Verification
Four things that can be proved without truth data, and one that cannot.
Verification asks whether the code solves the equations it claims to solve. Validation asks whether those equations describe reality. The first can be closed with care and open references; the second needs measured flight data. Everything on this page is the first kind. The second is still open and is named as such at the bottom.
Headline results
Evidence summary
Test 1 · closed-form comparison
The propagator against Allen–Eggers
Allen–Eggers gives a closed-form velocity profile for unlifted ballistic entry into an exponential atmosphere, neglecting gravity against drag. Driving the integrator into those same assumptions gives a check that depends on no other part of this build.
Test 2 · order of accuracy
Is the integrator actually fourth order?
Halving the step should cut a fourth-order method's error roughly sixteen-fold. Measured on a state variable at fixed time against a reference run at a step 512 times finer.
Test 3 · conservation
Energy, drag-free
Test 5 · adversarial
Red-teaming the novelty detector
Test 4 · cross-model transfer
De-circularising the classifier
A classifier trained on tracks from the same propagator it then constrains learns that propagator's habits. This one is trained on a deliberately different forward model and evaluated against the shipping one.
What this does and does not buy