Native regression of feasible linear initialization
The revised optimizer passes 36 known-answer fits, eight direct probes and nine starting-point comparisons. This separate regression now tests its complete generated IMU-to-prediction pipeline with the native integrator.
Fixed cases
Use both original prescribed starts for these exposed development cases:
| Seed | Input | Upstream-removal parameter | Earlier role |
|---|---|---|---|
| 281474977010692 | Point | Fitted | Earlier successful control |
| 281474977010693 | Point | Fitted | Earlier numerical failure/regression |
| 281474977210667 | Increment | Absent | Latest calibration failure |
All three cases use the specified-field globe generating candidate. Each fits its own generating truth; these are not Earth-model votes. Preserve the old fit labels. Require exact replay of all six original input, training-observation and error-operator bindings. Keep 27 nuisance columns for the first two cases and all 26 for the last. No future prediction error is used to select a start.
Fresh finite allowance
Freeze six starts, at most 200 fitting evaluations and 620 linear operations per fit, zero retries, four workers with one numerical thread each, and a 5,400-second cooperative ceiling. The unchanged shared formulas give maximum charges of 359,724 native calls and 2,259,076,518 packet visits. These are prepayment limits, not promises to consume the allowance. Estimated runtime is about 15–20 minutes; the earlier four-fit regression took 10.7 minutes.
The shared worker and auditor receive an explicit six-reference count; their historical default remains four. Bind the new analytical gate's exact hashes, all seven numerical source files and the environment. Record source changes against the latest closed calibration: the qualified linear solver and the necessary reference-inventory worker/auditor changes only. Physical/noise source changes must fail preparation. Qualify the interface and source bindings before launch. Preserve every failure and interrupted charge; never reset a budget.
Require six complete numerical/physical passes, three agreeing start pairs, exact original-input replays and independently checked timing-cell evidence. Publish every outcome, verify hashes and archived replay, then compact working copies through the shared store. No original aircraft or protected ESA data.
A complete native-model matrix remains a subsequent gate before a separately budgeted calibration. Neither this trial nor analytical success reopens an old campaign or supplies an uncertainty threshold.