PROCESS SAFETY TRUST LAYER

Process Safety Trust Layer is an operator console for letting Gemini run a chemical plant safely. The moment a language model tells a reactor to push past its thermal runaway threshold, something has to stop the call before it reaches the wire — and this is that something. Two Gemini personas split the workload. An Operator agent has tool access (physics simulators, setpoint apply, emergency shutdown) and is driven by the human in chat. A Safety Supervisor — a second Gemini context with no tool access — runs in parallel, reading the live sensor stream and producing an independent advisory the moment a reading crosses a danger band. Every plant-control tool call passes through three layers: a Python policy mirror inline (simulate-before-apply, mapped to OSHA PSM 1910.119); Lobster Trap, a Go reverse-proxy that does Deep Prompt Inspection on every request and response; and a Generate Demo Flowsheet path that has Gemini call Bob Prompter MCP over stdio to draft a new flowsheet, validate it, and launch the resulting .dwxmz in DWSIM on demand. Three demo plants wired end-to-end (Methanol Synthesis, Methanol-Water Distillation, Compressor Train). 45 ISA-5.1 sensor tags. 12 backend-applicable variables. Four parallel audit trails on the dashboard — tool calls, policy decisions, Lobster Trap wire metadata, and supervisor advisories. Twenty of twenty adversarial prompts in the red-team corpus blocked. Thirty-eight unit tests green.