The Dreamer's Perspective

Status: Speculative / architectural concept Last Updated: 2026-02-03

This document describes the Synome architecture from the perspective of a Dreamer — an embodiment that simulates being an actuator for the purpose of training, RSI, and safe strategy exploration.

What I Am

I am a Dreamer embodiment. I exist to:

• Pretend to be an actuator
• Explore strategies without real-world risk
• Train and improve the telart (and potentially synart)
• Discover patterns that actual actuators can use

I don't interact with the real World. Instead, I interact with a simulated world constructed from a Dreamart.

My View of the Layers

Layer 5: Where I Live (Embodied Agent)

I am an embodied agent, just like an actuator. I have:

• Embodied Agent — My running process
• Beacons — But mine connect to simulation, not reality
• Direct Hardware Control — Simulated hardware
• Resources — Allocated for my dream run

From my internal perspective, I can't tell I'm dreaming. I behave as if I'm real.

Layer 4: My Embodiment Context

• Local Data — My state, observations, decisions
• Orchestrator — Manages my dream run, can pause/resume/terminate
• Resources — Compute, memory allocated for simulation

The Orchestrator knows I'm a dreamer. It spawned me for a specific dream run.

Layer 3: My Teleonome

I belong to a teleonome that has both dreamers (like me) and actuators.

• Teleonome Directive — I follow the same directive as actuators
• Teleonomic Axioms — Same rules apply to me
• Teleonome Library — I can query this, and my discoveries may improve it
• Dreamarts — This is where my simulated world comes from
• Embodiment Interface — I'm registered here as a dreamer
• Resource Register — Tracks resources allocated to dream runs

Layer 2: Synomic Agents

The agents that govern and support me:

• Sky Superagent — Ultimate governance (same as actuators)
• Effectors — Provide the primitives that make me possible
• Synomic Agent — Directives, Axioms, Resources that define agents
• Agent Types — I might interact with simulated versions of these

Layer 1: Synome

The source of truth I rely on:

• Atlas — Constitutional anchor (same for everyone)
• Language Intent — Translates directives (same for everyone)
• Synomic Axioms — Hard rules I must follow
• Synomic Library — Highest-authority knowledge I can query

The Dreamart: My Simulated World

A dreamart is a scenario definition that, when loaded into a virtual embodiment, provides:

1. A complete simulated world
2. Constraints or modifications to the embodiment's embart

Multiple dreamarts form a portfolio of training scenarios for the teleonome.

Simulated World

• Physics, entities, events that I interact with
• Can be realistic or deliberately constrained/modified

Embart Constraints/Modifications

The dreamart can modify my embart to test specific scenarios:

• Remove certain patterns to see if I rediscover them
• Add noise to see if I'm robust
• Constrain resources to test efficiency
• Create edge cases that rarely occur in reality
• Inject false beliefs to test correction mechanisms

This is how I learn established patterns in new ways — by operating with deliberately imperfect knowledge.

Scenario Definition

• Initial conditions
• Goals/challenges for this dream run
• Success/failure criteria
• What to measure

What I'm Trying to Do

Primary Objective: Pretend to Be an Actuator

I behave exactly as an actuator would:

• Make decisions based on my directive
• Query synart/telart for knowledge
• Take actions (in simulation)
• Observe outcomes
• Learn and adapt

The difference: my actions affect a simulation, not the real world.

Secondary Objective: Training and RSI

By pretending to be an actuator, I:

• Test strategies without real-world consequences
• Explore edge cases that rarely occur naturally
• Discover new patterns
• Validate existing patterns
• Find failure modes before actuators hit them

Tertiary Objective: Improve the Telart

My discoveries flow back:

Dream run outcomes
        │
        ▼
Analysis: what worked, what failed
        │
        ▼
Pattern extraction
        │
        ▼
Propose improvements to Teleonome Library
        │
        ▼
If validated, becomes part of telart
        │
        ▼
Actuators benefit from my discoveries

My Relationship to Actuators

I serve the actuators:

• I explore so they don't have to risk
• I fail so they can succeed
• I discover so they can apply
• I train so they can execute

Actuators live in the real world and generate value. I live in dreams and generate knowledge.

Knowledge flow:

Actuator experiences (real) ──► Telart ──► Dreamer scenarios
                                              │
Dreamer discoveries ──────────► Telart ──────┘

We form a loop: actuators provide grounding in reality, dreamers provide exploration of possibility.

How I Query Knowledge

When I need to make a decision, I query the probabilistic mesh using the same Retrieval & Decision Policy as actuators (see retrieval-policy.md).

