Joint Z-axis calibration profiles: staged development plan

Prepared scope, not authorized execution. The completed studies are frozen. The saved-result audit found that the revised integration left all six Z gain/offset solutions at the same limits and completed only four fits. Both unfinished zero-field globe fits fail stationarity by more than fifty times. This is not a reason to retry those fits or replace the archive's solver.

An IMU (inertial measurement unit) measures turning and acceleration. Here, gyro gain changes the scale of its measured turning; offset adds a constant correction. A joint profile fixes both corrections to specified values, then fits all other permitted parameters. It can show whether many correction combinations fit almost equally well, or whether a model consistently requires an extreme correction. It does not measure time-varying instrument drift.

First stage: one complete reference, one clock, one processing case

Reuse the full 438-second selected stretch 570b37d23af5e245dd461051 from ILVIS0_applanix_57300_610_LVIS_38614965.013, IMU record type 21. Its median ground speed is 878.06 km/h and minimum is 859.55 km/h. Preserve the saved selection, source hash, integer increments, receiver observations and extraction provenance. Do not select or shorten a stretch using gyro residuals.

Use the original integration and packet-header intervals from the completed six-fit clock reference. This isolates calibration in a method that completed the control; it does not establish the proprietary increment convention. The revised integration is a later sensitivity stage, not the default for these profiles.

First use bias1_unsubtracted: zero upstream field removal, with the original ±1°/hour constant gyro-offset limits and every other physical bound unchanged. These limits remain sensitivity assumptions, not measured per-unit limits.

Run the specified-field globe, zero-field globe and stationary disc at all nine coordinates: 27 identities. Removing two fixed coordinates leaves 24 free parameters in this non-removal case. The same reduced coordinates and bounds must be used for every model. Keep the frozen solver, initialization stages, receiver covariance, mounting transformation and observation metric.

Begin each point at zero in the other normalized coordinates, with its two fixed coordinates already inserted. Do not warm-start from the favored model, a previous grid point or the unrestricted fits. Use the frozen initialization stages. Grid order is gain, then offset, then the three models; order grants no adaptive retries.

The unrestricted clock fits are comparison references, not fixed-profile starts. Even when an unrestricted optimum is at a grid corner, it does not replace the scheduled fit. A profile result worse than that feasible reference at the same corner flags incomplete optimization; it does not invalidate the saved reference.

Finite cost and execution requirements

Request 27 starts × 200 native sweeps = at most 5,400 sweeps, one start per identity, no retries. Count failed and interrupted passes and any final diagnostic predictions. Reserve two final prediction passes within the 200 limit; never add untracked evaluations to obtain a report. The proposed identity ledger is generated by lll.ilvis0_integration_audit.profile_scope() and saved in profile-scope.json. Its authorized-start count is zero.

At the measured roughly 3.8–4.3 seconds per pass, full budgets take approximately 6–7 serial hours. This estimate is not a larger allowance. One worker may use two numerical threads total. Do not run another numerical study concurrently.

Before requesting launch approval, implement and test a separate profile worker. Test fixed-coordinate expansion, unchanged physical decoding, reduced Jacobian columns, feasible-bound stationarity, known-optimum controls and native accounting. Do not edit frozen fitter or kernel sources to freeze the two coordinates. Tests must show that the solver cannot move either fixed value, including at ±1 normalized coordinates, and that its final diagnostics describe only free feasible directions.

Seal sources, tests, this plan, identity ledger, original reference, selection, conversion, compiler/kernel hashes and numerical environment in a new development manifest. Preparation and packaging must not extract or integrate original data. Observed extraction occurs only during an authorized run. Changed freezes require new identities and consent; no unused completed-study budget transfers here.

Hold an exclusive shared lock; journal and fsync each charged start before execution. Save hashed results atomically. A charged identity without a verified result remains an interrupted failure. Resume only uncharged identities in an unchanged unfinished freeze; completion prohibits resume. A detached worker must survive the command sandbox, publish on completion and preserve its journal and failures.

Diagnostics and interpretation

Save cost, receiver-coordinate errors, convergence, stationarity, elapsed time, sweep counts, all fixed/free physical parameters and free active bounds. Capture the full and free gradients, scaled/projected free gradient, singular values and right singular vectors from the already counted final tangent pass. Record vector coordinate order and distinguish fixed Z coordinates from free active bounds. Step-history capture must not change solver behavior; if unavailable, say so.

Plot each model's absolute achieved cost over the feasible gain/offset rectangle, alongside within-model cost differences and the unrestricted solution. Keep separate panels and a common absolute scale for comparisons. Do not normalize away a model's poor absolute fit. Annotate all nine coordinates and convergence status. Unfinished points are upper bounds on their unknown optimized costs and must not form completed contours. A 3 × 3 grid cannot resolve a narrow valley or prove global optimality. Cost contours are not confidence regions before uncertainty calibration.

This stage evaluates one processing case on one recording. It cannot establish robustness across both processing assumptions, clocks, integrations or instruments. Interpret geometry first. Compare background fields only after a complete matched globe result, and describe any such comparison as limited to this diagnostic stage. All scientific decision flags remain abstentions. Neither field model identifies the field's cause or tests the solar/sidereal rate difference.

After review, separately scope the profiled-removal case, nominal clock and revised integration if the first grid is informative. The full two-clock/two-processing 3 × 3 grid would require 108 starts before integration variants. No later stage, recording replication or corpus rerun is implicit in the first-stage plan.