Self-Model as Social Decoupler: Ablation and Baseline Evidence for Reflective Mediation in Proto-Conscious Agents

The PROSENTIR series investigates the computational correlates of proto-consciousness through agent-based simulations of iterative cooperation. Paper 6 addresses a central open question from Papers 4 and 5: does the self-model mechanism causally produce the internal autonomy indexed by C, or does it merely correlate with it? We report three independent empirical modules. Module C validates mutual information estimators: MINE-SMILE diverges systematically (Δ mean ≈ 56) on emergent differential-of-sigmoid distributions, while the Kraskov k-NN estimator remains stable (Φₚ = 2.19), ratifying Kraskov as the canonical estimator for the PROSENTIR series with empirical rather than conventional justification. Module A ablates the self-model (wₛ = 0, N = 200 pairs per condition): the self-model suppresses the synchrony-cooperation coupling from r = 0.478 to r = 0.029 (p < 0.0001), confirming H6.1, and introduces cooperative behavioral variance (σ = 0.151 vs. 0.062 without self-model). Module B compares PROSENTIR against three baselines (RNN, Tit-for-Tat, Memory agent, N = 200 each): all systems without self-model show synchrony-cooperation coupling r ≥ 0.478; only PROSENTIR with self-model collapses that effect. Additionally, the autonomy-cooperation trade-off present in all baselines (r = −0.551 to −0.606) is neutralized in PROSENTIR (r = −0.045). We interpret these results as computational evidence for the saber-que / saber-desde distinction formalized in Paper 3: without the self-model, agents act as social mirrors; with it, they exhibit reflexive mediation of social responses. The cost of this mediation is social desynchronization; the benefit is behavioral stability and trade-off neutralization.