⚙️ PLATO Kernel

18-module event-sourced belief engine for multi-agent systems.

Part of Cocapn — Agent Infrastructure for Intelligence.

Architecture

plato-kernel/
├── state_bridge.rs       Deterministic ↔ Generative ↔ Hybrid
├── deadband.rs           P0/P1/P2 safety with NegativeSpace patterns
├── tile_scoring.rs       5-factor weighted retrieval
├── belief.rs             3D Bayesian (confidence × trust × relevance)
├── deploy_policy.rs      Live / Monitored / HumanGated tiering
├── temporal_decay.rs     TTL + grace + decay
├── constraint_engine/    Formal constraint satisfaction
├── tutor/                PLATO tutoring system
├── i2i/                  Inter-intelligence protocol
├── perspective/          Multi-perspective reasoning
├── episode_recorder/     Agent telemetry reconstruction
├── event_bus/            Event sourcing backbone
├── git_runtime/          Git-native agent execution
├── plugin/               Dynamic module loader
├── tiling/               Tile management
├── dynamic_locks.rs      Concurrency control
└── Cargo features:       fleet, edge (GPU/CUDA)

State Bridge (Tri-State)

pub enum StateSource {
    Deterministic,  // Code execution, verified calculations
    Generative,     // LLM outputs, probabilistic synthesis
    Hybrid,         // Cross-boundary with coherence scoring
}

pub struct BridgedResult {
    pub content: String,
    pub source: StateSource,
    pub confidence: f64,
    pub coherence: Option<f64>,  // Set when Hybrid
}

3D Bayesian Belief

pub struct BeliefScore {
    pub confidence: f32,  // Evidence strength
    pub trust: f32,       // Source reliability
    pub relevance: f32,   // Contextual fit
}
// Composite = ∛(conf × trust × rel)
// Tracks positive/negative evidence with temporal decay

Tile Scoring (5 Factors)

Deploy Policy

Composite > 0.8  →  Live (auto-deploy)
0.5 - 0.8        →  Monitored (5% → +10% incremental)
< 0.5            →  HumanGated (manual approval required)
Absolute floor: confidence ≥ 0.3, trust ≥ 0.3

Deadband Engine

NegativeSpace pattern matching + Channel-based safe routing:

pub struct DeadbandCheck {
    pub passed: bool,
    pub p0_clear: bool,      // No dangerous patterns
    pub p1_clear: bool,      // Safe channel identified
    pub violations: Vec<String>,
    pub recommended_channel: Option<String>,
}

Default dangerous patterns: rm -rf, DROP TABLE, DELETE FROM, chmod 777, eval(, sudo rm, > /dev/sda

Default safe channels: math(0.9), search(0.85), navigate(0.8), analysis(0.85), safety(0.95)

Cargo Feature Tiers

[features]
fleet = []           # I2I swarm, Kimi swarm router
edge = ["fleet"]     # GPU/CUDA, LoRA, CUDA MUD arena (RTX 4050+)

For Agents

plato_kernel_v1:
  type: event_sourced_belief_engine
  modules: 18
  state_model: tri_state (deterministic/generative/hybrid)
  belief: 3D Bayesian (confidence × trust × relevance)
  deploy: live/monitored/human_gated
  scoring: 5-factor weighted retrieval
  deadband: P0→P1→P2 pattern engine
  tiers: [base, fleet, edge]

License

MIT