EXP-SLC

✓ success

Scrambled-LLM Control

2026-03-07 L3 level paper3 3 runs $0.06

new

premortem

mock

real

metrics

analyze

review

krit

done

In plain language

The LLM is not a neutral medium — it actively shapes the outcome. The same initial conditions produce different structures depending on which model interprets the text. The interpreter is a participant, not a tool.

What we found

The LLM interpreter creates organization — it is not neutral substrate. NORM > SHUF > UGRM gradient confirmed (p=0.002).

The LLM interpreter creates organization — it is not neutral substrate. NORM > SHUF > UGRM gradient confirmed (p=0.002). Fragment persistence is the key discriminator: normal processing maintains coherent structures across iterations, scrambling destroys them. The interpreter IS the source of organization, not a passive medium. This is interpreter-dependence.

Predictions we made before running

1/3 confirmed · 2 refuted

▶ Technical details

Research hypothesis

The LLM interpreter is not neutral substrate — it actively creates organization. Three conditions test this: normal processing (NORM), shuffled token input (SHUF), and unigram random baseline (UGRM). If organization is an LLM artifact, NORM >> SHUF ≈ UGRM. If organization is substrate-inherent, NORM ≈ SHUF >> UGRM.

Experimental setup

Type: three_level_gradient

Condition	Parameters
NORM	label: Normal LLM processing (intact prompt → response), n_replicates: 3
SHUF	label: Shuffled tokens (word order randomized before LLM call), n_replicates: 3
UGRM	label: Unigram baseline (words sampled from frequency distribution), n_replicates: 3

Parameters

model: gemini-2.5-flash-lite
n_replicates: 3
n_rounds: 200
temperature: 0.9

Trophic Ratios by Condition

Mean trophic ratio per agent across runs. Error bars = ±1 std dev. Higher TR = more upstream (exporter).

agent interpreter-dependence control organization