📊 Comparison

Design Recovery Control vs RL-Based Control vs LLM-Based Control

---

🎯 Purpose

This document provides a strict, explicit, and non-negotiable comparison between:

• 🛠 Design Recovery Control (DRC)
• 🔁 Reinforcement Learning (RL)–based control
• 🧠 LLM-based control systems

Its purpose is to prevent conceptual mixing,
especially in safety-critical, audited, or certified engineering contexts.

---

🔑 Fundamental Conceptual Difference

The decisive difference is what is being controlled.

Framework	What Is Directly Controlled
🛠 DRC	Control design assumptions
🔁 RL-based control	Control inputs or learned policies
🧠 LLM-based control	Control decisions or actions

This distinction is architectural, not stylistic.

---

🧩 Architectural Comparison

Aspect	🛠 DRC	🔁 RL-Based Control	🧠 LLM-Based Control
Real-time control	PID / FSM only	Learned policy	LLM inference
Learning element	None	Central	Central
LLM role	Design supervisor only	None	Primary controller
Execution timing	Asynchronous, discrete	Continuous / online	Continuous or event-driven
Safety authority	PID + FSM (explicit)	External or learned	Often implicit
Determinism	Deterministic	Often stochastic	Non-deterministic
Inspectability	Full	Partial	Low
Certification suitability	High	Low–Medium	Very low

---

🔒 Control Authority Boundary

🛠 Design Recovery Control (DRC)

• The LLM never touches:
  • control inputs,
  • actuator commands,
  • real-time execution paths.
• The LLM only modifies:
  • design parameters,
  • design assumptions,
  • configuration artifacts.

👉 Control authority remains fully classical and deterministic.

---

🔁 Reinforcement Learning–Based Control

• Learned policy directly outputs control actions
• Control logic is implicit and learned
• Safety is typically enforced via constraints or wrappers

⚠ Common risks:

• policy opacity,
• distribution shift sensitivity,
• certification difficulty.

---

🧠 LLM-Based Control

• LLM reasons about commands or actions
• Real-time guarantees are weak or absent
• Output variability is intrinsic

⚠ Common risks:

• non-deterministic behavior,
• hallucination under uncertainty,
• incompatibility with hard safety constraints.

---

🔄 Learning vs Recovery

Concept	🛠 DRC	🔁 RL	🧠 LLM Control
Online learning	❌ No	✅ Yes	⚠ Sometimes
Self-modifying behavior	❌ No	✅ Yes	❌ Often
Design intent preservation	✅ Yes	❌ No	❌ No
Assumption recovery	✅ Yes	❌ No	❌ No

DRC restores design validity,
not behavior.

---

⚠ Failure Handling Philosophy

🛠 DRC

• Assumes:
  • the physical system remains operational,
  • the control structure is still meaningful.
• Repairs:
  • violated design assumptions.

---

🔁 RL / 🧠 LLM Control

• Attempts to:
  • adapt behavior directly,
  • learn new control policies.

🚫 These philosophies are fundamentally irreconcilable.

---

🛡 Safety and Certification Perspective

Criterion	🛠 DRC	🔁 RL	🧠 LLM Control
Real-time determinism	✅	❌	❌
Explicit safety guards	✅ FSM	⚠ Optional	❌ Rare
Auditability	✅	⚠ Partial	❌
Formal verification	✅	❌	❌
Human approval gating	✅	❌	❌

---

🧭 When Each Approach Is Appropriate

Use 🛠 DRC when:

• safety certification is required,
• long-term degradation occurs,
• human review is mandatory,
• design intent must be preserved.

---

Use 🔁 RL when:

• the environment is well-bounded,
• exploration is acceptable,
• safety risk is low or externally mitigated.

---

Use 🧠 LLM-Based Control when:

• the system is non-critical,
• high-level autonomy is desired,
• failure consequences are minimal.

---

🚫 Explicit Non-Equivalence Statement

Design Recovery Control is NOT a form of reinforcement learning.
Design Recovery Control is NOT an LLM-based controller.

Any system that allows an RL agent or LLM
to directly influence control inputs
must not be described as DRC.

---

🔒 Design Intent Freeze

This document fixes the conceptual boundaries
between DRC, RL-based control, and LLM-based control.

Future documents may expand examples,
but must not blur, merge, or reinterpret these categories.

---

End of document.