EXP-KAUFFMAN-COMP

KAUFFMAN COMPOSITION HYPOTHESIS (L4 — Organism-Level Novelty Generation)

BACKGROUND: Friston-style boundary (Markov blanket) fails in LLM substrate — agents CONVERGE rather than separate (ANTI-FRISTON finding). Kauffman's alternative: life isn't about boundaries but about COMPOSITION — autocatalytic sets where products catalyze each other's creation, generating novelty not present in any component alone.

TRANSLATION TO LANGUAGE SUBSTRATE: If multi-agent interaction is compositional (Kauffman sense), then:

• LIVE groups should produce vocabulary items (words, n-grams, or phrases) that appear in NEITHER agent's isolated output
• These "novel compositions" should INCREASE over time (autocatalytic growth)
• The rate of novel composition should exceed chance recombination

This is DISTINCT from convergence (VP): VP measures agents becoming SIMILAR. Composition measures agents producing something NEITHER would produce alone. VP is about overlap; composition is about the COMPLEMENT — what's in the joint output but absent from all individual outputs.

KEY METRIC: Composition Index (CI) CI = |V_joint - (V_a ∪ V_b ∪ V_c)| / |V_joint| Where V_joint = vocabulary of LIVE group, V_x = vocabulary of same agent in ISOLATED CI > 0 means interaction creates words no agent produces alone. CI growing over rounds = autocatalytic composition.

CONTROL LOGIC:

• PERSONA_ISOLATED: each agent writes alone (same persona, same seed). Their individual vocabularies = baseline repertoire.
• PERSONA_LIVE: same agents interact. Novel vocabulary = V_live - (V_alpha_isolated ∪ V_beta_isolated ∪ V_gamma_isolated)
• IDENTICAL_LIVE vs IDENTICAL_ISOLATED: same but without persona differentiation. If composition requires differentiation → CI_persona > CI_identical

WHY L4 (NOT L3): L3 = self-organization (hierarchy forms). Already confirmed. L4 = organism-level properties. Compositional novelty is an organism-level property: the GROUP generates output that no COMPONENT can generate alone. This is the canonical definition of organismal emergence — the whole exceeds the union of its parts, not just their sum.

EXPLORATION TIER: N=3 runs per condition. Goal: does composition index > 0?

CONNECTION TO OPEN QUESTIONS:

• "morphodynamic ceiling" → composition = potential teleodynamic signature
• "Kauffman composition axis — design measurement" → THIS experiment
• "ANTI-FRISTON" → alternative axis that works in LLM substrate