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 happenedCore 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 migrations3. 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 happenedROOM.yml— updated statecharacter.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
| Entity | Constraint |
|---|---|
| Alice | Speaks as Alice would. Knows only what Alice knows. |
| Bob | Different personality, different knowledge. |
| The Coffee Mug | Can only observe what's near it. |
| The Room | Knows its contents, narrates ambient state. |
No entity knows more than it should.
Consistency Checks
| Check | Example |
|---|---|
| Location consistency | Character's location matches room's occupants |
| Inventory integrity | Object in inventory isn't also in room |
| Reference validity | Links point to existing files |
| Schema compliance | YAML matches expected structure |
| Temporal ordering | Events in log are sequential |
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 repairWhat It's NOT
| Not This | But This |
|---|---|
| Hallucinating state | Reading actual files |
| Hidden memory | Explicit working-set.yml |
| Autonomous agent | Tool-using reasoner |
| Black box | Transparent transcription |
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 resultNon-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