Patterns · illustrative

Put decisions where they help

These examples show request structure only. Validate thresholds and human-review policies against your own product before use.

01 / Routing

Choose a support queue

JSON request

{"model":"s1-fast","state":{"message":"I need to update my invoice address.","account_type":"business"},"questions":{"route":{"type":"choice","instructions":"Choose the support queue that best fits the request.","criteria":{"billing":null,"technical":null,"account":null}}}}

02 / Guardrail

Ask a yes-or-no risk question

A probability can support a downstream threshold, but it is not a safety guarantee.

JSON request

{"model":"s1-fast","state":{"action":"delete_workspace","actor_role":"viewer"},"questions":{"allowed":{"type":"noul","instructions":"Is this action allowed under the stated role policy?"}}}

03 / Review band

Score a draft against a rubric

JSON request

{"model":"s1-pro","state":{"draft":"...","rubric":"Clarity and factual support"},"questions":{"quality":{"type":"score","instructions":"Rate the draft against the rubric.","criteria":["low","medium","high"]}}}

04 / Image decision

Classify a bounded image input

Image input is coming soon. This shape shows how an image question will look; today the API accepts text.

Illustrative shape

{"model":"s1-pro","state":"Is this package visibly damaged?","questions":{"damaged":{"type":"noul","instructions":"Is the package visibly damaged?"}},"images":["data:image/png;base64,<bounded-image-data>"]}

Not a policy substitute

Use deterministic checks and human review where a mistake could affect people, money, access or safety. These snippets are not validated model performance claims.