Device23 · Q optimizer physical-trust rebuild

Blue: tracked-pole FEM Q on the persistent frozen reference connectivity. Gold circles: qualified Tidy3D median Q across three monitors and three fit windows. The old trajectory is rejected because these trends have opposite signs.

Matched canonical checkpoint evaluations. The replacement optimizer is held at step 2 while its eight smooth hole/sidewall directions are measured on fresh meshes.

The superseded trial projection moved hole contours by 1.805 nm at nominal zero step. The corrected line search is bitwise identity at zero, applies only the pure log(Q) gradient, calibrates realized normal motion, and records predicted versus measured gain.

This probe isolated the original failure: the first-order gain was positive, but the saved level-set candidate lost Q after rebuilding the million-DOF mesh.

The fixed-connectivity ALE solve agreed with the local prediction, but this was not sufficient: its smoothed displacement differed from the saved boundary. This led first to the exact-checkpoint remesh gate below, and then to the persistent-reference formulation used by the live optimizer.

Production-fidelity gate decisions versus requested boundary motion. Positive ALE gains cannot promote a saved checkpoint whose paired fresh-remesh gain is negative.

The repaired optimizer freezes trusted step 2 connectivity and represents every later boundary as an absolute ALE deformation. Steps 3 and 4 consecutively improve tracked Q on that common operator family with 94.9% and 94.2% derivative agreement; replay, mode, residual, boundary transfer, feature, and element-distortion gates all pass.