Route the work.
Choose a workflow, tool, or review path using company evidence and explicit policies. Escalate when the available information is insufficient.
HELM / SYSTEM 1 MODELSHELM DECISION MODELS
Specialized models
for recurring decisions.
Specialized models for the choices inside your workflow. Define the situation, question, and allowed actions; evaluate the decision and the uncertainty around it.
A BOUNDED DECISION
ContextQuestionAllowed choices
Act. Ask. Defer.SPECIALIZE THE REPEATED CHOICE
Specify the decision the model makes.
The broader agent plans and recovers. A decision model handles a bounded question inside that process.
Improve an agent component.
Compare specialized choice models with rules and base models, then measure their effect on the complete task.
Select an allowed skill.
Use high-level skill selection from observations and task context, with independent control and physical verification.
MAKE THE CHOICE INSPECTABLE
Inspect the input and output.
Explore three example contracts. These examples illustrate the interface; they do not call a model.
Should this case proceed or be reviewed?
The requested credit exceeds the recorded authority. The allowed options include requesting approval.
SCHEMA CONCEPT / FICTIONAL EXAMPLEPROPOSED REQUEST → RESPONSE
Probabilities require task-specific calibration. This is a schema example, not a live API.
FROM LABELED EVIDENCE TO A SERVED MODEL
Train on judgment. Evaluate on outcomes.
Define the decision and the labels.
Specify allowed choices, evidence, constraints, and the consequences of an error. Include expert corrections and uncertainty where available.
Adapt to the task.
Use supervised adaptation and lightweight adapters using relevant text, documents, screens, images, or video.
Calibrate and evaluate independently.
Separate training, calibration, and test data. Compare against rules, base models, and the existing decision process.
Serve within a bounded workflow.
Configure inference with model and data versioning, explicit escalation, and evaluation of the full agent or robot task.
“System 1” describes the role of a specialized decision component. Speed, cost, and quality advantages must be demonstrated on the target workload. A choice probability is not automatically a calibrated prediction of human behavior.
DECISION-MODEL PILOT
Start with a recurring decision.
Start where the choices are explicit and outcomes can be checked.
Discuss a System 1 pilot- YOU BRING
- Representative states, candidate actions, reviewed labels, and operating constraints.
- WE SCOPE TOGETHER
- A decision schema, comparative evaluation, integration plan, and uncertainty thresholds.
- THE SUCCESS TEST
- Improved task outcomes at acceptable latency and cost, with failures and limitations visible.
Engagement details. Scope, timing, price, interfaces, data rights, and acceptance criteria are agreed before work begins.
Part of the Helm model portfolio
Domain foundations.
Company-specific decisions.
Connect vertical foundation models with company knowledge and bounded decision components for harness routing, computer use, workflow choices, and robot skill selection.