PROSENTIR Phase 5: Internal Autonomy vs. Social Sensitivity — The Negative Coupling Between Integration and Cooperation
Authors/Creators
Description
We present PROSENTIR Phase 5, reporting the results of the
largest experiment in the series (N=200 independent agent pairs,
T_game=1000) and establishing the clearest empirical pattern
across five papers: internal information integration measures
negatively predict social cooperation.
The central hypothesis — that inter-agent mutual information
MI(M3_A, M3_B) via MINE predicts mutual cooperation — is
definitively falsified (r=-0.007, p=0.926, N=200). Two secondary
metrics yield highly significant results: C index v2.0
(r=-0.203, p=0.004) and Learning-Action coupling L
(r=-0.258, p=0.0002) both negatively predict cooperation,
consistently across three independent runs totaling N=400
simulations.
We name this the Internal Autonomy Hypothesis: agents with
higher internal information integration (measured by the digital
signatures τ, Φ, C, L introduced in Phases 1-4) are less
sensitive to their partner's social state, producing lower
mean mutual cooperation. This is a structural consequence of
the dominant self-model weight (w_S=0.45) consuming social
input bandwidth.
Four explicit limitations are acknowledged before results:
no ablation studies, no hyperparameter sensitivity analysis,
no baseline comparisons, and restriction to a single social
environment (IPD). These define the Phase 6-8 research agenda.
Biological grounding is established in the H01 human cortical
connectome (Shapson-Coe et al., Science 2024): the cortical
laminar hierarchy is the conceptual analog of PROSENTIR's
abstraction hierarchy, and non-random multisynaptic connections
in human cortex mirror PROSENTIR's dynamic synchrony finding.
The 69%/31% excitatory/inhibitory ratio motivates Phase 7's
three-agent scenario with asymmetric E/I dynamics.
Related work:
- Phase 1: 10.5281/zenodo.19912924
- Phase 2: 10.5281/zenodo.20019260
- Phase 3: 10.5281/zenodo.20090917
- Phase 4: 10.5281/zenodo.20265133
Reproducible code: prosentir_v15.py (included).