Automation runs safety-critical processes well inside their design envelope. Abnormal situations still fall to human operators, and their mismanagement is a leading contributor to process-safety incidents. Full autonomy in other domains usually leans on a safe stop. A chemical plant does not have that recourse because stopping is itself a hazardous and expensive operation.

This project asks whether a general-purpose large reasoning model (LRM), with no task-specific training and only the information available to an operator, can manage abnormal situations at run time, provided it acts through an interface that cannot express an unsafe command.

A large reasoning model sits in an abnormal-situation-management layer above a regulatory control layer, connected through a programmatic validator and a digital-twin forward simulation.

The LRM operates as an abnormal-situation-management layer above the decentralized regulatory (PID) controllers. It sees only a measured process-state summary and proposes actions from a bounded interface, and every proposal passes a programmatic validator backed by a digital-twin shadow rollout.

How it is guarded

  • Bounded action interface. The model chooses among a small set of generic supervisory moves. It can adjust setpoints, reconfigure control, run a procedure, distrust a suspect instrument, do nothing, or recommend shutdown. Direct actuator manipulation is not representable.
  • Programmatic validator. Every proposal is checked against static bounds, rates, cooldowns, and a per-episode budget. The validator returns machine-readable reasons for a single revision.
  • Shadow rollout. A nominal-model digital twin serves two roles. The model queries it to reason about a candidate action, and the validator runs it as a shadow simulation to veto any move predicted to trip the plant.

Knowledge is split in two. A process-agnostic reasoning doctrine lives in the system prompt, while a generated, plant-specific description supplies the numbers and mechanisms. Moving to a new process means regenerating one document rather than re-engineering the method.

The benchmark covers 39 episodes on a plant-wide industrial process built on the Tennessee Eastman simulator. The episodes span cases where no intervention is needed, cases where a safety limit is approached, and cases where a commanded quality change exceeds what the baseline can track. The model kept the plant within every hard constraint in every episode.

Code and benchmark are available at github.com/khalidlabs/safetep-core, built on the TEP Studio simulation kernel.