【Control】🧩 14. Explicit Responsibility and Reliability Boundaries in Control Architectures

topics: [“AITL”, “control”, “architecture”, “AI”]

🧭 Introduction

In control system design, it is surprisingly rare to see a clear explanation of:

• What is guaranteed by design
• What is treated as an external assumption
• Where design responsibility explicitly ends

In recent years, terms such as:

• AI
• Machine learning
• Autonomous optimization

have entered the control domain, often blurring the boundary between
control, decision-making, and redesign responsibility.

This article organizes the architectural thinking behind
AITL (Architecture for Integrated Technology Logic) Controller A-Type, focusing on:

• Architectural responsibility separation
• Explicit definition of the Reliability Boundary

This is not about code or equations.
It is about why the architecture is structured this way
and where responsibility is intentionally placed.

🏗️ Fundamental Structure of AITL Controller A-Type

AITL Controller A-Type is based on a strict separation of responsibilities:

1. PID Control Layer (Inner / Real-Time Layer)
2. FSM Layer (Middle / Mode Management Layer)
3. NN / RL Assistance Layer (Real-Time, Bounded Adaptation)
4. LLM Layer (Outer / Non-Real-Time Redesign and Interpretation)

This is not a layered structure for visual sophistication.

It exists to physically separate control, adaptation, and redesign responsibilities,
and to prevent them from interfering with one another.

🧠 Design Responsibility of Each Layer

⚙️ PID Control Layer: Stability and Immediate Response

The PID layer is responsible only for:

• Real-time stabilization
• Performance within predefined design conditions
• Continuous and deterministic response to input variations

It explicitly does not handle:

• Interpretation of environmental changes
• Mode-switching decisions
• Structural redesign or learning

PID must not become intelligent.
Its role is simply to keep the physical system stable.

🧾 FSM Layer: State and Order

The FSM (Finite State Machine) layer is responsible for system order:

• Definition of control modes
• Management of state transition conditions
• Switching between approved PID parameter sets
• Permission control for NN / RL adaptation

FSM decides, but does not reason.

It does not perform inference, learning, or creative responses
to undefined situations.

🔧 NN / RL Layer: Bounded Adaptation

The NN / RL layer is not a replacement for PID or FSM.

Its operation is strictly constrained:

• Activated only when permitted by FSM
• Provides bounded, real-time adaptive assistance
• Preserves the fundamental structure and stability of PID

Its purpose is to extend the range in which:

“The system can still be handled by control”

Before adaptation becomes excessive,
authority is returned to FSM or higher layers.

🧠 LLM Layer: Redesign and Semantic Interpretation

The LLM layer, as the outermost layer, is responsible for:

• Semantic interpretation of anomalies and degradation
• Determining policies for PID gain re-identification
• Revising FSM transition rules
• Re-evaluating original design assumptions

Critically:

The LLM never participates in real-time control.

It is a thinking layer, not a controlling layer,
and is structurally isolated to preserve control-loop stability.

🛑 Defining the Reliability Boundary

The most important architectural element is the
explicit definition of the Reliability Boundary.

✅ Guaranteed by Design

• Stability of the PID layer (within design assumptions)
• Consistency of FSM state transitions
• Bounded operation of NN / RL adaptation
• Formal consistency of interfaces between layers

❌ Explicitly Not Guaranteed

• Truthfulness of sensor measurements
• Physical health of actuators
• Optimal performance in undefined environments
• Intrinsic correctness of LLM-generated outputs

This explicit boundary allows engineers to discuss:

Where design responsibility ends when something breaks

in technical, non-ambiguous terms.

📐 Why Explicit Reliability Boundaries Matter

Architectures without clear boundaries tend to produce:

• Unclear responsibility during failures
• The illusion that “AI will fix it”
• Entanglement of control, adaptation, and decision-making

In AITL Controller A-Type, responsibility is explicitly assigned:

• Stability → PID
• Order and permission → FSM
• Bounded adaptation → NN / RL
• Redesign and interpretation → LLM

This is both an engineering choice and a
formal declaration of design responsibility.

🧩 Closing Remarks

What this article presents is not an implementation trick,
but a design philosophy.

Detailed specifications, diagrams, and updated documents are available here:

👉 https://samizo-aitl.github.io/aitl-controller-a-type/

Before making control systems “intelligent,”
first make responsibility explicit.

If that message is conveyed,
this article has achieved its goal.

End of Article