Bounded calibration: what passed and what remains

The study completed its full frozen allowance of 228 mathematical starts using 40,236 residual/Jacobian calls in 14.39 seconds. There were no observed fits, original-file reads or native integrations. Every failure is preserved; the study will not restart or spend its unused allowance.

Known solutions passed; the full matrix did not qualify

All 12 known-answer starts passed, including exact-boundary, near-boundary, interior and rank-deficient controls. Analytic derivative and fresh endpoint checks passed their unit tests before the freeze.

In the matched constructed-history matrix, 136/216 endpoints passed, but only 32/72 three-start comparisons passed every required check:

All 80 unsuccessful endpoints exhausted the 200-evaluation allowance. Their residual RMS values were approximately 1.15 × 10⁻¹⁰ to 3.50 × 10⁻⁷ m/s². Most failures therefore occur where almost nothing remains to fit, not where a physical candidate visibly misses the constructed force magnitude.

Of those 80 budget failures, 53 satisfy the separately saved fresh stationarity limit. They remain failures: the frozen rule also requires solver completion. Saved diagnostics do not retroactively replace that rule. The other 27 failures also miss the fresh stationarity limit. Successful optimization cannot be inferred merely from a small scalar residual.

Changing motion can preserve a scalar difference

All 24 wrong-model changing-direction/climbing comparisons passed their three-start checks. Wrong-model RMS spans about 0.00551–0.0131 m/s² in the direction-changing controls and 0.000137–0.00215 m/s² in the climbing controls. These are feasible finite numerical comparisons, not globally certified optima, noise-normalized evidence or results from the aircraft archive.

The separate exact constant-gain construction already demonstrates straight-leg ambiguity. The zero complete-comparison counts above do not create a new stationary/eastbound model finding; they show that this numerical method is not qualified across both informative and nearly uninformative inputs.

Consequence for the roadmap

C02 remains open; C04 observed channel fitting is blocked. A follow-up needs a separately identified candidate that handles near-zero residuals and flat calibration directions honestly, with unchanged bounds and repeat acceptance. No larger observed budget, relaxed stopping check or retry of this study follows.

Independent C03 quantization/timing controls and metadata work may continue. The main result stays 53 globe / zero flat / zero unresolved; the separate IMU21 horizontal-only sensitivity remains unresolved. None of these mathematical controls changes an airborne disposition.

Every start · Completed report · Frozen scope · Authoritative backlog.