Why the first trust-region candidate failed qualification

The original qualification tests are retained as historical evidence, with their exact failed-study source hash. They are separate from the current regression matrix; the original four failures remain recorded.

The failed candidate missed the same known-answer control from all four starts. The control has 27 parameters, an invertible monotone nonlinear transform and a linear map with singular values from 10,000 to 0.0001. The known feasible optimum has objective 1. Weak parameter directions make it a demanding numerical test; it is not an archival flight or an Earth-model simulation campaign.

Four separately bounded diagnostic replays reproduced the failures. Every start used 199 of its 200 allowed evaluations and stopped without passing stationarity. Final objectives ranged from 1.00005157 to 1.00076924. The solver was still accepting small corrections, with late normalized step lengths around 0.0037–0.0121. Its damping and trust radius could remain unchanged when actual improvement was between one quarter and three quarters of the local prediction. Progress along weak directions was therefore repeatedly restricted. The preserved replay measurements include every history and endpoint.

Removing routine damping improved an initial prototype from 44/48 to 46/48, but two starts still exhausted their allowance. A short box-constrained correction can point in a different direction from the full correction in a curved, poorly conditioned valley. That prototype and its failed screen remain preserved.

The final revised candidate first tests the full feasible bounded Gauss–Newton correction. Only a rejected full correction invokes the smaller trust-region proposal; damping is a fallback for an unfinished local solve. Every evaluated proposal is counted. This retains the established full-step path where it works and supplies a different bounded direction when it does not.

The separately frozen revised qualification passed all 48 original checks, including six independently constructed raw-increment motion controls and two negative controls. Physical models, parameter bounds and the final undamped stationarity, cost and actual-position checks are unchanged. Small trust-region steps cannot establish convergence by themselves.

This explains the observed software-control failure; it does not establish that the same mechanism caused every unfinished archival fit. The separate matched archival trial tests that transfer.