Architectural Cognitive Diversity
Why different foundational models produce structurally different thinking — and what that means for deliberation.
Cognitive Quarantine solves Persona Collapse at the prompt level. But a second-order problem remains: even with structurally separated agents, if every agent runs on the same foundational neural network, the agents share a deeper consensus — the consensus of a single model's biases, training distribution, and alignment tuning.
The Model Signature
Each foundational model has a characteristic "signature" — patterns of reasoning, risk assessment, hedging, and commitment that emerge from how it was trained.
Claude Sonnet produces structured, cautious, schema-compliant output. It excels at synthesis and JSON compliance. It hedges more than other models and tends toward qualitative precision.
GPT-4o produces expansive, systemic, exploratory output. It builds broader maps and generates more parallel hypotheses. It is more willing to speculate.
Llama 3 (open-weight, different safety alignment) produces more direct, less hedged, more contrarian output. Its safety training is less aggressive, which makes it a stronger devil's advocate.
Gemini Pro produces generative, system-building output. It naturally gravitates toward operational frameworks and strategic structuring.
These are not quirks. They are architectural properties of how each model was trained — what data it saw, what RLHF preferences shaped its outputs, what safety constraints were applied.
Why This Matters for Deliberation
When all agents run on the same model, the deliberation inherits that model's biases uniformly. A full-Sonnet deliberation will be cautious, well-structured, and slightly risk-averse. A full-GPT-4o deliberation will be expansive, speculative, and slightly unfocused. The "Strategist" and "Critic" disagree on the surface, but their underlying probability distributions agree about what constitutes reasonable discourse.
Architectural Cognitive Diversity is the principle that assigning different foundational models to different agents creates adversarial friction at the model-architecture level, not just the prompt level. The friction is deeper because the models' training distributions are genuinely different — not just their instructions.
The Ideal Mapping
Based on each model's characteristic signature, the theoretical ideal assignment is:
| Agent | Ideal Model | Why |
|---|---|---|
| Strategist | GPT-4o | Expansive landscape mapping, parallel hypothesis generation |
| Critic | Claude Sonnet | Structural precision, analytical rigor |
| Devil's Advocate | Llama 3 | Less hedged contrarianism, more willing to commit to extreme positions |
| Synthesizer | Claude Sonnet | Schema compliance, structural synthesis, JSON reliability |
| Researcher | Gemini Pro | Broad knowledge grounding, systematic evidence gathering |
| Operator | GPT-4o | Practical operational framing |
Validation
The Validation Suite tests this empirically across multiple model configurations:
Test 1 (All Sonnet) — the V1 default. Single-model deliberation with Cognitive Quarantine only.
Test 2 (Mixed Frontier) — Sonnet + GPT-4o + Gemini. True architectural diversity across frontier models.
Test 3 (Strong Mix) — Opus + Sonnet. Different capability tiers from the same provider.
Early results suggest the Mixed Frontier configuration produces genuinely different unresolved_points than any single-model configuration. The diversity of training distributions surfaces tensions that a single model smooths over — because different models disagree about what constitutes a "reasonable" position.
Why V1 Is Single-Model
Despite the theoretical promise, V1 ships with Sonnet locked for all agents. The reason is practical, not philosophical.
The Conclusion Protocol depends on strict JSON schema compliance. The Synthesizer in Conclusion Mode (temperature 0.2) must produce valid, schema-compliant JSON every time. The Blind Critic must produce exactly "PASS" or "FLAG: [Field] - [Objection]" with no conversational filler. Different providers handle schema enforcement, escape sequences (especially \n inside JSON strings), and output formatting differently.
Mixing models at the Synthesizer or Blind Critic layer introduces a class of failure modes — malformed JSON, missing fields, inconsistent escape handling — that can corrupt an otherwise valid deliberation. V1 must ship with a stable conclusion pipeline.
The per-agent model configuration UI (BYOK — Bring Your Own Key) is designed but deferred to V2. The Validation Suite proves the principle offline; the product exposes it after the conclusion pipeline has been hardened.
The Pedigree Connection
The Vitakka integration path adds another dimension: once conclusions are first-class memory items in Vitakka, users will want to see which model produced which insight. The Pedigree of Thought includes model provenance — knowing that the Strategist was GPT-4o and the Critic was Claude is part of understanding why the deliberation reached its conclusion.
That is V2 territory.
Next: The Three Rounds — the core architecture of a Vāda deliberation.