Examples

The language layer has three canonical example families, plus a judgment example for measured confidence.

​

Classification

Demonstrates:

• Implicit ctx
• ask()
• Empty-reply fallback

The workflow classifies a message and falls back to "unknown" when the reply is empty. It does not gate on confidence, because a plain ask() reply has no measured confidence.

​

Confidence-gated triage

Demonstrates:

• judge(...) with a choice question
• to_probable()
• A real probability gate

examples/judgments/triage.voss asks Jev which support queue should handle a request and returns the queue only when its probability is at least 0.80. Running it needs a Jev key.

​

Semantic support routing

Demonstrates:

• match similar
• within budget
• Fallback control flow
• Context budgeting

The workflow routes customer messages by meaning and falls back when budget limits are exceeded.

​

Parallel research

Demonstrates:

• agent
• spawn
• gather
• @tool
• memory
• Budget fallback

The workflow launches one researcher per topic, gathers findings, and degrades to a simpler summary if needed.

​

Success criterion

For the language layer, the examples should pass voss check, compile to readable Python, and run with hermetic defaults where possible.