Concept Acquisition Through Attractor Dynamics: Formation, Generalization, Discrimination, Causal Influence, and Cross-Family Replication in a Minimal Cognitive Architecture

Description:Can a minimal cognitive architecture built on attractor dynamics support concept-like representations across formation, generalization, discrimination, causal behavioral influence, and system-level learning? We built one and asked. In a 32-dimensional field-space runtime with a frozen safety projection and MiniLM embeddings, six linked experiments spanning five semantically distinct concept families and 150 independent seeds produced a chain of results. Structured experience reliably produces attractors, but formation alone does not distinguish concept-relevant from concept-irrelevant populations. Generalization provides the key filter: attractors with target-family provenance activate on held-out within-category prompts, while control attractors do not (77% vs 0%). The same attractors discriminate within-category from near-category prompts (AUC = 0.683; 74% of testable seeds AUC ≥ 0.60), though boundary tightness varies across seeds. Selective ablation of target-provenance attractors causally changes behavioral output in 96% of seeds. After formation, every seed produces different behavioral output on identical held-out prompts, with target-family prompts shifting 15.8x more than filler prompts. This learning effect replicates across all five families (98.7% verdict change, 94.0% family-specific), though the medical misinformation family shows a qualitatively distinct pattern. A prototype centroid baseline achieves 100% generalization (versus ADAM's 77%) and equivalent discrimination, indicating that the attractor mechanism's contribution is dynamical: field-state dynamics, causal intervenability, and system-level behavioral change, not static representation. Under minimal conditions, an attractor-based cognitive system can support concept acquisition with measurable causal effects — modifiable, ablatable, and functionally coupled to a downstream readout in ways a static representation cannot. This is v1, assembled from raw Markdown source. The canonical manuscript will be formatted in LaTeX for journal submission.