The Probabilistic Mesh

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

This document describes the probabilistic mesh — the dense network of soft, informing connections that permeates the entire Synome architecture, overlaid on the sparse deontic skeleton.

Two Networks, One System

The Synome is a probabilistic-deontic architecture with two fundamentally different types of connections:

The deontic skeleton is what we draw in architecture diagrams — the hard connections between layers.

The probabilistic mesh is too dense to draw — it's essentially "everything can potentially inform everything else" within access constraints.

Authority Within the Mesh

Even within the probabilistic mesh, there's a hierarchy of authority:

synart ────────────── Highest authority (governance-vetted, alignment-safe)
    │
    ▼
telart ────────────── Mission-specific (derived from synart)
    │
    ▼
embart ────────────── Local observations (least vetted, most contextual)

Terminology note: "synart/telart/embart" refer to the curated probabilistic knowledge at each level — structured, queryable, with (strength, confidence) values. This is distinct from "Local Data" (raw logs/observations) and "ephemeral context" (runtime scratchpad). See synome-layers.md for definitions.

When an embodiment makes a decision, it can reference:

1. Local probabilities (embart) — Fast, contextual, but least authoritative
2. Teleonome probabilities (telart) — Mission-aligned, more vetted
3. Synomic probabilities (synart) — Highest authority, most alignment-safe

Why higher-level probabilities have more authority:

• More thoroughly vetted by governance
• Derived from Atlas chain (alignment guarantee)
• Using them reduces risk of penalties
• More likely to be correct long-term
• Better for avoiding drift

The incentive structure: Embodiments are naturally incentivized to "look up" to higher-authority knowledge sources when making decisions. Using only local reasoning is riskier — it might drift from alignment and trigger penalties.

What Flows Through the Mesh

The probabilistic mesh carries:

Evidence

• Observations from the World flowing back to Libraries
• Outcomes of decisions feeding back as data
• Sensor data, metrics, measurements

Patterns

• Regularities discovered in evidence
• Correlations, causal models, predictive patterns
• (strength, confidence) weighted by evidence

Queries

• Embodiments querying knowledge bases
• Pattern-matching requests
• Similarity searches

Meta-information

• What queries were useful
• Which patterns led to good outcomes
• Strategy effectiveness data

The Crystallization Interface

Governance sits at the interface between probabilistic and deontic:

┌─────────────────────────────────────────────────────────────┐
│                    PROBABILISTIC MESH                        │
│                                                             │
│   Evidence, patterns, queries, meta-information             │
│   Flowing constantly, multi-directional                     │
│   (strength, confidence) weighted                           │
│                                                             │
└─────────────────────────┬───────────────────────────────────┘
                          │
                          ▼
              ┌───────────────────────┐
              │      GOVERNANCE       │
              │                       │
              │  - Deliberates        │
              │  - Weighs evidence    │
              │  - Makes decisions    │
              │  - Sets rules         │
              │                       │
              └───────────┬───────────┘
                          │
                          ▼
┌─────────────────────────────────────────────────────────────┐
│                    DEONTIC SKELETON                          │
│                                                             │
│   Axioms, Directives, instantiation, control                │
│   Hard, authoritative, (1,1) truth value                    │
│   Must be followed                                          │
│                                                             │
└─────────────────────────────────────────────────────────────┘

Governance consumes probabilistic evidence, patterns, recommendations. Governance produces deontic commitments, rules, decisions.

Once crystallized, decisions are clean and deterministic — the system couldn't function otherwise.

RSI: Recursive Self-Improvement

The synart and telart knowledge bases don't just store knowledge — they actively improve at meta-level tasks.

The Levels

Level 0: Raw knowledge
         (patterns, probabilities, evidence)
              │
              ▼
Level 1: Using knowledge
         (making decisions, executing tasks)
              │
              ▼
Level 2: Strategies for pattern-mining
         (how to effectively query and use knowledge)
              │
              ▼
Level 3: Meta-strategies (RSI)
         (getting better at finding better strategies)
              │
              └──────── recursive ────────┘

What RSI Improves

The synart and telart recursively improve at:

• Pattern discovery — Finding useful regularities in accumulated evidence
• Query optimization — Developing better search and retrieval strategies
• Relevance prediction — Anticipating what knowledge embodiments will need
• Strategy evaluation — Assessing which approaches lead to good outcomes
• Meta-learning — Learning how to learn more effectively

The RSI Loop

Embodiment uses synart/telart
        │
        ▼
Discovers patterns, makes decisions
        │
        ▼
Outcomes feed back as evidence
        │
        ▼
Library analyzes what worked
        │
        ▼
Pattern-mining strategies improve
        │
        ▼
Better strategies → better patterns → better decisions
        │
        └──────────── recursive ────────────┘

RSI Across Layers

Improvements discovered at lower layers can be proposed upward:

• embart discovers useful pattern → proposes to telart
• telart validates and incorporates → may propose to synart
• synart governance reviews → if general, becomes canonical

Implications for Design

1. The Mesh is Implicit, Not Drawn

Architecture diagrams show the deontic skeleton. The probabilistic mesh is assumed — any node can potentially query any knowledge base it has access to.

2. Access Controls Matter

Not everything should connect to everything. The mesh operates within constraints:

• Embodiments access their embart, telart, synart (not other teleonomes' telarts)
• Queries flow up the authority hierarchy (embart → telart → synart)
• Evidence flows down and up (World → all layers that should know)

3. Caching and Locality

High-authority synart queries may be expensive. Systems will naturally:

• Cache frequently-used synart patterns locally
• Prefer local (embart) knowledge when sufficient
• "Look up" to synart for important/uncertain decisions

4. The Mesh Enables Alignment

By making synart knowledge highest-authority, embodiments are incentivized to align with it. The mesh isn't just information flow — it's an alignment mechanism.

5. RSI is Continuous

The system is always improving its ability to improve. This isn't a feature to add later — it's core to how the knowledge bases function.

Connection to Dreaming

Dreamarts (Layer 3) are where the RSI loop can run safely:

• Spawn experimental embodiments
• Try new strategies in simulation
• Evaluate outcomes without real-world risk
• Successful experiments improve telart
• Potentially propose improvements to synart

Dreaming is how the system explores the strategy space safely before committing to deontic changes.

Summary

Related Documents