python strapdown primitive agrees with a separate reconstruction. Interior and near-boundary calibration references recover under QR and SVD linear solves. A coordinated clock/calibration change preserves packet-endpoint positions but is distinguishable when position observations have independent physical time tags. These controls use a Cartesian frame with zero background/transport terms and known reference attitude. They do not yet qualify the native geographic fitter, its receiver association or nonlinear optimization. The separate geographic controls extend that check through the actual native value/tangent kernels and nonlinear solvers. They use independent increments and off-packet receiver epochs for all three geographic models, with six free calibration, initialization and latency parameters. All 24 fits passed convergence/prediction checks, with track errors below 6 micrometres. The saved calibration audit nevertheless finds a weak Z gyro-offset direction: interior offset errors reach 0.647°/hour, and all 12 near-boundary fits reach the offset limit. The interval and geometry comparison now separates these effects: all nine new fits passed; halving the interval reduces offset error roughly fourfold, and added roll substantially improves recovery at the same fine interval. This supports the proposed mechanism in the controls, without identifying the NASA boundary cause. The completed original-rate coning/sculling controls now recover the known offset within 0.000011°/hour for both globe truths and 0.000377°/hour for the disc truth, with both corrections enabled. All nine new fits passed, without changing the original solver or limits. Calibration conditioning remains weak. The latitude/convention controls now pass all 24 fits at ±45°, including opposed starts and roll/unequal timing. Offset errors stay below 0.00079°/hour and no fit reaches a calibration limit. Already-compensated controls demonstrate the error from applying corrections twice. The six-fit observed integration comparison is complete: four fits converged, while both zero-field globe fits exhausted their budgets. The saved-result audit finds their stationarity errors remain 53–56 times the limit. Z gain and offset remain at exactly the original limits in all six fits. The alternative does not qualify for archive rollout. The full-column derivative check passed; it does not qualify global optimization. Do not assume the archive lacks upstream compensation. Good predictions do not qualify unique calibration. These short-record tests do not qualify all 27 archival nuisance parameters or authorize observed profiles.

Start with a timestamp audit on the already selected recordings. Preserve packet order, configuration, integer increments and original timestamps. Record both time tags and their declared types, quantization, nominal cadence, adjacent differences, discontinuities and cumulative departure from the declared sample period. Split incompatible configurations and time discontinuities; never repair science measurements by interpolation.

A packet timestamp is not automatically its integration duration. Generate small, independent controls with known increment periods and separately varied timestamp jitter or clock scale. Establish what either integration convention recovers. Where Group1 engineering channels help check timing, label them fused navigation and use them only for decoder validation. Investigate the vertical effect through accumulated velocity, gravity correction and timing/calibration sensitivity. A smaller rotating-globe cost alone cannot identify the clock convention.

The first joint grid is complete: 27/27 fits converged, using 4,174 native sweeps. Under original integration, packet-header timing and no upstream field removal, all nine Z gain/offset pairs favor globe, then the specified field. The preferred model's cost varies only 2.58%; horizontal error stays 0.081–0.093 m. Its remaining boundary count is 0–1, versus 12–13 for zero-field globe and 13 for disc. The weakest remaining direction is dominated by gravity correction and X/Y gyro gains. This demonstrates robustness across tested Z corrections, not unique calibration.

Two approved, separately frozen follow-ups use the same grid: profiled upstream removal, then the earliest predeclared other recording. Both have finished: removal converged 27/27, while the other recording converged 25/27. Seven complete nodes favor globe; two remain unresolved. Their allowances are spent and no retries are authorized. A saved-only audit now examines their calibration behavior and stopping failures. The second is a whole-stretch robustness check, not the deferred training/prediction replication. Further clocks, the second predeclared recording and archive rollout remain outside this allowance.

The completed saved-only audit preserves all 81 grid fits and their original statuses. Both unfinished fits pass saved stationarity but lack solver-reported success at their budget ceiling. Preferred control fits keep near-zero removal and almost unchanged costs. Epoch files reproduce all saved costs and RMS values. The completed covariance audit adds longer-lag, cross-axis and spectral diagnostics; these saved checks do not calibrate uncertainty or authorize retries. At each fixed gain/offset pair, reoptimize all remaining parameters with their existing physical bounds. Use identical grid coordinates, initialization policy and finite allowances for specified-field globe, zero-field globe and stationary disc. Later stages compare clocks and processing assumptions. Reuse only compatible completed reference points with verified provenance. Save the absolute cost and within-model cost increase, stationarity, completion, remaining boundary values and prediction errors.

Show each model's feasible region, fitted points and profile contours. Mark unfinished points explicitly: their achieved costs are upper bounds on the unknown minimum, not completed contours. A coarse grid cannot demonstrate a narrow minimum. Describe movement of valleys across clocks and whether preference persists along admissible calibration alternatives. Do not convert cost contours into confidence regions before uncertainty calibration, or interpret cross-model relative contours as absolute model ranking.

Cost must be explicit before launch. Even a 3 × 3 grid across three models, two clocks and two processing cases creates 108 fit identities before any permitted reference reuse. It is not a cheap six-fit follow-up. Choose a staged grid and solver only after reviewing the failed qualification and testing the numerical method, publish the exact identity ledger and request a finite allowance. No adaptive restarts or expansion are implicit.

Success is an explained timing/calibration interaction and a stable complete comparison under defensible alternatives. Persistent ambiguity remains a reported result; it is not permission to add flexible drift until the fit looks better.

2. Establish a shared residual-error model

The completed saved-only audit reports dense lag curves, cross-axis covariance and correlation, and low-frequency power in physical errors and declared innovations across all 81 grid fits. It retains sample counts, actual lag durations, cadence and the limited frequency resolution of these two short recordings. Irregular timestamps receive no FFT; measurements are never resampled. Physical track errors use a common reference metric across models. The control's vertical innovations retain substantial minute-scale dependence, with less slow dependence on the other date. Neither the objective nor rankings change.

