Deterministic multi-start software qualification

The horizontal-only IMU21 diagnosis finds differing locally stationary solutions, not a gross timing-derivative fault. The next candidate retains the existing ridge-probe solver and adds a fixed timing-start list. No score or physical stopping threshold changes. Selection considers every attempt: if the lowest returned cost belongs to an unfinished fit, the multi-start result is unfinished.

Known answers

Use four bounded, four-column constructed least-squares systems. Their timing cost is an asymmetric double well with a known algebraic minimum; a curved calibration coordinate tracks timing. Include both well orientations, a true timing optimum at the bound, and an exactly invisible nuisance coordinate. Another coordinate has its unconstrained optimum outside the box. All columns remain present, including the zero column. Expected minima are calculated from polynomial stationary points and box endpoints, independently of the optimizer.

For each system use two independently fixed non-timing seeds, each with timing starts −0.8, 0, +0.8: 24 starts /4,800 residual-tangent calls maximum, 200 calls per start, four numerical threads, a 120-second wall ceiling. These are cheap constructed functions, not synthetic flights, native airborne predictions or observed-data fits. Charge calls and starts before execution, freeze sources/environment, preserve all outcomes, and never retry identities.

Qualification

Both selected solutions per system must pass the unchanged fresh numerical gate, agree with the independent best cost within 1e−8 relative to max(cost,1), and agree with the known predicted residuals within 1e−5 absolute. The two best solutions must also agree within the existing 1 mm axis-RMS /1 cm epoch-maximum prediction limits. Individual local-well solutions and failures remain visible. Do not require unique values of the invisible nuisance parameter.

A pass qualifies this start-selection mechanism only on these controls. It does not establish global optimality, likelihood calibration or archive-wide performance. Any observed test needs a separate source/input/environment freeze, fixed starts and a finite allowance. The original 53-stretch findings stay fixed.

Separate compatible receipt freeze. Earlier preparations stopped before any function evaluations; preserve their snapshots and charged identities. The ledger uses its existing tangent charge kind for constructed derivatives; results explicitly report zero native calls. No earlier allowance is reused.