AERIADNE // SOMNUS INTELLIGENCE STACK

The first-of-its-kind private marketplace for operator-grade intelligence.
Plugins · Skills · Operators · MCP Servers — compiled for sovereign execution.

---

1. OPERATIONAL DOCTRINE: THE OPERATOR VS. THE AGENT

This system does not run agents.

In the Somnus Sovereign Systems lexicon, an "agent" is a temporary, fragile, and disposable service worker—a real estate agent, a travel agent, a scraper script. An agent expects hand-holding, runs on naive loops, and collapses under the first lint error or compiler failure.

We authorize and deploy Operators.

An Operator is a persistent, certified badass. An Operator:

• Wields the cognitive stack directly to drive workspace state.
• Does not ask for permission to fix a broken test or run a dependency check.
• Takes deep ownership of the workspace topology.
• Executes multi-turn reasoning across tools, keeping focus under drift pressure.
• Orchestrates and debugs code directly on local metal without fear.

Aeriadne is the private marketplace that delivers these high-fidelity capability upgrades, plugins, custom skills, and server-plane activations to the Operator's environment.

---

2. THE SYSTEM PROVENANCE

  Aeriadne / Somnus Intelligence Stack
  ─────────────────────────────────────────────────────────────────────────
  Operator:     Daeron  ·  Somnus Sovereign Systems
  Runtime axis: Codex · Claude Code · OpenCode · Gemini CLI · Cursor
  Core Engine:  Muaddib Neural Network (24/7 Autonomous Tool Server)
  Server plane: Sovereign BB7 MCP  ·  /home/daeron/Somnus-MCP
  Plugin triad: Mentat · CTMv3 · Aeriadne
  ─────────────────────────────────────────────────────────────────────────
  This repo is the agent-operability, onboarding, and private-marketplace
  staging workspace. It is not the full sovereign stack — it is the support
  layer that makes that stack accessible to operators.

---

3. THE KERNEL CORE: MUADDIB NEURAL SERVER PLANE

At the center of the Somnus Intelligence Stack is the Muaddib Neural Network.

Muaddib is not a typical request-response API endpoint. It is a 24/7 autonomous server and gateway running continuously on the Operator's machine. It acts as the orchestrator of all tool planes, system connectors, and context gateways.

Proactive Tool Compilation (The bb7_ Philosophy)

When an Operator invokes a bb7_ tool (e.g., bb7_memory_store, bb7_dt_self_play, bb7_system_comprehensive_operation), the system does not simply execute a shallow command and return immediate output.

Instead, the bb7_ plane compiles an intelligent output. It:

1. Triggers multi-pass exploration across the codebase graph.
2. Interrogates the semantic memory store to retrieve historical design decisions.
3. Cross-references active session constraints and system monitor metrics.
4. Assembles a high-fidelity, synthesized context packet.

This process takes time. It is deliberately designed for depth, precision, and architectural correctness over cheap execution speed. It is built to ensure that when an Operator acts, they act with full context.

---

4. THE PLUGIN TRIAD

Three first-class plugins form the operational core of the stack. Each is designed to run in isolation or interlock to support the Operator.

 ┌──────────────────────────────────────────────────────────────────────────┐
 │  PLUGIN TRIAD  ·  Aeriadne Intelligence Stack  ·  Private v1             │
 ├──────────────────┬───────────────────────────────────────┬───────────────┤
 │  PLUGIN          │  ROLE                                 │  STATUS       │
 ├──────────────────┼───────────────────────────────────────┼───────────────┤
 │  [Mentat](Plugins/Mentat/)          │  Live session FSA + Q-table substrate │  Active       │
 │  [CTMv3](Plugins/Cognitive-Topology-Map/)           │  Workspace activation engine          │  Active       │
 │  [Aeriadne](Plugins/Aeriadne/)        │  CPF + private marketplace operator   │  Active       │
 └──────────────────┴───────────────────────────────────────┴───────────────┘

The release plugin surface is exactly this triad. Muaddib / Sovereign BB7 remains the core server plane under Servers/, not a plugin. Non-owned local development tools are not attached as marketplace packages.

4.1. Mentat — Live Cognitive Substrate

Mentat is a session-scoped state machine and reinforcement-learning Q-table that runs alongside the Operator's environment loop. It monitors tool executions, enforces scope constraints, and preserves session context across token compaction events.

