Matched horizontal-only multi-start comparison

Purpose and unchanged observations

The first two-recording sensitivity study leaves IMU21 unresolved at its repeat gate. Fixed-point diagnosis finds coupled timing/calibration solutions, with no gross timing-derivative fault. The separate known-answer multi-start controls must pass before this study can launch. This is not a new corpus selection.

Use the exact preselected 4.07-minute IMU21 recording and original measurements. Retain both IMU channels, nominal 200 Hz integration, the existing horizontal covariance, all nuisance columns, ±1°/hour offset bounds, both processing cases, and unchanged fresh stationarity/cost/position stopping limits. The earlier IMU8 comparison and 53-stretch full-position findings remain preserved.

Starts and selection

Give every model and processing case the same three normalized timing starts: −0.8, 0, +0.8, corresponding to −120, 0 and +120 ms. Use two independent non-timing seed families: the original full-position fit, and the original receiver/IMU initialization used by the first study. Do not copy the new primary solution into its repeat. The grid is identical to the qualified software grid.

This gives 3 models ×2 cases ×2 seed families ×3 timing starts = 36 target starts. Add two fresh IMU8 specified-field/unsubtracted controls, initialized at the earlier matched horizontal primary and repeat endpoints. Eight separately charged projected residual/Jacobian checks must reproduce the parent arrays exactly before fitting. Controls must reproduce the earlier score/predictions and pass their own repeat gate.

Each family selects its lowest returned unmodified cost, with deterministic identity tie-breaking. Preserve every start and failure. A lower-cost unfinished fit cannot be ignored in favor of a higher-cost converged result; a missing or invalid outcome makes that group unresolved. Selected solutions must separately pass the original numerical gate and the unchanged 1e−8 relative cost /1 mm axis-RMS /1 cm epoch-maximum repeat checks. Only complete case/model comparisons enter the shape-first hierarchy. This search is not a global-optimum proof.

Finite allowance and runtime

• 38 optimizer starts, at most 200 tangent calls per start.
• One full prediction diagnostic per start, plus eight preflight calls.
• 7,646 native calls maximum, four total numerical threads.
• 20-hour wall ceiling, no retries, continuation or expansion.

The preceding IMU21 fits took about 1.93 hours for twelve starts. Tripling those starts suggests roughly six hours, but convergence paths may differ. At the measured slowest IMU21 call rate, the full allowance is about 18.5 hours. The longer IMU8 controls are warm-started; the wall ceiling applies regardless.

Freeze sources, inputs, environment, original containers, start vectors and qualification receipts before launch. Use exclusive locks and append-only pre-call charges; save atomic hashed fit, tangent and full-prediction artifacts. An interrupted start stays charged and cannot restart. Completion closes unused allowance. Publication uses saved artifacts only and runs automatically.

Limits of interpretation

Large unscored height errors and flat vertical null directions remain visible. Even a completed horizontal comparison is not adequate full-flight reconstruction, gyro-only evidence, coordinate independence or calibrated significance. Existing shape and retained-background-field names do not identify the field's cause.