Guarded orchestration for scientific models
Letting scientists run trained prediction models through natural language without the model guessing missing inputs.
- Status
- Accepted · WORKS26 at SC26
- Paper
- Guarded, State-Aware Orchestration for Schema-Constrained Scientific Model Execution
- Authors
- Venue
- WORKS26, held with SC26 in Chicago on 15 November 2026
The problem
LLM function calling maps a request onto a tool call. A scientific prediction model needs more than that: its inputs often arrive over several turns of conversation, and a missing value has to stop the run rather than be guessed. When a scientist corrects an input, asks a what-if question, or starts over, earlier values must not leak into the new run.
The approach
I designed a state-aware layer between the conversation and the model backend. It keeps the workflow state explicit, so the language model can interpret requests but cannot run a model on incomplete or invalid inputs.
- AccumulateCollect required inputs across conversational turns.
- ValidateCheck every argument against a typed schema.
- GateBlock execution while required state is incomplete.
- ExecuteRun the trained prediction model on complete inputs.
↻ Correction, what-if, and reset requests are explicit state transitions back into the workflow.
Where it runs
The workflow was instantiated for protein-corona prediction over a backend of trained prediction models.
Results
The results are in the WORKS26 paper. A link will be added here when the proceedings are online.