Skip to content

Architecture

ArqonHPO is built as a control primitive, not a search algorithm. This page explains how the pieces fit together.

System Overview

graph LR
    subgraph "ArqonHPO Runtime"
        A[Telemetry] --> B[Probe]
        B --> C[Classify]
        C --> D[Refine]
        D --> E[Safety Executor]
        E --> F[Apply]
    end

    F --> G[Your System]
    G --> A

The PCR Loop

ArqonHPO uses a Probe → Classify → Refine architecture:

1. Probe (Tier 0)

Goal: Efficiently scan the parameter landscape.

  • Uses Low-Discrepancy Sequences (LDS) for uniform coverage
  • Prime-sqrt-slopes rotation for high-dimensional spaces
  • Stateless, deterministic, shardable via ArqonProbe

2. Classify (Tier Ω)

Goal: Detect whether the landscape is structured or chaotic.

  • Analyzes probe results for patterns
  • Determines optimal strategy for refinement
  • Detects: smooth gradients vs. noisy/multimodal surfaces

3. Refine (Tier 2)

Goal: Converge to optimal parameters.

Strategies:

  • Nelder-Mead — Simplex method for smooth functions
  • Multi-Start Nelder-Mead — Parallel restarts for multimodal
  • TPE — Bayesian optimization for noisy landscapes

Safety Layer (Hotpath)

The Safety Executor enforces governance on every parameter update:

graph TD
    A[Proposal] --> B{Guardrails Check}
    B -->|Pass| C{Rate Limit Check}
    B -->|Fail| D[Violation]
    C -->|Pass| E{Bounds Check}
    C -->|Fail| D
    E -->|Pass| F[Apply Atomically]
    E -->|Fail| D
    F --> G[Audit Log]
    D --> G

Key Components

Component Responsibility
SafetyExecutor Enforces all safety rules
Guardrails Configures bounds, rate limits, delta limits
RollbackPolicy Defines when to revert changes
ControlSafety Implements safe mode and exemptions
AuditQueue Lock-free event logging

Crate Structure

crates/
├── core/           # Solver, Probe, Classify, Strategies
│   └── strategies/ # Nelder-Mead, Multi-Start NM, TPE
├── hotpath/        # Safety Executor, SPSA, Telemetry, Audit
├── cli/            # CLI, TUI, Dashboard
├── ffi/            # C FFI bindings
└── ship/           # ArqonShip DevSecOps tool

bindings/
└── python/         # PyO3 bindings (ArqonSolver, ArqonProbe)

Data Flow

  1. User calls ask() → Solver returns candidate parameters
  2. User evaluates candidates → calls tell() with results
  3. Solver updates internal state, refines strategy
  4. Repeat until budget exhausted

For online mode, use ask_one() + seed() for single-candidate flow.

Performance Characteristics

Metric Value
Overhead per trial ~3ms
Throughput ~33,000 trials/sec
Memory O(history_size)
Deterministic Yes (with fixed seed)

Next Steps