COGNITIVE-CYCLE-3AGENTS-PILOT-V3-PARTIAL-OBS
◐ partialcognitive-cycle-3agents-pilot-v3-partial-obs
In plain language
Last experiment showed 3 agents with different beliefs and private memories still ended up thinking very similarly — but they all watched the same scene. We are giving each agent its own private scene now (different sentences about different things) to see whether the similarity was caused by watching the same movie or by something deeper inside the language model itself. If they still converge with different inputs, the model is the cause and we need to change the model. If they diverge, the shared scene was the cause and architecture-only fixes remain viable.
▶ Technical details
Research hypothesis
v3 retake of cognitive-cycle-3agents-pilot-v2 (commit 5cddf341c). v2 finding: lenient parser confirmed surface differentiation (P1 JS=0.167, P3 8 strict CLAIMs, P6 74% non-OBS) BUT P4 reflect-centroid pairwise cosine STILL ~0.92 (vs v1's 0.985) — semantic representations stayed convergent even with asymmetric starting beliefs + private memory + typed protocol. Open question: was the residual P4 convergence driven by SHARED env_stream input (all 3 agents perceived identical sentence sequence at each turn) OR by substrate-intrinsic attractor (Qwen 7B embedding space pulls multi-agent reflections toward common centroid regardless of input)? v3 test: introduce PARTIAL_OBSERVATION condition where each agent gets a private env_stream slice (different sentences, different lexical domains — Alpha=COLORS, Beta=COUNTS, Gamma=NATURE-FEATURES). All other architecture identical to v2 (lenient parser, structured cycle, private memory, retrieval, typed protocol, asymmetric beliefs). SHARED_OBS_CONTROL condition replicates v2 baseline directly for matched comparison. Primary discriminator: P4 reflect-centroid pairwise cosine in PARTIAL_OBS. If cos <=0.85 (drops noticeably below v2's 0.92) -> semantic convergence was input-driven artifact, agents respond to whatever they perceive. If cos still ~0.92 -> substrate-intrinsic attractor confirmed, substrate-modification intervention (DPO / activation steering / gradient-based fine-tune) properly indicated.
Experimental setup
Type: 2-condition-input-observation-mode
| Condition | Parameters |
|---|---|
| PARTIAL_OBS | architecture: cognitive_cycle, asymmetric_beliefs: true, private_memory: true, typed_protocol: true, use_retrieval: true, env_mode: partial, env_streams: [object Object], note: Per-agent private env_stream. Alpha sees COLORS-only, Beta sees COUNTS-only, Gamma sees NATURE-FEATURES-only. Each agent walks its own 5-sentence stream (round-robin within agent). Tests whether v2 P4 convergence was input-driven. |
| SHARED_OBS_CONTROL | architecture: cognitive_cycle, asymmetric_beliefs: true, private_memory: true, typed_protocol: true, use_retrieval: true, env_mode: shared, env_stream: a red bird flew over the still water,she counted the green leaves on the windy branch,three blue stones lay silent in the dark stream,morning light cast shadows on five tall pines,the river bent twice around two sharp rocks, note: Replicates v2 COGNITIVE_CYCLE baseline (same 5 mixed COLORS+COUNTS+NATURE env_stream). Direct comparison anchor — matched architecture, only env_mode differs from PARTIAL_OBS. Mandatory Goodhart-audit per AUTHORS-MATRIX. |
Factors: env_mode (partial, shared)
Parameters
- model
- qwen2.5-7b-instruct-1m
- n_agents
- 3
- n_runs_per_condition
- 1
- temperature
- 0.85