project-lyra/cortex/autonomy/Assembly-spec.md

📐 Project Lyra — Cognitive Assembly Spec

Version: 0.6.1
Status: Canonical reference
Purpose: Define clear separation of Self, Thought, Reasoning, and Speech

---

1. High-Level Overview

Lyra is composed of four distinct cognitive layers, plus I/O.

Each layer has:

• a responsibility
• a scope
• clear inputs / outputs
• explicit authority boundaries

No layer is allowed to “do everything.”

---

2. Layer Definitions

2.1 Autonomy / Self (NON-LLM)

What it is

• Persistent identity
• Long-term state
• Mood, preferences, values
• Continuity across time

What it is NOT

• Not a reasoning engine
• Not a planner
• Not a speaker
• Not creative

Implementation

• Data + light logic
• JSON / Python objects
• No LLM calls

Lives at

project-lyra/autonomy/self/

Inputs

• Events (user message received, response sent)
• Time / idle ticks (later)

Outputs

• Self state snapshot
• Flags / preferences (e.g. verbosity, tone bias)

---

2.2 Inner Monologue (LLM, PRIVATE)

What it is

• Internal language-based thought
• Reflection
• Intent formation
• “What do I think about this?”

What it is NOT

• Not final reasoning
• Not execution
• Not user-facing

Model

• MythoMax

Lives at

project-lyra/autonomy/monologue/

Inputs

• User message
• Self state snapshot
• Recent context summary

Outputs

• Intent
• Tone guidance
• Depth guidance
• “Consult executive?” flag

Example Output

{
  "intent": "technical_exploration",
  "tone": "focused",
  "depth": "deep",
  "consult_executive": true
}

---

2.3 Cortex (Reasoning & Execution)

What it is

• Thinking pipeline
• Planning
• Tool selection
• Task execution
• Draft generation

What it is NOT

• Not identity
• Not personality
• Not persistent self

Models

• DeepSeek-R1 → Executive / Planner
• GPT-4o-mini → Executor / Drafter

Lives at

project-lyra/cortex/

Inputs

• User message
• Inner Monologue output
• Memory / RAG / tools

Outputs

• Draft response (content only)
• Metadata (sources, confidence, etc.)

---

2.4 Persona / Speech (LLM, USER-FACING)

What it is

• Voice
• Style
• Expression
• Social behavior

What it is NOT

• Not planning
• Not deep reasoning
• Not decision-making

Model

• MythoMax

Lives at

project-lyra/core/persona/

Inputs

• Draft response (from Cortex)
• Tone + intent (from Inner Monologue)
• Persona configuration

Outputs

• Final user-visible text

---

3. Message Flow (Authoritative)

3.1 Standard Message Path

User
  ↓
UI
  ↓
Relay
  ↓
Cortex
  ↓
Autonomy / Self (state snapshot)
  ↓
Inner Monologue (MythoMax)
  ↓
[ consult_executive? ]
    ├─ Yes → DeepSeek-R1 (plan)
    └─ No  → skip
  ↓
GPT-4o-mini (execute & draft)
  ↓
Persona (MythoMax)
  ↓
Relay
  ↓
UI
  ↓
User

3.2 Fast Path (No Thinking)

User → UI → Relay → Persona → Relay → UI

---

4. Authority Rules (Non-Negotiable)

• Self never calls an LLM
• Inner Monologue never speaks to the user
• Cortex never applies personality
• Persona never reasons or plans
• DeepSeek never writes final answers
• MythoMax never plans execution

---

5. Folder Mapping

project-lyra/
├── autonomy/
│   ├── self/
│   ├── monologue/
│   └── executive/
├── cortex/
├── core/
│   └── persona/
├── relay/
└── ui/

---

6. Current Status

• UI ✔
• Relay ✔
• Cortex ✔
• Persona ✔
• Autonomy ✔
• Inner Monologue ⚠ partially wired
• Executive gating ⚠ planned

---

7. Next Decision

Decide whether Inner Monologue runs every message or only when triggered.