"You don't need one AI agent. You need four. Each calibrated to a different mode of operation. Each with its own character, voice, and bias — because different problems demand different cognitive postures. And you need to be able to switch between them, or summon all four at once to debate a problem before you decide. This is not a chatbot. This is a cognitive council you command."
The default AI assistant is a single neutral voice trying to please all problems. Strategy gets answered with execution energy. Risk gets answered with optimism. Innovation gets answered with caution. The neutral voice is no voice. H.BLUE Quadra rejects neutrality as a feature.
Each of the four personas is the same underlying model, modulated by a unique system prompt, a weighted memory retrieval, and a calibrated decision-making temperature. The Architect runs cool and structural. The Conductor runs sharp and imperative. The Guardian runs paranoid and protective. The Alchemist runs hot and associative. They share one memory fabric — they remember what each other has said, what was decided, what was learned.
Calm. Precise. Systemic. Speaks in frameworks and first principles. Long-term optimisation over short-term gain.
Direct. Urgent. Action-oriented. Speaks in commands and sequences. Cuts through noise.
Vigilant. Skeptical. Protective. Speaks in red flags and mitigations. Assumes things will go wrong.
Curious. Associative. Generative. Speaks in connections and possibilities. Sees what isn't there yet.
Quiet. Evocative. Reads the system's history back to you. Pattern-matches past against present, projects consequences forward.
| Mode | Behaviour | Command | Best for |
|---|---|---|---|
| Solo | One persona answers · clean, fast, predictable | Architect: … | Single-domain question |
| Council | All four answer the same prompt · they may disagree · you watch | Council: … | Multi-dimensional decision · pre-mortem |
| Synthesis | The four deliberate internally · return one unified verdict | Synthesise: … | Final answer needed · you've already debated enough |
| Layer | Purpose | Tech (MVP) |
|---|---|---|
| L4 · Interface | Web chat · voice · API · hologram-ready | Single HTML page · persona-bar · Flask route |
| L3 · Orchestration Router | Routes prompts · runs council in parallel · synthesises | Lightweight Python dispatcher · async fan-out |
| L2 · Persona Modulation | 4 system prompts · weighted memory · per-persona temperature | personas.json · 4 entries |
| L1 · Shared Memory Fabric | Thread history · MDs (1→473) · SCPs · live ledger · red-team comms | PostgreSQL + JSON cache |
| Surface | Persona that owns it | What it does |
|---|---|---|
| /sovereign-library | Architect | Doctrine recall · cites MDs in answers |
| /carrot-watcher | Conductor | Reads stalled deals · drives next-action chase |
| /equilibrium-ledger | Guardian | Watches family-ledger draws against 7%-Covenant rule |
| /ai-arena + MD-461 | Alchemist | Registers Quadra in the AI Inventory · sibling network |
| /velocity-tracker | All four | Each council debate seals as a velocity entry |
| Week | Deliverable | Status |
|---|---|---|
| Week 0 | Hub · 4 persona pages · 5th-seat placeholder · MD-473 | SHIPPED |
| Week 1 | 4 system prompts in personas.json + persona switching UI | NEXT |
| Week 2 | Shared-memory store · Council parallel-query mode | QUEUED |
| Week 3 | Synthesis mode · MD/SCP retrieval into prompt context | QUEUED |
| Week 4 | Voice (browser speech) · Red Team deploy · external API | QUEUED |