What can you use Jev for?

Find the decision before choosing the model

A useful starting task has a bounded question, enough evidence to answer it, and a clear next step. For a customer message, those might be the requested service, the relevant account record, and a suggested queue. For a document, they might be a reader’s query, the text, and a location to inspect.

The workflows here are design suggestions unless explicitly labeled as official examples or community reports. We have made live calls to check this website’s functionality, but have not independently reproduced the third-party cases or published quality or performance benchmarks. A working demo and an individual model result do not establish accuracy for these applications.

Community cases: how builders use Jev

These public reports show specific decisions that builders have tested. Each summary links to its author. Inclusion does not imply endorsement by TypeSafe or the featured teams.

Try a writing check in the official Playground

Original fictional teaching example; no recorded model results. This is a separate exercise inspired by the writing-review use case, not Every’s prompt or a reproduction of its experiment. Orbit Notes is fictional. The English input and questions stay identical across languages for copying.

Open the official Playground

An external TypeSafe page. Sign in with your own account and the required access; joining the waitlist does not itself grant access.

1. Paste the sample below into the state field.

2. Add three Noul questions using the names and instructions below. The official quickstart explains how to enter a state and questions.

3. Run the request in the official Playground. Read each returned probability alongside its question, then edit the sample and run it again if useful.

Fictional input

Orbit Notes saves your drafts locally.
Your drafts are stored on your device.
Click Export to download a copy.

Questions to enter

repetition (Noul)
Does the text repeat a claim without adding new information?

clear_action (Noul)
Does the text explain what happens when the reader clicks Export?

guaranteed_safety (Noul)
Does the text claim that a draft can never be lost?

Inspect whether the first two sentences add different information, whether the Export action is explained, and whether the text promises that drafts cannot be lost. The questions are independent; their probabilities do not need to sum to one. Use the judgments as prompts for human review. This exercise supplies no expected scores or automatic decision threshold.

Four roles, four starting points

Developers: review a semantic rule

Try a narrowly written convention that ordinary linting misses: does a change introduce a user-facing error without explaining how to recover? Supply the relevant diff and the rule, and return a review signal. The official use-case map includes semantic code linting; this particular check is our illustrative proposal.

Keep the compiler, test suite, and exact lint rules. Let a reviewer inspect the flagged lines and decide whether the concern is valid. Begin with advisory comments so you can learn the cost of false alarms before making the check a merge requirement. This is a suggested integration, not an available PR bot supplied by this website.

Support teams: separate ownership from urgency

A frustrated customer may need billing rather than engineering. An apparently calm message may describe an urgent outage. Ask separately about the destination and time sensitivity, then apply a queue policy. The official quickstart provides the ticket example below.

Your application still has to retrieve account facts, deduplicate tickets, and enforce the rules for refunds or account changes. A classification does not confirm that a reported failure happened.

Search and RAG teams: choose evidence before writing

Retrieval-augmented generation (RAG) supplies retrieved material to a text generator. Jev can be evaluated between retrieval and generation. TypeSafe’s RAG passage cookbook checks relevance, usable evidence, contradictions, and attempted instructions separately, then uses code to include, flag, or exclude a passage.

Keep the source identifier with the judgment. Contradictory evidence may deserve a visible warning rather than silent deletion. Evaluate whether your filter removes the only passage needed to answer a difficult query. A generator is still responsible for the final wording; neither stage should override document access permissions.

Security teams: prioritize analyst attention

TypeSafe’s guardrails cookbook demonstrates checking incoming messages and generated replies, with probabilities for hazards and an ordered severity assessment. Code then applies the response policy.

For an initial integration, retain your existing detection path and compare the proposed review queue with analyst decisions. Track missed incidents and unnecessary escalations separately. A low model score should not grant a tool permission, disable an existing control, or establish that an attachment is harmless. See the adversarial-input limitations.

Complete example: find an answer in a document

1. Task and input

Search the cookbook’s GitHub Terms of Service text: 218 lines tagged with IDs, using jev-1.12. The cookbook and full script link the complete input and explain replay.

2. Questions

One request asks where (Choice over line IDs) and exists (Noul: does the document contain an answer?).

3. Published output excerpt

These are selected values, not a complete API response:

query: who owns the code I upload?
exists: 0.98
L052: 0.95

The matched source line begins: L052 | You own Your Content.

4. Postprocessing

The cookbook sorts line probabilities and joins IDs back to source text. Its existence policy marks values at least 0.7 as answered, below 0.35 as absent, and the interval between them as partial. This example therefore points to L052 and passes the answer-existence check.

5. Limits and source

The thresholds and earlier model version belong to this example. Replay is not a new measurement. It demonstrates retrieval, not legal interpretation or accuracy on other documents. Original case and displayed output

Why use two signals?

Choice distributes probability over the supplied options, which together sum to one. A leading option is therefore a relative winner; it is not an independent guarantee that an appropriate option exists. TypeSafe documents a maximum of 255 Choice options. Choice semantics and limits

Our implementation advice is to preserve both the location and the existence judgment in a result record. Do not turn the top line into an unconditional answer. Show the source text for inspection, handle missing or partial evidence explicitly, and retain the document version so later edits do not silently change what an ID means.

When adapting this design, include documents with no answer in your review set. Also include answers spanning several lines and questions whose wording contains a false assumption. Those cases test the retrieval policy you actually need, beyond whether the first-ranked line looks plausible.

Complete example: triage a support ticket

Input, questions, and documented output

The sample customer describes a failed Stripe connection over three days, lost sales, and a need for urgent help. The request asks for a department, frustration level, and urgency. The message is paraphrased here.

Source: quickstart request and response.

Turn the result into a proposal

The following is our explanatory code. Its thresholds are illustrative policy choices, not validated settings:

function proposeRoute(response) {
  const department = response.answers.department;
  const urgency = response.answers.is_urgent.noul;

  return {
    queue: department.confidence >= 0.7
      ? department.choice
      : "manual-triage",
    priority: urgency >= 0.9 ? "urgent" : "normal",
    suggestedTeam: department.choice,
  };
}

For the published values, this function proposes urgent manual triage and suggests technical support. The leading department does not clear our chosen confidence threshold. No ticket is actually assigned by this example.

Probability and confidence are different fields. TypeSafe derives Choice and Score confidence from the distribution; Noul has no separate confidence field. Confidence documentation

Before connecting such a proposal to a ticketing system, validate the response, define how conflicts are resolved, and make failures observable. Evaluate wrong routes and missed urgent tickets separately. The sample does not establish the customer’s account state or resolve the integration problem.

Community report: exploring phishing-email triage

A Discord participant described an early Python SDK experiment on September 17, 2026, at 14:24–14:27 UTC, hoping to assist security operations and later connect to a SOAR workflow. They described encouraging early results and sensitivity to the wording and detail of criteria. Experiment introduction, follow-up observations

This is a community self-report. It does not establish detection accuracy, false-positive rates, speed, savings, deterministic behavior, or production integration. We have not reproduced it. No verified official response was observed in the captured exchange; the research covered selected discussions, not the complete community history. The links may require community access.

Pick one next step

Choose one decision with an inspectable outcome and a manageable review process. Start with getting access and sending a request, budget the path with the pricing guide, and read the limitations before connecting the output to automatic actions.