Bounded-method software screen
This compares two actual optimization strategies on 60 small software starts: TRF (the completed trial's method) and dogbox (a rectangular trust-region method). It reads no archival observations and performs no native integrations. Both use the same original parameter box, Jacobian scaling, 600-evaluation ceiling, fresh gradient check and unchanged cost/residual stability checks.
| Known-answer control | TRF | Dogbox |
|---|---|---|
| interior | 3/3 | 3/3 |
| boundary | 3/3 | 3/3 |
| redundant | 3/3 | 3/3 |
| scaled | 3/3 | 3/3 |
| nonlinear | 3/3 | 3/3 |
| gain_offset | 3/3 | 3/3 |
| mixed_interior | 3/3 | 3/3 |
| mixed_boundary | 3/3 | 3/3 |
| coupled_interior | 3/3 | 3/3 |
| coupled_boundary | 2/3 | 3/3 |
| Total | 29/30 | 30/30 |
Each control has three predetermined starts. The last two controls have 27 coupled parameters, a 100-million-to-one singular-value ratio, and a near-boundary start. Known prediction and objective recovery are checked separately from numerical termination. Unique parameters are checked only where the equations identify them. Both strategies call the real SciPy optimizer; temporary dispatch leaves the completed trial's source unchanged.
Failed software checks
trf::coupled_boundary::2: 29 evaluations; known answer passed; final checks failed; linear_correction_not_stable.
What this establishes
This is a screen, not qualification for the airborne archive. The controls demonstrate behavior on specified equations, not the cause of a particular saved flight's stall. Complete observed Jacobians and rejected-step traces were not saved in the recovery trial. A new observed diagnostic needs a separate source freeze and finite allowance. Neither a successful software test nor an unused completed allowance permits replaying the 41 unfinished stretches.
Run screen_methods.py --render-only to regenerate this page from saved results.
Running it without that flag performs this small software screen again; it never
starts archival fits. Input/source/environment hashes are in the screen receipt.