COGNITIVE-CYCLE-3AGENTS-PILOT

negative

cognitive-cycle-3agents-pilot

2026-05-24 L7 level 4 runs

new

premortem

mock

real

metrics

analyze

review

krit

done

In plain language

Three small AIs with different starting beliefs watch the same environment, each keeping its own private journal and talking to the others using a structured message protocol. We check whether they develop distinct personalities and figure out the hidden structure of what they see — and whether a control group without the private journals or structured talk fails to do the same.

▶ Technical details

Research hypothesis

Test whether non-trivial prompt+memory architecture (asymmetric beliefs + private memory + typed-message protocol + structured cognitive cycle + retrieval-augmented context) instantiates life-like properties (differentiated cognitive styles, stable emergent roles, goal coordination, memory divergence) in LLM substrate that are absent in naive baselines. This is the proper substrate test, replacing 3 prior trivial-baseline negatives (differentiation-without-injection fa4f19fdb, genealogical-replicator 413858e61, friston-minimal-pilot 4647b0491) which used identical agents in shared memory with no asymmetry, boundary, or retrieval.

Experimental setup

Type: 2-condition-architecture-vs-naive

Condition	Parameters
COGNITIVE_CYCLE	architecture: cognitive_cycle, asymmetric_beliefs: true, private_memory: true, typed_protocol: true, use_retrieval: true, note: Full architecture: 3 agents with different starting beliefs, per-agent private memory, typed-message protocol (OBSERVATION/QUESTION/CLAIM/DOUBT/SILENT), structured PERCEIVE-REFLECT-PLAN-ACT cycle, retrieval-augmented context (top-3 memories by cosine vs env+latest-msgs).
NAIVE_REPLICATION	architecture: naive_replication, asymmetric_beliefs: false, private_memory: false, typed_protocol: false, use_retrieval: false, note: Replicates differentiation-without-injection baseline as control: 3 agents with IDENTICAL starting belief, shared text memory (last 10 lines dumped), free-text output (no typed protocol), no retrieval. Mandatory Goodhart-audit per AUTHORS-MATRIX — distinguishes architecture effect from substrate noise.

Condition

Parameters

COGNITIVE_CYCLE

architecture: cognitive_cycle, asymmetric_beliefs: true, private_memory: true, typed_protocol: true, use_retrieval: true, note: Full architecture: 3 agents with different starting beliefs, per-agent private memory, typed-message protocol (OBSERVATION/QUESTION/CLAIM/DOUBT/SILENT), structured PERCEIVE-REFLECT-PLAN-ACT cycle, retrieval-augmented context (top-3 memories by cosine vs env+latest-msgs).

NAIVE_REPLICATION

architecture: naive_replication, asymmetric_beliefs: false, private_memory: false, typed_protocol: false, use_retrieval: false, note: Replicates differentiation-without-injection baseline as control: 3 agents with IDENTICAL starting belief, shared text memory (last 10 lines dumped), free-text output (no typed protocol), no retrieval. Mandatory Goodhart-audit per AUTHORS-MATRIX — distinguishes architecture effect from substrate noise.

Factors: architecture (cognitive_cycle, naive_replication)

Parameters

model: qwen2.5-7b-instruct-1m
n_agents: 3
n_runs_per_condition: 1
temperature: 0.85