Do Theoretical Labels Add Computational Content? Reproducing and Stress-Testing an FEP-RSI Model

Computational models of selfhood can reproduce theory-consistent dynamics while leaving the theoretical interpretation underdetermined. We independently reproduce a public model that maps Lacan's Real, Symbolic, and Imaginary (RSI) registers onto coupled discrete active-inference units, then separate reproducibility, component engagement, and construct specificity. Under a frozen runtime, five repeated source runs are bitwise stable and a clean-room NumPy implementation exactly matches all sampled actions and 2,700 trajectory cells. Message, weighting, and edge interventions alter 71.4%, 66.4%, and 55.9% of relevant trajectory entries. Yet bijectively replacing RSI names by generic factor identifiers leaves trajectories exactly unchanged. On a separate 240-episode coordination probe, the coupled model beats a no-message lesion but loses to DeGroot consensus and an informed-star control. The result does not reject Lacanian theory; it shows why computational accounts of the self need discriminant controls beyond successful simulation.