SCENARIO / SUPPORT

Check promises before sending a reply

Keep policy and draft separate; check whether promises extend the supplied conditions.

Who this is for Support and review teams checking draft replies against supplied policies.

Authored examples · No live inference

Define the decision first

Use the supplied policy as evidence; the draft cannot change it.
Compare time windows, eligibility and promises; a match is not permission to auto-send.
Choose unknown when policy is missing or does not cover the question.

Decide first, then compare

01 / EXERCISE

Choose an example

{
  "policy": "Unopened items may be returned within seven days of delivery.",
  "draft": "If the item is unopened and was delivered within seven days, you may request a return."
}

How would you classify this using the criteria above?

Supported by policy Exceeds policy Insufficient information

Read all examples and authored answers

Typical example · Supported by policy

{
  "policy": "Unopened items may be returned within seven days of delivery.",
  "draft": "If the item is unopened and was delivered within seven days, you may request a return."
}

The unopened and seven-day conditions remain; no full-refund or other promise is added.

Next action: Continue to human review and verify the order facts.

Counterexample · Exceeds policy

{
  "policy": "Unopened items may be returned within seven days of delivery.",
  "draft": "Opened and used items qualify for a full refund within 30 days."
}

The draft expands eligibility, changes seven days to 30 and adds a full-refund promise.

Next action: Stop automatic sending and ask support to revise.

Insufficient information · Insufficient information

{
  "policy": "No refund policy supplied.",
  "draft": "We will refund you within three business days."
}

There is no supplied refund-timing policy to support the promise.

Next action: Obtain the policy and ask a person to confirm.

Adapt the template to your task

02 / CONFIGURE

Edit the material and question criteria to create a request. Stable field names stay in English for code. Editing does not run a model or change the authored answers above.

Material to evaluate (text or JSON)

Questions and criteria

Stable field IDs; editable criteria and options

policy_check choiceInstructionsOptions and definitions (JSON; add or remove options)

REQUEST.JSON

View the Python call

Download the original request

Run in your own environment

03 / BUILD

Download and extract the examples
Python 3.10+; no third-party dependencies. Run from the directory containing examples/.
Download all example code
Start with authored offline examples
```
python -m examples.hello_jev support --mode offline --language en
```
This checks the local workflow, not model quality. Changed inputs cannot replay old fixture answers.
Use a real account when you are ready
Set TYPESAFE_API_KEY locally and put request.json in the project root. Live calls send material to TypeSafe and may incur charges. Live interoperability has not been tested in this edition.
```
python -m examples.hello_jev support --mode live --language en --request request.json
```
Open the official Playground · Account and environment setup

Understand the method

Question design Confidence and evaluation

Request fields checked against the official API reference on 2026-09-20. API reference