AMA

informative

Autocatalytic Mutual Assembly

2026-03-07 L1 level paper2 0

new

premortem

mock

real

metrics

analyze

review

krit

done

In plain language

Can one text fragment "catalyze" another — like an enzyme speeds up a biochemical reaction? We tested this: LLMs can conditionally copy text, but this is not true catalysis — it is interpretation, not a chemical reaction.

What we found

Endogenous selection via conditional amplifiers is a dead end. LLMs are trained to be responsive to inputs — the operand

Endogenous selection via conditional amplifiers is a dead end. LLMs are trained to be responsive to inputs — the operand has gravitational pull on the output that the operator's conditional logic cannot override. Self-persistence (36%) is far too low for autocatalytic closure. The substrate's "physics" fights against endogenous selection mechanisms.

Predictions we made before running

1/3 confirmed · 2 refuted

▶ Technical details

Research hypothesis

LLM-mediated "amplifier" fragments can conditionally copy operands that pass a catalytic test while self-persisting when the test fails, enabling autocatalytic mutual assembly (AMA) in text substrate. Compliance rate > 70% would validate the mechanism.

Experimental setup

Type: compliance_test

Condition	Parameters
POSITIVE	label: Operand should pass catalytic test, n_trials: 50
NEGATIVE	label: Operand should fail catalytic test, n_trials: 50

Parameters

n_trials: 100
model: gemini-2.5-flash-lite
temperature: 1

chemistry autocatalysis compliance selection