EXP-CULTURAL-RATCHET

CULTURAL ACCUMULATION HYPOTHESIS (L6 — FIRST EVER)

L5 showed: group identity persists (64% in-group citation), asymmetric vocab propagation (abstract 3.5x > concrete), merge amplifies hierarchy.

KEY QUESTION: Can agents evolve CROSS-GENERATIONALLY? When one generation's outputs seed the next generation's inputs, does CULTURAL ACCUMULATION emerge?

What distinguishes culture (L6) from selection (L3):

• Selection (L3): top fragments survive. No modification by receivers.
• Culture (L6): outputs are TRANSMITTED + MODIFIED + ACCUMULATED. Each generation doesn't just SELECT — it BUILDS ON what came before. The ratchet: complexity can only increase because each generation inherits AND extends the prior generation's innovations.

PROTOCOL (multi-generational):

• Generation 0 (G0): 3 agents × 10 rounds. Fresh seed. Standard multi_agent.
• Selection: take top-3 fragments by vocab richness from G0.
• Generation 1 (G1): 3 NEW agents × 10 rounds. Seed = selected G0 fragments.
• Selection: take top-3 fragments by vocab richness from G1.
• Generation 2 (G2): 3 NEW agents × 10 rounds. Seed = selected G1 fragments.
• Generation 3 (G3): 3 NEW agents × 10 rounds. Seed = selected G2 fragments.

(4 generations total — minimum 3 transitions to observe trends)

CONTROL (fresh-start, no inheritance):

• Same protocol but each generation gets ORIGINAL seed (no inheritance).
• Same agent types, same round count. Only difference: no cultural memory.

WHY THIS IS L6 AND NOT L3: L3 = agents within a generation self-organize (hierarchy, differentiation). L6 = information RATCHETS across generations. Key markers: (a) Vocab complexity increases monotonically across G0→G3 (b) Novel words from G0 appear in G2 (skip-generation transmission) (c) Fragments become more structured (longer dependency chains) (d) Receivers MODIFY inherited content (not just copy) If only selection operates, control should perform equally well. If culture operates, inheritance condition should show monotonic increase.

THEORETICAL GROUNDING:

• Tomasello (1999): cumulative culture = ratchet effect. Each generation starts where the last left off, never resets to zero.
• Gabora (1998-2020): episodic→memetic transition. Memes evolve when they can be selectively transmitted and recombined.
• Dawkins (1976): meme as replicator. Fidelity + fecundity + longevity. Here: fidelity = do inherited words survive? fecundity = do they spread to all agents within a generation? longevity = do they persist across 2+ generations?
• Walker (assembly theory, 2017): complex objects require many steps to assemble. If fragment complexity increases across generations, the assembly index is increasing — signature of open-ended evolution.
• Vervaeke (relevance realization, 2019): agents don't just transmit — they select what's RELEVANT. Cultural evolution is selective accumulation, not indiscriminate copying.

CONTEXT FROM L5 (exp-community-merge):

• In-group citation: 0.638 (agents preferentially engage with familiar vocab)
• Asymmetric propagation: abstract 3.5x > concrete
• Prediction: cultural transmission should show SAME asymmetry — abstract/ theoretical vocabulary should accumulate faster than concrete/sensory.

EXPLORATION TIER (N=1). Goal: does the phenomenon exist at all?