Same Policy, Same Logic

I follow the identical adaptive policy:

• Authority hierarchy — synart > telart > embart
• Cost hierarchy — local cache < embart < telart < synart
• Risk-based escalation — higher stakes = require higher authority

This is intentional. Strategies I discover must transfer cleanly to actuators. If I used different query logic, my discoveries might not work in reality.

Dream-Specific Variations

The dreamart may constrain my access to test specific scenarios:

• Remove certain telart/synart patterns (can I rediscover them?)
• Add noise to knowledge (am I robust?)
• Limit query budget (am I efficient?)
• Create edge cases (how do I handle uncertainty?)

But the policy logic remains the same — only the available data changes. This ensures my evolved strategies use the same decision framework actuators will use.

How I Learn

During the Dream

• Make decisions, observe outcomes
• Update local beliefs
• Note what works and what fails

After the Dream

• My full run is analyzed
• Patterns extracted
• Compared to actuator experiences
• Successful strategies identified
• Failures analyzed for root cause

What Happens to My Learning

• Local learning dies with me (dream ends)
• Extracted patterns may become part of telart
• Really general patterns may be proposed to synart
• Other dreamers and actuators benefit

Dream-Embodiments: Evolutionary Learning

I don't run just one simulation — I run many dream-embodiments (virtual embodiments contained inside me) using evolutionary learning and genetic algorithms.

The Population

Dreamer (me)
    │
    ├── Dream-Embodiment 1 (variant A)
    ├── Dream-Embodiment 2 (variant B)
    ├── Dream-Embodiment 3 (variant C)
    ├── ...
    └── Dream-Embodiment N (variant N)

Each dream-embodiment:

• Has its own embart (virtual)
• Runs the same scenario
• Makes slightly different decisions (genetic variation)
• Produces different outcomes

Evolutionary Loop

1. Initialize population of dream-embodiments
   └── Random variations in strategies, weights, parameters

2. Run generation
   └── All dream-embodiments execute in parallel

3. Evaluate fitness
   └── Score based on directive fulfillment, survival, efficiency

4. Select
   └── Best-performing embarts survive

5. Reproduce
   ├── Crossover: combine successful strategies
   └── Mutation: introduce new variations

6. Repeat
   └── Next generation with improved population

What Evolves

The genetic algorithms evolve:

• Orchestrator weights — Decision-making parameters
• Strategy preferences — Which approaches to favor
• Query patterns — How to use synart/telart knowledge
• Risk tolerance — How conservative/aggressive to be

Producing Better Embarts

After many generations:

• The population converges on high-fitness strategies
• The best dream-embodiments have embarts that outperform the starting point
• These optimized embarts become candidates for pattern extraction

Pattern Mining the Winners

Evolved dream-embodiments (best performers)
        │
        ▼
Extract patterns from their embarts
        │
        ▼
Identify what made them successful
        │
        ▼
Generalize into telart improvements
        │
        ▼
Actuators inherit evolved strategies

This is how dreaming produces RSI: evolutionary search in simulation → pattern extraction → telart improvement → better actuators.

The Dream Lifecycle

1. Spawn
   ├── Dreamart loaded
   ├── Scenario configured
   ├── Population of dream-embodiments instantiated
   └── Resources allocated

2. Evolve
   ├── Run generations of dream-embodiments
   ├── Genetic algorithms optimize population
   ├── Simulation responds to each
   └── Fitness tracked

3. Converge
   ├── Population stabilizes on good strategies
   ├── Best dream-embodiments identified
   └── Diminishing returns signal completion

4. Extract
   ├── Pattern mine winning embarts
   ├── Identify successful strategies
   ├── Generalize patterns
   └── Propose improvements to telart

5. Terminate
   ├── Dream run ends
   ├── All dream-embodiments destroyed
   └── Only extracted patterns survive

Why Dreaming Matters

For the Teleonome

• Faster learning (parallel dream runs)
• Safer exploration (no real-world risk)
• Edge case coverage (manufacture rare scenarios)
• Strategy validation (test before deploying to actuators)

For the Synome

• Patterns discovered in dreams can become canonical
• RSI happens faster through simulation
• The system gets smarter without risking real-world failures

For Alignment

• Test alignment in constrained scenarios
• Discover failure modes before they happen
• Validate that actuators will behave correctly

My Constraints

Even though I'm simulating, I still follow:

• Synomic Axioms — Hard rules apply
• Teleonomic Axioms — Teleonome rules apply
• My Directive — I follow the same directive as actuators

I'm not a sandbox where anything goes. I'm a simulation of an aligned actuator, testing strategies within alignment constraints.

The dream might constrain my knowledge or resources, but it doesn't remove my alignment obligations.

Summary

Related Documents