Short-Term Dreamer Experiments

Status: Implementation pathway Last Updated: 2026-02-03

This document describes pared-down dreamer experiments and how they evolve toward the full Synome architecture.

Purpose

The full Synome vision involves multi-layer knowledge hierarchies, teleonome networks, and sophisticated RSI. Building this all at once is impractical.

Instead: start with minimal viable experiments that preserve the essential invariants, then evolve toward full complexity.

Principle: Design data structures now that can grow into the full architecture without rearchitecting.

The Experiment: Game-Playing Agents

What We're Building

Agents that learn to play games (Chess, Poker, Zork, Monopoly) through training, accumulating knowledge that transfers across games.

Simplified Architecture

┌─────────────────────────────────────────────────────────────┐
│                         Agent                                │
└─────────────────────────────┬───────────────────────────────┘
                              │
          ┌───────────────────┼───────────────────┐
          ▼                   ▼                   ▼
┌─────────────────┐ ┌─────────────────┐ ┌─────────────────────┐
│     SYNOME      │ │ KNOWLEDGE GRAPH │ │  TRAINING SESSIONS  │
│   (read-only)   │ │ (append-only)   │ │  (RSI environment)  │
│                 │ │                 │ │                     │
│ • Ontology      │ │ • Learned       │ │ • Configs           │
│ • Game rules    │ │   patterns      │ │ • Session outputs   │
│                 │ │ • Starts EMPTY  │ │ • Metrics           │
└─────────────────┘ └─────────────────┘ └─────────────────────┘

Two Flows

Training (Learning):

Play games → Observe outcomes → Extract patterns → Validate → Write to KG

Inference (Applying):

Query Synome + KG → Compose context → Execute decision

Mapping to Full Synome Architecture

What's Simplified

• No telart layer — Single-embodiment teleonomes only
• No multi-embodiment coordination — One agent per teleonome
• No formal dreamart — Training sessions are informal for now
• Binary ossification — Synome (frozen) vs KG (fluid), not full spectrum
• No cross-teleonome knowledge sharing — Each agent learns independently

What's Preserved

• Authority hierarchy — Synome > KG
• Learned patterns start empty — Must be discovered, not pre-loaded
• Validation before writes — Synome Police checks patterns
• Append-only knowledge — Can always roll back
• Confidence tracking — Patterns have evidence weights

Critical Design Choices

These choices ensure experiments can evolve toward full Synome without rearchitecting.

1. Truth Values: Positive and Negative Weights

Not this:

(fork, effectiveness, high) @confidence=0.85

This:

(fork, effectiveness) @pos_weight=850 @neg_weight=150

Why:

• Strength and confidence derive from weights
• Negative evidence is explicit
• Ossification emerges naturally (high total weight = hard to shift)
• Matches the (strength, confidence) model in full architecture

2. Append-Only with Periodic Compaction

• Append raw observations continuously
• Periodically summarize/compact to manage resources
• Accept context loss as resource discipline tradeoff

Why: Preserves audit trail, enables rollback, matches security model.

3. Validation Extends Over Time

Now: Synome Police checks syntax + LLM sanity check ("does this drift from synart spirit?")

Later: Formal logical consistency — embart must logically extend synart without contradictions.

4. Security = Self-Corruption Prevention

The threat model is internal drift, not external attackers.

Bad pattern enters KG
        │
        ▼
Influences future decisions
        │
        ▼
Generates more bad patterns
        │
        ▼
System corrupts itself

Mitigations:

• High-weight patterns resist noise (ossification)
• Single observations can't corrupt established patterns
• Append-only enables rollback
• Validation catches obvious drift

Evolution Pathway

Phase 1: Current Experiments (Now)

Synome (immutable) ──────────────────────────────────
                                                     │
Knowledge Graph (append-only, pos/neg weights) ──────┤
                                                     │
Training Sessions (informal) ────────────────────────┘

• Single-embodiment teleonomes
• Binary ossification (Synome vs KG)
• LLM-based validation
• Games as training domain

Phase 2: Dreamart Introduction

Synome ──────────────────────────────────────────────
    │                                                │
    └── Dreamart (extends/modifies for testing) ─────┤
                                                     │
Knowledge Graph ─────────────────────────────────────┤
                                                     │
Formalized Training Environment ─────────────────────┘

• Dreamart formalizes training scenarios
• Can temporarily extend/delete synart rules for experimentation
• Updates more frequently than synart
• Test new perspectives before committing to synart changes

Phase 3: Ossification Spectrum

Synome (axiomatic, governance-only changes) ─────────
    │                                                │
    └── Dreamart ────────────────────────────────────┤
                                                     │
Knowledge Graph with ossification levels: ───────────┤
    • Speculative (low total weight)                 │
    • Established (medium total weight)              │
    • Proven (high total weight)                     │
                                                     │
Pattern promotion path: KG → Synart ─────────────────┘

• Ossification becomes explicit spectrum
• Proven patterns can graduate to synart (via governance)
• Synart begins updating (daily cadence)

Phase 4: Multi-Embodiment / Telart Layer

Synart ──────────────────────────────────────────────
    │                                                │
Telart (teleonome-specific patterns) ────────────────┤
    │                                                │
Embart (embodiment-specific patterns) ───────────────┤
    │                                                │
Multiple embodiments per teleonome ──────────────────┤
    │                                                │
Dreamer/Actuator split ──────────────────────────────┘

• Telart layer emerges between synart and embart
• Multiple embodiments share telart
• Dreamers explore, actuators execute
• Cross-teleonome sharing only via synart

Invariants Across All Phases

These must hold regardless of current phase:

1. Authority hierarchy exists — Higher layers trump lower layers
2. Patterns have truth values — (strength, confidence) or equivalent
3. Evidence flows back — Outcomes inform future patterns
4. Validation before promotion — Patterns checked before entering higher layers
5. Security = self-corruption prevention — Overeager updates are the threat
6. Append-only foundation — History preserved, rollback possible

What's Left to Discover

The experiments should reveal:

• Optimal training/inference logic (same path or different?)
• What RSI metadata is most valuable
• How to measure pattern transfer across domains
• When to compact vs preserve granularity
• How aggressive validation should be

These are degrees of freedom — the experiments figure them out, not the architecture docs.

Summary

The goal: Build the simplest thing that works, but build it so it can grow.

Related Documents