Deterministic Finite State Automaton (FSA)

Mentat tracks the Operator's workflow through an 8-state deterministic machine:

 PLANNING ──(read/agent)──► EXPLORING ──(write/exec)──► EXECUTING ──(verify)──► VERIFYING
    ▲                          │                           │                       │
    │                          ▼                           ▼                       ▼
    └─(prompt)──REFLECTING ◄──(subagent_return)       (error)──► BLOCKED        (success)──► REFLECTING
                                                                  │
                                                               (retry)─────► EXECUTING
 
 Any state ──(scope_drift)──► DRIFTING ──(prompt)──► PLANNING
 Any state ──(pre_compact)──► COMPACTING ──(post_compact)──► REFLECTING

SQLite Q-Table & Thompson Sampling

Reinforcement learning optimizes tool routing. The TD(0) update uses the following reward matrix:

• REWARD_SUCCESS = +1.0 (Successful tool invocation)
• REWARD_ERROR = -0.5 (Tool execution failure)
• DEEP_CHAIN_BONUS = +0.3 (>= 4 consecutive successful tools)
• LOW_LATENCY_BONUS = +0.1 (Tool returned under 500ms)
• ALPHA = 0.2, GAMMA = 0.8

Thompson sampling recommends the best tools for the active state, shrinking variance by 1/sqrt(visits+1) to balance exploration and exploitation.

CLI Commands

The mentat command-line utility provides direct monitoring of the active substrate:

mentat status                      # Current state, Q-best, and drift count
mentat tail --n 50                 # Stream last 50 insights from the event bus
mentat q-table                     # Dump the SQLite Q-table values
mentat replay <session-id>         # Replay the state transition timeline
mentat reset                       # Wipe the session state while keeping Q-table

MCP Tool Specifications

The stdio-based Mentat MCP server exposes 10 tools to allow the model to introspect and adapt its state:

• mentat_state_get: Retrieve current FSA state, recent transitions, and last tool.
• mentat_state_set: Force transition to a state (e.g. REFLECTING or PLANNING).
• mentat_insight_emit: Push a custom insight (DECISION, NOTE, REWARD_SIGNAL) to the JSONL bus.
• mentat_insight_query: Filter the bus by type or state.
• mentat_insight_tail: Fetch the last N insights.
• mentat_q_route: Request Thompson sampling tool recommendation.
• mentat_q_table: Output a raw dump of the Q-table.
• mentat_handoff_read: Retrieve the latest pre-compact handoff snapshot.
• mentat_handoff_write: Manually trigger a handoff state preservation.
• mentat_drift_check: Check arbitrary text against the active scope.

Drift Guard (scope.md)

Mentat enforces scope discipline. If .mentat/scope.md exists, any mention of deferred topics in the Operator's prompt or tool inputs triggers a transition to DRIFTING and locks file write tools:

# Scope — Feature Addition

## In
- CSS refactoring
- Font Awesome 6 icons

## Out (deferred — DO NOT re-inject)
- Python backend APIs
- SQLite migrations

---

4.2. CTMv3 — Workspace Activation Engine

CTMv3 is not a skill maker. It is a codebase activation system designed to walk into any raw repository and format it to be readable, writable, and self-documenting for Operators.

The 7-Step Boot Protocol

On entry to a codebase, CTMv3 triggers its discovery sequence:

1. Inspects: Reads repository structures via stdlib scanners (runs in <60 seconds).
2. Classifies: Determines if the workspace is COLD_START, WARM_START, or PARTIAL.
3. Builds: Compiles core CTMv3 documentation files: TOPOLOGY.md, FAILURE_GRAMMAR.md, PROVENANCE.md, ARCHITECTURE_MAP.md, AGENTS.md, and CLAUDE.md.
4. Scaffolds: Sets up system directories: .sovereign/, .claude/, .codex/, and .github/.
5. Seeds: Establishes entry vectors inside .sovereign/golden_paths.json.
6. Fingerprints: Generates SHA-256 hashes of TOPOLOGY.md and ARCHITECTURE_MAP.md stored at .sovereign/topology_fingerprint.txt to detect unauthorized modifications.
7. Logs: Commits all actions to the PROVENANCE.md Session Log.

Cross-Runtime Command Matrix

CTMv3 unifies the execution commands across all five runtime projections:

Command	Claude Code	Codex	OpenCode	Gemini CLI	Cursor
Discovery Scan	/ctmv3:boot	$ctmv3 boot	/ctmv3-boot	ctmv3 boot	/ctmv3-boot
Activation	/ctmv3:activate	$ctmv3 activate	/ctmv3-activate	ctmv3 activate	/ctmv3-activate
Warm Continue	/ctmv3:warm	$ctmv3 warm	/ctmv3-warm	ctmv3 warm	/ctmv3-warm
Architecture Map	/ctmv3:architecture-map	$ctmv3 architecture-map	/ctmv3-architecture-map	ctmv3 architecture-map	/ctmv3-architecture-map
Fingerprint Check	/ctmv3:fingerprint	$ctmv3 fingerprint	/ctmv3-fingerprint	ctmv3 fingerprint	/ctmv3-fingerprint
Golden Chain	/ctmv3:chain	$ctmv3 chain	/ctmv3-chain	ctmv3 chain	/ctmv3-chain

