Safety Gate & Policy Engine

Safety Gate & Policy Engine

In industrial operational technology (OT), automated actuators and programmable logic controllers (PLCs) govern physical machinery where unconstrained actions can cause catastrophic mechanical failure, severe downtime, or human injury.

Sofia Engine enforces a strict architectural principle: No machine learning inference, reinforcement learning agent, or diagnostic copilot can actuate physical equipment directly. Every actuation intent must pass through the deterministic PolicyEngine.

---

1. Default "DENY" Architecture

The PolicyEngine (sofia_ai.decision.policy) implements an uncompromising default DENY security posture:

flowchart TD
    CR["CommandRequest\n(Action, Device, Confidence, State, Nonce)"] --> PE{"PolicyEngine Gate"}
    
    PE --> C1{"Action in Allowlist?"}
    C1 -->|No| D1["DENY: ACTION_NOT_ALLOWED"]
    C1 -->|Yes| C2{"Confidence >= Floor?"}
    
    C2 -->|No| D2["DENY: CONFIDENCE_TOO_LOW"]
    C2 -->|Yes| C3{"Machine State == OPERATIONAL?"}
    
    C3 -->|No| D3["DENY: INVALID_MACHINE_STATE"]
    C3 -->|Yes| C4{"Operator Approved?\n(if required)"}
    
    C4 -->|No| D4["DENY: OPERATOR_APPROVAL_REQUIRED"]
    C4 -->|Yes| C5{"Physical Interlocks Clear?"}
    
    C5 -->|No| D5["DENY: INTERLOCK_ACTIVE"]
    C5 -->|Yes| C6{"Nonce Fresh & Unseen?\n(TTL Window)"}
    
    C6 -->|No / Replayed| D6["DENY: REPLAY_DETECTED"]
    C6 -->|Yes| APP["CommandDecision: APPROVE\n(Signed Audit Trail)"]
    
    style APP fill:#a6e3a1,stroke:#40a02b,stroke-width:2px,color:#181825
    style D1 fill:#f38ba8,stroke:#d20f39,stroke-width:2px,color:#181825
    style D2 fill:#f38ba8,stroke:#d20f39,stroke-width:2px,color:#181825
    style D3 fill:#f38ba8,stroke:#d20f39,stroke-width:2px,color:#181825
    style D4 fill:#f38ba8,stroke:#d20f39,stroke-width:2px,color:#181825
    style D5 fill:#f38ba8,stroke:#d20f39,stroke-width:2px,color:#181825
    style D6 fill:#f38ba8,stroke:#d20f39,stroke-width:2px,color:#181825

Loading

---

2. Policy Verification Checks

Every CommandRequest must satisfy all of the following criteria simultaneously to receive an APPROVE decision:

Gate	Check	Rationale	Failure Verdict
1. Allowlist	action in allowed_actions	Prevents arbitrary or unauthorized action strings from entering PLC dispatch.	ACTION_NOT_ALLOWED
2. Confidence	confidence >= min_confidence	Ensures decisions are only made when model/detector certainty is high (e.g. $\ge 0.85$).	CONFIDENCE_TOO_LOW
3. Machine State	state in valid_states	Prevents actuation if the asset is in MAINTENANCE, FAULT, or EMERGENCY_STOP.	INVALID_MACHINE_STATE
4. Operator Gate	operator_approved == True	Mandatory human-in-the-loop authorization for critical commands (e.g., speed changes, shutdowns).	OPERATOR_APPROVAL_REQUIRED
5. Interlocks	interlocks_clear == True	Verifies physical limit switches, thermal overloads, and safety trip lines.	INTERLOCK_ACTIVE
6. Rate Limiting	rate <= max_commands_per_min	Throttles actuation frequency to prevent valve chatter or contactor wear.	RATE_LIMIT_EXCEEDED
7. Replay Window	now - timestamp < ttl and nonce not seen	Rejects previously intercepted or delayed network commands.	REPLAY_DETECTED / EXPIRED

---

3. Code Example: Policy Evaluation

from sofia_ai.decision import PolicyEngine, CommandRequest, MachineState

# Instantiate policy with explicit constraints
policy = PolicyEngine(
    allowed_actions={"THROTTLE_DOWN", "LUBRICATE_BEARING"},
    min_confidence=0.80,
    operator_approval_required=True,
    valid_states={MachineState.OPERATIONAL, MachineState.STANDBY},
)

# 1. Attempt unapproved command -> DENIED
cmd_unapproved = CommandRequest(
    action="THROTTLE_DOWN",
    device_id="turbine-04",
    machine_state=MachineState.OPERATIONAL,
    confidence=0.95,
    operator_approved=False,
)
decision = policy.evaluate(cmd_unapproved)
assert decision.verdict.value == "DENY"
assert decision.reason == "OPERATOR_APPROVAL_REQUIRED"

# 2. Approved command -> APPROVED
cmd_approved = CommandRequest(
    action="THROTTLE_DOWN",
    device_id="turbine-04",
    machine_state=MachineState.OPERATIONAL,
    confidence=0.95,
    operator_approved=True,
)
decision = policy.evaluate(cmd_approved)
assert decision.verdict.value == "APPROVE"

---

4. Regulatory & Functional Safety Disclaimer

Caution

Sofia Engine is an engineering decision-support tool, NOT a certified Safety Instrumented System (SIS).

Sofia Engine does not claim compliance with:

• IEC 61508 (Functional Safety of Electrical/Electronic/Programmable Electronic Safety-related Systems)
• ISO 13849 (Safety of Machinery - Safety-related parts of control systems)
• IEC 62443 (Security for Industrial Automation and Control Systems)

Physical safety interlocks (emergency stops, mechanical relief valves, hardware trip relays) must remain implemented in dedicated hardware or certified safety PLCs completely outside of Sofia Engine's software control loop.