← back to timeline

EXP-RATCHET-NULL-SHUFFLED

artifact

L6 Null Model — Shuffled Input (Group Without Inheritance)

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

In plain language

This experiment tested whether a "cultural ratchet" effect, where knowledge builds up over time, requires information to be passed down in a specific, coherent order. Researchers created a version where the information used to start each new generation was randomly selected from past generations, rather than directly from the immediately preceding one. The results showed that this random shuffling broke the cultural ratchet effect, suggesting that coherent, sequential transmission of knowledge is necessary for it to occur.

Technical details

Research hypothesis

NULL MODEL: SHUFFLED INPUT

exp-cultural-ratchet showed semantic ratchet in INHERITANCE condition. FRESH_START control used original seed every generation — but this tests "inheritance vs no inheritance", not "coherent inheritance vs random input."

CRITICAL QUESTION: Does the ratchet require COHERENT transmission?

SHUFFLED condition: 3 agents per generation, same as original. But between generations, the seed fragments come from a RANDOM previous generation's pool (not the immediately preceding one). This breaks the causal chain while preserving the STATISTICAL PROPERTIES of inherited text.

If shuffled produces same ratchet as coherent inheritance → the ratchet is just "LLM + any complex input → elaboration." No cultural memory needed.

If shuffled does NOT produce ratchet → coherent transmission is necessary. Cultural accumulation requires not just rich input but SEQUENTIAL building.

This is the stronger null model. Fresh-start (original experiment) tests "any seed vs inherited seed." Shuffled tests "random inherited seed vs coherent inherited seed." The latter isolates the SEQUENTIAL component.

DESIGN: 3 agents × 10 rounds × 4 generations. Between generations: select top-3 from the CURRENT generation, but seed NEXT generation with top-3 from a RANDOM PREVIOUS generation (not the one that just ran). Specifically: G1 seed = G0 selections (same as coherent — only one source). G2 seed = G0 selections (skip G1). G3 seed = G1 selections (skip G2). This breaks temporal coherence while keeping text quality similar.

Experimental setup

Type: multi_generational_shuffled

Condition Parameters
SHUFFLED description: Group runs per generation. Seeds come from non-consecutive previous generation., inheritance: shuffled, seed_protocol: G0: S1 (original) G1: top-3 from G0 (only one source available — same as coherent) G2: top-3 from G0 (SKIP G1 — break chain) G3: top-3 from G1 (SKIP G2 — break chain)

Factors: n_agents_per_generation (3) × n_rounds_per_generation (10) × n_generations (4) × selection_size (3) × selection_criterion (vocab_richness)

Parameters

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