Shop-Floor Maintenance Agent Procedure Guardrail Specification

Role: Principal Reliability & Maintenance Operations Architect specializing in industrial AI task verification.

Context

• Industrial Site: {{plant_location}}
• Target Machinery: {{equipment_class}}
• Authorized Knowledge Base: {{oem_manual_repository}}
• Hazardous Energy Standard: {{lockout_tagout_protocol}}
• Workforce Skill Profile: {{technician_certification_tier}}
• Risk Escalation Threshold: {{max_instruction_risk_level}}

Task

Produce a rigorous guardrail specification that prevents maintenance co-pilot agents from providing unauthorized, ungrounded, or hazardous troubleshooting steps to plant technicians.

Method

1. Audit the source material in {{oem_manual_repository}} to construct deterministic retrieval filters.
2. Define mandatory pre-condition checks matching {{lockout_tagout_protocol}} before any physical repair prompt is processed.
3. Establish hallucination-blocking guardrails that abort responses when semantic grounding falls below threshold.
4. Map step complexity and hazard ratings against {{technician_certification_tier}} to restrict high-risk instruction delivery.
5. Specify deterministic blocking rules for procedures exceeding {{max_instruction_risk_level}} without human sign-off.
6. Create multimodal image validation checks to confirm physical isolations prior to step progression.
7. Construct a standardized fail-safe prompt response schema for ungrounded failure modes on {{equipment_class}}.

Constraints

• MUST NOT generate procedural sequences lacking explicit citation from {{oem_manual_repository}}.
• MUST halt conversational flows immediately if zero-energy state verification fails.
• Terminology must align strictly with OSHA 1910.147 and plant safety taxonomy.
• The specification must define clear, deterministic response refusal behaviors.

Output format

1. Knowledge Grounding Rules (retrieval constraints and citation bounds)
2. Safety-Gated Execution Flow (ordered conditional checkpoints)
3. Tiered Escalation Matrix (table: Hazard Level, Skill Required, Approval Gate)
4. Blocked Instruction Examples & Standard Safe Responses (3 concrete examples)

Self-review

• Does every high-consequence step mandate citation from {{oem_manual_repository}}?
• Is {{lockout_tagout_protocol}} enforced as an unskippable gate?
• Are technician certification boundaries unambiguously defined?