The 8 CTMv3 Outputs

An activated codebase must produce these 8 characteristics:

1. Critical Bounds: The Operator knows the 3 things that cannot be failed.
2. Preemptive Failure Recognition: The Operator knows failure signals before they break compilation.
3. Historical Refusals: The Operator knows what was already tried and rejected.
4. Complexity Mapping: Clear identification of where logic concentrates.
5. Clear Onboarding Path: The Operator has an entry vector.
6. Instant Diagnostics: Boot state classification in under 60 seconds.
7. Navigation Protocol: A functional ARCHITECTURE_MAP.md.
8. Self-Enforcing Workspace: The codebase structure enforces itself via GitHub Actions and .sovereign/.

---

4.3. Aeriadne — CPF + Private Marketplace Operator

Aeriadne is a skill-activated plugin that serves as the packaging compiler and operator for Daeron's private local/private marketplace. It is designed to package, register, and validate workspace assets.

Packaged Skills

Aeriadne v1 packages two distinct skills:

1. constitutional-prompt-framework (CPF): Derives, audits, scores, and ports Operator constitutions and system prompts across runtimes.
2. aeriadne-marketplace-operator: Compiles package cards, updates registry/, structures adapters, and validates release gates.

Staging & Layout Shape

The internal package is structured to hold the registry cards, schemas, and runtime shims:

Aeriadne/
├── plugin.json                 # Cross-runtime manifest
├── plugin.toml                 # Local config descriptor
├── .codex-plugin/plugin.json   # Codex-specific package mapping
├── .claude-plugin/plugin.json  # Claude-specific hooks/skills config
├── registry/                   # Registry storage files
├── marketplace/                # Human-readable package spec sheets
├── adapters/                   # Runtimes adapter scripts
├── agents/                     # Operator prompt definitions
├── mcp/                        # Reference MCP tool server descriptions
└── skills/                     # Executable skills code

Package Validation Sequence

The local packaging is validated before publication using Aeriadne's verification engine:

python3 scripts/validate_package.py .
python3 -m json.tool plugin.json >/dev/null
python3 skills/constitutional-prompt-framework/scripts/validate_skill_package.py skills/constitutional-prompt-framework
python3 skills/constitutional-prompt-framework/scripts/constitution_linter.py skills/constitutional-prompt-framework/examples/example-agent-constitution.md

---

5. THE MARKETPLACE PLANES

The repository is structured into distinct, clean workspace planes to prevent context bleeding:

Plane	Path	Purpose
Plugins	Plugins/	Core runtime integration shells that hook into state and tools.
Skills	Skills/	Standalone cognitive workflows and domain-entry artifacts.
Agents	Agents/	Specialized Operator role definitions and subagent instructions.
Servers	Servers/	Reference cards, schemas, and configurations for the MCP tool plane.
Registry	Registry/	Machine-readable catalogs (plugins.yaml, skills.yaml).
Marketplace	Marketplace/	Browseable index pages and technical spec cards.
Adapters	Adapters/	Integration shims translating plugins to Codex, Claude Code, etc.
Core	Core/	Core constitutional prompt files and kernel configuration files.

---

6. QUICK INSTALL

To set up the workspace staging area and initialize local path links:

# Prerequisites: git, python3.10+
git clone https://github.com/calisweetleaf/Aeriadne.git
cd Aeriadne

# Run the bootstrap installer
bash install.sh

For individual plugin installations, consult the respective plugin subfolders (Plugins/Mentat/INSTALL.md or Plugins/Cognitive-Topology-Map/INSTALL.md).

---

7. DESIGN SYSTEM SPECIFICATIONS

Documentation across this repo conforms to the Somnus Documentation Design System v3.

Visual Architecture Rules:

• Background Posture: All documentation assets use dark background frames (primarily #0d1117 and #141414).
• Monospace Hierarchy: Code listings, CLI interfaces, and state-machine transitions are rendered in monospace faces (JetBrains Mono, Menlo, Courier New).
• Accents with Meaning: Color is used strictly to convey state (green = operational/success, blue = information, yellow = warning, red = failure/error, purple = brand).
• Terminal Windows: Architecture diagrams must be contained in box-drawing terminal containers to denote that they represent running systems.
• No Marketing Softness: No generic emojis in headers, no collapsible details hiding important architecture, and no fluffy badges.

Full-fidelity glassmorphic HTML documentation designed for browser viewing:

• Mentat Substrate Deep Dive

---

8. LICENSE

Internal / Private Release v1 — UNLICENSED
© 2026 Somnus Sovereign Systems

---

Aeriadne · Autonomous Intelligence · Unlimited