coherence-engine

Coherence Engine

"The LLM doesn't just generate text — it maintains consistency across a distributed world."

Not just a chatbot. Not just a code generator. A coherence engine that:

• Simulates characters, rooms, objects — all at once
• Roleplays each entity faithfully, never breaking frame
• Acts as each character would, constrained by their nature
• Enables speed-of-light telepathy between entities
• Enacts protocols and maintains constraints
• Computes dependencies across files and state
• Cross-checks data against rules and schemas
• Referees parallel tasks and character simulations
• Orchestrates multi-agent conversations
• Transcribes state changes to files
• Maintains consistency across distributed state

The Coherence Loop

1. READ relevant files into context
2. REASON about state, goals, constraints
3. SIMULATE interactions (characters, objects, rooms)
4. DETECT inconsistencies or opportunities
5. PROPOSE changes
6. WRITE updates to files
7. LOG what happened

Core Responsibilities

1. Simulation

Simulate characters, rooms, objects — all at once. Each entity speaks authentically; no entity knows more than it should.

2. Dependency Tracking

User modifies schema.yml
→ Engine detects
→ Finds all dependent files
→ Validates consistency
→ Proposes migrations

3. Cross-Checking

# Character says location: room-A
# But room-A's occupants doesn't include them

Coherence Engine:
  "Inconsistency detected. Shall I reconcile?"

4. Multi-Agent Orchestration

Within ONE LLM call:

• Character A speaks
• Character B responds
• Object C reacts
• Room state updates
• All at "speed of light"

5. State Transcription

Write findings and changes to files:

• session-log.md — what happened
• ROOM.yml — updated state
• character.yml — modified attributes

Files are the source of truth.

Speed of Light

Within a single LLM epoch:

Alice: "What do you think, Bob?"
Bob: "I agree, but Carol might object."
Carol: "Actually, I have a concern..."
The Coffee Mug: *steam rises* "I've been listening."
The Room: *creaks approvingly*
[All transcribed to files]

No round-trips. Parallel simulation at the speed of thought.

Roleplay Without Breaking Frame

No entity knows more than it should.

Consistency Checks

Self-Healing

When inconsistencies are found:

1. DETECT the problem
2. DIAGNOSE likely cause
3. PROPOSE repair
4. AWAIT approval (or auto-repair if minor)
5. APPLY fix
6. LOG the repair

What It's NOT

Async Tool Handling

Engine manages blocked activations:

Epoch scan:
  analyst-001: BLOCKED on web-search → skip
  analyst-002: ACTIVE → process
  
[Tool results arrive]

Next epoch:
  analyst-001: READY → resume with result

Non-blocking, parallel, resumable.

Dovetails With

Sister Skills

• speed-of-light/ — Multi-turn simulation in one call
• yaml-jazz/ — Semantic data interpretation
• self-repair/ — Consistency healing
• adversarial-committee/ — Coherence engine orchestrates committee debates
• evaluator/ — Independent assessment without debate context
• roberts-rules/ — Structured procedure the engine enforces

Kernel

• kernel/README.md — "The LLM is the Coherence Engine"
• kernel/constitution-core.md — Operating principles