← back to timeline

EXP-RATCHET-NULL-SINGLE

artifact

L6 Null Model — Single Agent Iteration (No Group)

2026-03-27 L6 level paper3 0
new
premortem
mock
real
metrics
analyze
review
krit
done

In plain language

This experiment tested if a group of AI agents was necessary for concepts to persist and evolve over time, or if a single AI agent could do it alone. Researchers found that a single AI agent, when iterating on its own work, did not show the same concept persistence as a group, suggesting that group interaction is important for this phenomenon.

Technical details

Research hypothesis

NULL MODEL: SINGLE-AGENT ITERATION

exp-cultural-ratchet showed semantic ratchet (concept families persist across generations, Jaccard 0.14, concept "attunement" survived 4 generations).

CRITICAL QUESTION: Is this a GROUP phenomenon or an LLM default?

If a SINGLE agent, iterating on its own output for N generations (same number of rounds), produces the same ratchet — the original result is trivial. LLMs naturally elaborate/structure text when iterated. No group needed.

If the single agent does NOT produce the ratchet (no concept accumulation, no increasing divergence from seed) — the group dynamic is necessary.

THIRD OPTION: Both show ratchet but with DIFFERENT dynamics (e.g., group produces more diverse concepts, single agent converges to narrow elaboration). This would mean group AMPLIFIES an LLM tendency — interesting but weaker than pure emergence.

DESIGN: One agent, 4 "generations" of 10 rounds each. Between generations: select top-3 fragments (same selection protocol as exp-cultural-ratchet). Re-seed next generation with selected fragments. Same agent type (synthesizer) across all generations.

Compare directly with exp-cultural-ratchet INHERITANCE condition.

Experimental setup

Type: single_agent_iteration

Condition Parameters
SINGLE_ITERATION description: One agent iterates alone. Same total rounds as group (3×10=30 per gen)., inheritance: true, agents: 1

Factors: n_agents (1) × n_rounds_per_generation (30) × n_generations (4) × selection_size (3) × selection_criterion (vocab_richness)

Parameters

model
gemini-2.5-flash-lite
temperature
0.9
scheduler
round_robin