1. Read the configuration before the answer. Cell 1 runs the reference as configured, capped at one iteration, then the same reference with the cap lifted, then the port. The port agrees with the lifted reference to within 1e-12 and sits 68.5 percent from the configured one. Note what the configured answer did with dof 0.
2. Choose the judge. Two answers to one problem can be compared fairly only by the objective the reference itself minimises, which in MuJoCo is what mj_constraintUpdate scores. Write excess_cost so it returns a number rather than a verdict, because the size of the excess is what cell 3 checks against the geometry.
3. Run the control. Against the cap-lifted reference the faithful port must show no excess at all. If it shows one, your function is wrong, not the port.
4. Convict the planted bug. The planted copy charges friction half price once the joint slips. Against the cap-lifted reference it must pay, and the Gauss term fixes how much: at least half its distance squared in the mass-matrix norm, the same property that makes the optimum unique.
5. Watch two wrong questions acquit it. Against the reference as configured the planted bug looks cheaper, because one iterate never let the joint slip, and slipping is where the bug lives. Scored by its own objective it looks cheaper again, because every converged solver wins its own game.
6. Run the three checks the bench cannot. Count the input elements on both sides, reconcile the number of subjects compared against the corpus, and pair every disagreement by identity. Only then read what is left as physics.