B-Type Architecture

PID × FSM × LLM Reliability-Oriented Control Stack

Architectural Intent

The B-Type architecture is designed to explicitly separate control execution from reliability judgment.

Rather than embedding adaptation logic directly into the control loop,
B-Type introduces a layered structure in which:

PID guarantees baseline stability
FSM decides whether adaptation is allowed
LLM supports redesign and explanation outside real-time constraints

This separation ensures that reliability constraints dominate performance optimization.

Layered Architecture Overview

flowchart TD
    R[Reference]
    P[Plant]

    PID[Inner Loop\nFixed PID Controller]
    FSM[Middle Layer\nReliability FSM]
    LLM[Outer Layer\nLLM Design Support]

    R --> PID
    PID --> P
    P --> PID

    PID --> FSM
    FSM --> PID

    FSM -. redesign rules .-> LLM
    LLM -. update parameters .-> FSM

Inner Loop: PID Controller

Role

The PID controller forms the core real-time control loop.

Design Characteristics

Fixed, conservative gains
Designed with sufficient gain and phase margins
No online structural modification

Design Rationale

The PID loop must remain stable regardless of higher-layer decisions.

In B-Type, the PID controller is never optimized aggressively.
It exists to guarantee a minimum safe control authority.

Middle Layer: FSM (Reliability Guard)

Role

The FSM supervises the PID loop and determines whether adaptive actions are permitted.

FSM Responsibilities

Monitor reliability metrics derived from plant response
Compare metrics against predefined thresholds
Select control modes accordingly

Representative FSM States

NORMAL: Nominal operation, no significant degradation
DEGRADED: Degradation detected, adaptation under consideration
ADAPT_ALLOWED: Adaptation permitted within limits
ADAPT_BLOCKED: Adaptation explicitly prohibited
SAFE_MODE: Forced fallback to fixed PID operation

Key Principle

The FSM must explicitly support
“the decision to block adaptation.”

This is the defining difference between A-Type and B-Type.

Outer Layer: LLM (Design-Time Intelligence)

Role

The LLM operates outside the real-time control loop and acts as a design assistant.

Typical Functions

Propose new FSM guard thresholds
Suggest alternative PID parameter sets
Analyze long-term reliability trends
Generate human-readable explanations of control decisions

Non-Responsibilities

No direct real-time control
No continuous parameter injection into the PID loop

The LLM influences design evolution, not instantaneous control.

Information Flow and Time Scales

Layer	Time Scale	Responsibility
PID	Continuous / real-time	Stabilization and tracking
FSM	Discrete / supervisory	Reliability-based mode selection
LLM	Offline / asynchronous	Redesign and explanation

This strict separation prevents unstable feedback between learning, supervision, and control.

Architectural Guarantees

The B-Type architecture guarantees that:

The control loop remains stable even if adaptation is disabled
Adaptive behavior is constrained by explicit reliability metrics
Long-term degradation does not lead to uncontrolled compensation
System behavior remains interpretable and explainable

Relation to A-Type Architecture

A-Type focuses on how adaptation can be performed
B-Type focuses on when adaptation must be restricted

B-Type should be understood as a supervisory envelope placed around the A-Type adaptive mechanisms.

Summary

The B-Type architecture formalizes a key control principle:

Performance optimization must never override reliability constraints.

By layering PID, FSM, and LLM with clearly separated roles and time scales,
B-Type provides a robust foundation for adaptive control under long-term plant degradation.

The next section details:

Formal definitions of reliability metrics
FSM guard conditions and threshold logic