Inputs arrive over time
A scientific task may require several values that a user supplies across multiple turns.
ACCEPTED RESEARCH CASE STUDYWORKS26 @ SC26
Guarded Scientific Model Orchestration
A state-aware layer that extends LLM function calling with explicit workflow semantics so domain scientists can invoke trained prediction models through constrained natural language.
SYSTEM PROBLEM
Function calling is not a workflow engine.
Missing state must block execution
The system should identify incomplete requirements instead of inventing arguments or invoking a model prematurely.
Exploration changes state
Corrections, what-if questions, and resets need explicit semantics so prior values do not leak into the wrong run.
ORCHESTRATION PATH
Accumulate, validate, execute, and steer.
The orchestration layer keeps the workflow state visible between natural-language input and scientific model execution.
Accumulate
Collect required inputs across multiple conversational turns.
Validate
Constrain tool arguments with typed schemas and domain requirements.
Gate
Block execution while required workflow state is incomplete.
Steer
Handle correction, what-if, and reset requests as explicit state transitions.
MY CONTRIBUTION
Workflow semantics around the model boundary.
Guarded execution
I designed the state-aware orchestration behavior and the conditions that prevent incomplete execution.
Scientific integration
The workflow was instantiated for protein-corona prediction over a backend of trained prediction models.
Evidence boundary
Measured accuracy and reliability results remain omitted from this preview until their primary evaluation artifact is approved for publication.
Accepted to WORKS26 at SC26.
The paper is scheduled to appear in November 2026. A public paper link will be added after an authoritative publication page is available and verified.