【Control】🧩 17. AITL Control Architecture — Final Specification

― Clear Responsibility Separation for PID, FSM, Adaptation, and LLM

topics: [“control”, “PID”, “FSM”, “adaptive control”, “architecture”]

🎯 Purpose of This Article

This article summarizes only the finalized design specification
of the AITL control architecture, derived from Articles 01–16.

Philosophy / background: ❌ Not included
Value judgments: ❌ Not included

👉 The goal is to immediately verify:

What each layer is responsible for
What must never be done
Where the system breaks if violated

🧱 Overall Architecture (Final)

🔁 Layered Structure

flowchart TB
    LLM[🧠 LLM<br/>Non-real-time<br/>Analysis & Proposals only]
    FSM[🧾 FSM<br/>Decision / Permission / Stop]
    PID[⚙️ PID<br/>Real-time Control]
    PLANT[🏭 Plant]

    LLM -. Design proposals .-> FSM
    FSM -->|Permit / Block| PID
    PID --> PLANT
    PLANT --> PID

    %% Forbidden paths
    LLM x--x PID
    LLM x--x PLANT

⚙️ Layer Responsibilities (Final Table)

📌 Responsibility Matrix

Layer	Primary Role	Real-Time	Allowed	Forbidden
⚙️ PID	Control	✅	Stabilization, tracking, V–I response	Decisions, learning, optimization
🧾 FSM	Judgment	✅	State transitions, permit/block	Control input generation, inference
🔁 Adaptation	Assistance	✅	Limited gain tuning (conditional)	Always-on, unlimited compensation
🧠 LLM	Analysis	❌	Root-cause analysis, design proposals	Real-time control, safety decisions

⚙️ PID Layer Specification

✅ What PID Does

Real-time stabilization
V–I response within design assumptions
Deterministic behavior

❌ What PID Must Never Do

State judgment
Mode switching
Gain redesign
Learning or optimization

👉 PID must “stabilize silently.”

🧾 FSM Layer Specification

✅ What FSM Does

Control mode definition
State transition management
Permit / block adaptation
Forced stop and fallback

❌ What FSM Must Never Do

Control signal computation
Gain calculation
Learning or optimization

👉 FSM is for judgment only.

🔁 Adaptive Control (A-Type / B-Type)

📊 Positioning Comparison

Item	A-Type	B-Type
Purpose	Limit exploration	Operation
Adaptation	Always ON	FSM-permitted only
Reliability	❌ Not guaranteed	✅ Lower bound maintained
Usage	Experiments only	Real operation

🧠 LLM Layer Specification

✅ What LLM Does

Analyze abnormal logs
Explain degradation causes
Propose design improvements

❌ What LLM Must Never Do

Real-time control
Stop / recovery decisions
Safety judgment

👉 LLM only “thinks.” It never acts.

🚫 Forbidden Design Patterns (Critical)

Any of the following is considered design failure:

❌ Inserting LLM into the control loop
❌ FSM performing optimization or learning
❌ Always-on adaptive control
❌ Undefined numeric stop thresholds

✅ Applicable Domains

Industrial control systems
Robotics
AI-assisted control architectures

📌 Final Summary

Control: ⚙️ PID
Judgment: 🧾 FSM
Adaptation: 🔁 Conditional
Analysis: 🧠 LLM (non-real-time)

👉 This responsibility separation must not be violated.

🔜 Next Article

18. AI Control Safety Checklist
— Safety Envelope, Stop Logic, and Recovery Design (with diagrams and tables)