Experiments

V8: Protocell Agency

Protocell Grid World16×16 portion — resource patches, agents, observation windows★Resource patchFocal agentSame typeOther type
Protocell Agent Architecture (V8–V18)extractinputW₁·tanh(W₂·h)softmax(W_a·h)action→observation loop (ρ wall)MSE(ŷ, y)Grid World32×32, N agents, R resourcesObservationlocal window + compassGRU CoreH=16 or H=32Hidden State heff. rank 5–11Prediction Headlinear or 2-layer MLPAction Head4 dirs + eat + (comm)SGD Stepwithin-lifetime learning

Small recurrent agents, about 3,400 parameters each, on a discrete grid with a bounded 5×5 sensory field and real actions: move, consume, emit. Their observations are shaped by their own actions. ρsync≈0.21\rho_{\text{sync}} \approx 0.21 in all three seeds from cycle zero, seventy times Lenia, which is true by construction: the loop exists because the architecture has one. Self-model salience crossed threshold in two of three seeds, through a self-prediction head I installed. The hidden-state analyses originally reported for this study were on zeroed buffers and are retracted (below).

The Necessity Chaineach step necessary, not contingent (V8)1MembraneBoundary separating self from non-selfdiscrete grid bodynecessitates2Free Energy GradientViability requires energy harvestingresource depletion → deathnecessitates3World ModelPrediction of own future energyC_wm = 0.10–0.15necessitates4Self-ModelAgent encodes own state more than environmentSM_sal > 1.0 in 2/3 seedsnecessitates5Affect GeometrySimilarity structure on internal statesRSA ρ > 0.21, eff. rank 5–11

Study record: metadata, result path, seeds, source code, and key finding for .