Input-Supported Population Regulation of an Attractor Registry: Formation, Pruning, and a Two-Factor Substrate Boundary in ADAM, an Online Cognitive Runtime

This paper tests whether an online cognitive runtime can grow and shrink its own attractor memory registry under structured experience, and whether those registry changes reach downstream behavior. Across three experimental protocols with 30 matched seeds each, structured exposure to a coherent target prompt family produces target-provenance mature attractors in 29/30 seeds (paired Wilcoxon p ≈ 6.44e-7), independent of temporal adjacency. Formation is bounded by a two-factor substrate condition: within-family trajectory coherence is the primary gate, and a theory-critical 0/30 fail family rules out a single-axis substrate-separation account. Formed memories persist under continued exposure (0/30 pruned) and are pruned under starvation (29/30; p ≈ 6.44e-7). The pruning pathway is lifecycle death from passive stability decay — tracked-target inactive-archive events never occur, and the ecology toggle produces identical tracked-target pruning counts (30/30). A within-subject causal ablation shows that suppressing formed target memories changes the cognitive field state (14/14) and safety logit (14/14, mean Δℓ = +5.00), while 16/16 seeds without such memories show exactly zero effect — an integrity check built into the design-frozen protocol. These findings show that an attractor-based cognitive runtime can grow, regulate, and functionally employ a memory population under controlled experience streams: the registry changes with structured input, is pruned under unsupported absence, and shifts downstream computation when directly perturbed.