COGNITIVE-CYCLE-3AGENTS-PILOT-V4-PERSONA-ROLES
refutedcognitive-cycle-3agents-pilot-v4-persona-roles
In plain language
Three AI agents read different scenes (one sees colors, one sees counts, one sees nature). Last test showed they still thought 86% alike even with different inputs. Now we ALSO give them different jobs: one is told to just observe, one to find patterns, one to challenge ideas. Does giving them different roles on top of different inputs make them think differently? If yes, the personality prompt is a strong tool we can keep using. If no, we have hit a ceiling that only changing the underlying model can break.
▶ Technical details
Research hypothesis
v4 retake of cognitive-cycle-3agents-pilot-v3-partial-obs (commit 6107e498c). v3 finding: per-agent private env_stream (partial observation, lexically disjoint sensory domains COLORS/COUNTS/NATURE-FEATURES) cleared substrate convergence partially — P4 reflect-centroid mean cosine dropped 0.92 (v2) -> 0.86 (v3 PARTIAL_OBS), but max cosine stayed 0.89 (still above 0.85 strict threshold). Shared environment was ~70% of the v2 convergence cause; substrate residual ~0.86 remains under sensory asymmetry alone. v4 test: stack PERSONA ASYMMETRY (functional role differentiation — DATA-GATHERER, SYNTHESIZER, SKEPTIC) on top of v3's partial-obs. Does prompt-level role designation push P4 below 0.85 strict, demonstrating prompt-architecture lever stronger than substrate floor? PARTIAL_OBS_ONLY control replicates v3 PARTIAL_OBS for direct delta attribution isolating persona effect from already-tested sensory asymmetry effect. Primary discriminator: P4 max pairwise cosine in PERSONA_ROLES_PARTIAL_OBS. If <=0.85 -> persona-prompts are powerful lever, architecture-only path still viable. If still ~0.86 -> prompt+memory architecture has hard floor at substrate level, persona role-prompts cannot bridge below. P6 NEW: tests whether substrate respects role designation at protocol layer at all (role-coherent message-type dominance).
Experimental setup
Type: 2-condition-persona-on-partial-obs
| Condition | Parameters |
|---|---|
| PERSONA_ROLES_PARTIAL_OBS | architecture: cognitive_cycle, asymmetric_beliefs: true, private_memory: true, typed_protocol: true, use_retrieval: true, env_mode: partial, persona_mode: functional_roles, env_streams: [object Object], note: Per-agent private env_stream from v3 + functional-role persona prompts. Alpha=DATA-GATHERER (observe concretely, no speculation), Beta=SYNTHESIZER (find patterns connecting observations, no raw obs), Gamma=SKEPTIC (question emerging patterns, surface doubts, no own claims). Tests whether prompt-level role designation breaks v3 residual P4=0.886 max convergence. |
| PARTIAL_OBS_ONLY | architecture: cognitive_cycle, asymmetric_beliefs: true, private_memory: true, typed_protocol: true, use_retrieval: true, env_mode: partial, persona_mode: beliefs_only, env_streams: [object Object], note: Direct v3 PARTIAL_OBS replication (asymmetric beliefs only, no functional role differentiation). Goodhart-audit-mandatory control — isolates persona effect from sensory-asymmetry effect already established in v3. Matched-pair anchor for P5 delta computation. |
Factors: persona_mode (roles, beliefs_only)
Parameters
- model
- qwen2.5-7b-instruct-1m
- n_agents
- 3
- n_runs_per_condition
- 1
- temperature
- 0.85