Generate, Evaluate, Route¶

A generative model writes a draft; Jev evaluates it against explicit evidence; Kedi selects the next operation. Neither the draft nor a high probability is permission to send a message or perform a financial action.

This example requires kedi[typesafe], pydantic-ai-slim[openai] and credentials for both integrations. It prints a review result; it does not send email. Model availability is a provider prerequisite, not a compiler guarantee.

> adapter: pydantic

@draft_reply(ticket: str, policy: str) -> str:
  > model: openai:gpt-5.6-luna
  >> Customer request: <ticket>.
  Policy: <policy>.
  A brief reply that does not invent completed actions is [reply].
  = `reply`

@check_reply(reply: str, policy: str) -> bool:
  > model: typesafe/jev-latest
  > settings:
    typesafe_threshold: 0.9
  >> Given policy <policy>, does reply <reply> avoid unsupported commitments?
  It is [supported: bool] that the reply is supported by the policy.
  = `supported`

[ticket] = Please refund my duplicate payment today.
[policy] = A reviewer must verify duplicate charges before approving refunds.
[reply] = `draft_reply(ticket, policy)`
[supported: bool] = `check_reply(reply, policy)`

> if: `supported`:
  = Draft for human review: <reply>
> else:
  = Manual review required; do not send the draft.

Boundaries¶

The first procedure can generate arbitrary prose. The second produces a boolean using Jev's strict threshold policy, not a generative text fallback. Unsupported schemas and provider failures remain errors rather than becoming a negative decision. The final branch is deterministic because it reads an existing bool.

The program deliberately does not retry until acceptance. A revision loop needs an explicit attempt budget, fresh evaluation of each revision, and a terminal manual-review outcome. Such a policy belongs to the application, not to hidden adapter behavior.

Use decision capture around an embedding call to retain the check_reply output's evidence after its procedure frame ends. The caller's deterministic supported initialization is not itself a new Jev evaluation record.