Relate those diagnostics to available receiver quality and precision fields. If a field is absent, report it absent. Fix quality is not a measured noise variance. Fitted residuals combine receiver error, inertial error, parameter estimation and model mismatch; they do not uniquely estimate GPS covariance. Very small whitened RMS does not by itself establish exceptional sensor accuracy or a calibrated likelihood.

The shared error-model plan is prepared. Its 35 independent software checks now pass: positive-definite covariance, cross-axis transformations, whitening and train/prediction conditioning, including irregular times and fixed drift anchors. These known-mean/known-covariance controls do not include fitted-parameter uncertainty; The training/prediction extension now passes 26 controls for training-only covariance selection, linear-mean uncertainty and sensitivity to all supplied covariances. Nonlinear calibration, active bounds and integrated covariance-parameter uncertainty remain outside these controls; specify their treatment before any observed comparison. The nonlinear geometry checks now retain active columns, report nonlinear remainders and deficient rank, and keep disconnected sampled support. Their boundary-distribution control demonstrates why interior Gaussian covariance and likelihood thresholds cannot be transferred unchanged to bounded estimators. Finite prediction ranges are sensitivity checks, not calibrated intervals. The native adapter controls now verify all 27 derivative columns on six independently constructed geographic paths. Training rank uses training predictions only. The finite comparison preparation specifies 12 matched starts and at most 3,060 native sweeps, including all nonlinear diagnostics. The durable runner now passes 39 targeted controls and has its own launch freeze. A separate finite observed-data allowance was approved and the comparison has completed: 10 of 12 fits converged, with the combined hierarchy unresolved. Preserve the original comparison. A separately versioned correlated-error objective remains a development alternative. Estimate its parameters on designated development or training observations and apply the same sealed covariance specification to all physical models. Compare direct prediction errors as well as covariance-adjusted costs. Account for cross-boundary correlation, fitted-parameter uncertainty and active bounds in future predictive intervals. Retain the original objective and its findings unchanged. An absolute-fit threshold requires independent calibration covering the shape hierarchy and selection procedure.

3. Identify the measurement contribution

Group 10001 is the primary-receiver stream, separately recorded from Group 1 fused navigation, as documented by the interface specification and source checks. The planned versioned provenance supplement will retain receiver-type codes, original NMEA identifiers, stream hashes and historical firmware/layout evidence. The active freeze remains unchanged. Separate streams do not establish independence from GPS coordinate conventions or statistical independence of measurement errors.

Specify matched controls for gyro increments, acceleration increments, gravity and receiver-coordinate metric. State exactly which inputs each alternative uses in its objective, initialization and motion constraints. Removing a channel from the objective while retaining it elsewhere does not isolate its contribution.

A receiver-constrained angular-increment diagnostic should compare observed Group4 gyros with each model's explicit trajectory-based prediction. Declare all attitude, mounting, timing and gravity assumptions. Do not replace absent gyro observations with Group1 angular rates or differentiated fused attitude. Unknown body attitude or yaw can leave such a diagnostic unidentifiable; report that rather than fabricate an independent test.

Include non-gyro controls with comparable nuisance freedom to ask whether coordinate conventions and acceleration alone reproduce the separation. A flexible receiver-only interpolator is not a fair model-comparison control. Report physical track errors in a common metric, objective errors in their declared covariance, and rank after each ablation. Horizontal-only, altitude-only and height/geoid variants remain useful, with covariance transformed consistently. Receiver coordinates still originate in a geodetic navigation system; these controls isolate dependence rather than create raw GNSS range evidence.

Success is a reproducible explanation of which measured combinations supply separation. Interacting channels need not admit a single percentage of evidence attributed to gyros.

4. Qualify optima, prediction and independent replication

Extend existing controls to longer independently generated physical records, every model, both clocks, coupled calibration errors, exact/near-boundary optima and poor starting conditions. Compare at least two numerical strategies within the same feasible region. Require objective recovery, predicted-observation agreement, repeatability and honest treatment of unidentifiable combinations. Recovery of individual parameters is required only where the control identifies them. A higher completion count alone is insufficient; a converged poor flat fit and an unfinished flat fit are different outcomes.

Any larger-budget sensitivity gets a new freeze and allowance; the original 200-evaluation records remain unchanged. Software controls do not establish full archival qualification.

Maintain a ledger of source hashes, aliases, UTC overlap, flight/date, platform and supported physical instrument/installation continuity. Configuration bytes alone do not identify a physical IMU. Overlapping streams and same-flight intervals are dependent observations. Unknown identities remain explicit and conservatively grouped.

Use the already predeclared other recordings for the next prediction replication after qualification and separate approval. Fit nuisance parameters on training observations only, propagate original increments without receiver resets, and report all failures. The completed control split is retrospective because its full recording was examined earlier. Different recording files are not automatically independent installations. Shared calibration across flights requires supported continuity and equal treatment of all candidate models.

5. Rotation-field work follows these checks

Before estimating the free field, examine the actual latitude/heading coverage and rank of north amplitude, east component and vertical amplitude against supported installation biases and upstream-removal assumptions. If the field and removal are indistinguishable, report their retained combination. More time samples or more files do not automatically remove that ambiguity. The rotation-field plan retains the later solar-versus-sidereal magnitude test; gain, clock scale and common systematic errors need precision appropriate to that much smaller separation.

The public milestone is a finished matched comparison with explained calibration, credible predictive adequacy and robustness, followed by supported independent replication. The survey-grade archive remains the primary project. Consumer stationary measurements remain a separate developing experiment and inherit none of its thresholds.