7 Production Use Cases That Jev Unlocks

Most software that needs a judgment has been borrowing a chatbot to get it. The team sends a ticket or an invoice to a large language model, waits seconds for prose, then parses that prose back into a value the code can branch on.

Typesafe AI’s Jev replaces that loop with a different interface. You send state as text or JSON and a set of typed questions, and Jev returns typed answers with probabilities in 70 to 500 milliseconds, at $0.042 per million input tokens with free output, so the check can live inside the request path.

Jev does not generate text. It answers three kinds of questions you define: Choice picks one of up to 255 options you name, Score rates the state against two to ten ordered levels, and Noul returns a yes-or-no probability.

Every question in one call shares one reading of the same state and is evaluated in parallel, so adding a second or tenth question costs almost no extra time and only the tokens for the questions themselves. The rest of the work stays where it belongs, in deterministic code that applies thresholds, permissions, dates, sums, and side effects.

The seven builds below use that division. Each one names the decision schema, the question types, how thresholds move the workflow, and a reported example with numbers where a public run exists.

1. Support ticket triage

Triage is a classification and routing problem that repeats every few minutes and punishes delay. The state is the ticket subject and body plus context the router already holds, such as plan, prior failures, and account tier. One batched call carries a Choice for owning team, a Noul for urgency, and a Score for frustration, so one reading produces all three signals.

Code owns the policy. If the Choice lands on billing with high probability and urgency is near one, the ticket goes to the billing queue and pages on-call when tier and delay match the runbook.

If probabilities split across two teams or confidence is low, the case stays in a shared review queue for a person to assign. Dates, entitlement checks, and idempotency remain in code and are never asked of the model.

A public pattern from TypeSafe’s docs puts a support ticket and its questions at roughly 300 input tokens, which is about $0.0000126 per call at published pricing, or about $1.26 for 100,000 tickets. The more operational signal comes from TypeSafe’s customer service workflow eval, which covers the same decisions at system level: Jev agreed with the averaged frontier reference on 76.0 percent of cases at about $0.0001 and 0.4 seconds per case. That level is mid-tier judgment at infrastructure timing, which is why the design routes the confident calls automatically and keeps the close calls with a person.

2. Lead scoring and routing

Lead scoring fails when one prompt tries to do ranking, weighting, and routing at once. The schema here separates atomic judgments from arithmetic. Jev scores stable dimensions with Score questions such as product fit, technical depth, buying intent, completeness of the request, and risk, each against a short rubric you write in plain language.

Your code then combines those scores with weights that reflect the business. A product-led motion may weight buying intent and technical depth heavily, while an enterprise motion may weight budget signals and team size. A Choice question selects the route, for example sales-assisted, self-serve, or partner, and a Noul flags cases that need human review, such as conflicting signals or a high-value account with missing required fields.

Thresholds are fitted on your own labeled leads. Route automatically when the winning route has high probability and every weighted score sits above its band.

Hold the middle band for a sales development representative, and return the lowest band to marketing nurture. The economics match any high-volume scoring surface: one call per lead, parallel questions, and no output tokens to meter, so scoring 50,000 leads in a month costs on the order of cents for the judgment layer plus whatever retrieval or enrichment you already run.

3. Invoice data extraction and payment decisions

Invoices look like extraction but behave like workflow. A payable decision touches fraud, duplication, vendor identity, sums, dates, and approvals, and no model should do arithmetic on money. TypeSafe’s invoice processing eval makes the split explicit: seven rounds of questions read one subject at a time, while sums, dates, account numbers, and statuses are computed in code.

A practical build sends the document text and linked records as state, then asks a set of Noul questions in one call: is this actually an invoice, is there a vendor mismatch, does a duplicate exist, is a required field missing, is a fraud pattern present, and does the amount exceed an approval threshold. Choice or Score questions can classify the document type or the reason for a hold, while the payment amount itself is read and summed deterministically.

Code applies the lists and sequences that follow. Fraud, paid duplicate, or wrong vendor ends the review immediately. Holds, line disputes, and requests for a corrected copy are added for every rule that matches.

A final series of yes-or-no checks decides whether the payload is scheduled, sent for approval, held for signature, or paid in the undisputed part. In TypeSafe’s four-workflow aggregate, invoice processing was the weakest for Jev at 61.8 percent agreement at $0.0011 and 0.5 seconds, which shows why fitting thresholds on your own invoices and keeping a manual review band for close calls carries weight beyond any headline average.

4. Real-time content moderation

Moderation must decide in the live path, not after the post has spread. The state is the candidate text, image caption, or listing plus the uploader’s declared category and relevant policy excerpts. Questions are all Noul and Score: does this contain disallowed content, does it contain off-platform contact, does the description match the declared category, and how severe is the violation on a small ordered scale.

Thresholds drive four exits, since a single binary block is too coarse. High severity with high probability blocks outright.

Mid-level probability or mid severity routes to human review with the probability and category attached. Low scores pass with logging, and a warning insert handles the band where the model is uncertain but risk is limited. Because moderation deals with user-controlled text, the state should be treated as potentially adversarial, so injection and persuasion attempts are explicit questions and are tested deliberately.

The same mechanism protects retrieval. In a RAG pipeline, TypeSafe’s cookbooks score each retrieved passage for relevance, usability as evidence, contradiction, and manipulation.

One run across 72 passages showed an injected passage ranked first by embedding similarity at 0.584 that Jev flagged with an injection probability of 0.99. That filtering keeps a poisoned passage from ever reaching the answering model, and it runs at the same per-pair cost as any other Choice or Noul.

5. Per-turn LLM verification and guardrails

Verification is the most direct complement to a reasoning model. A frontier model writes the tool call, the citation, or the answer, and Jev checks it on the next hop at a fraction of the cost.

The state is the model’s input, its recent reasoning trace, the tool call or diff it proposes, and the evidence it cites. Questions stay atomic: does this cite support the claim, does this command write to production, does this message attempt to jailbreak, does the trace contain a destructive or irreversible action.

Code gates the action on the numbers. Hold for approval when any risk Noul exceeds 0.5, route the middle band to a stronger checker or to a person, and let low-risk actions through with an audit event. Latency fits inside the agent loop because typical calls return near a tenth of a second, so checks can run before the tool executes and catch risks before damage occurs.

Public runs show the shape of this use case. TypeSafe’s citation cookbook checked eight quotations against RFC 7519, returning the four real passages as verified at probabilities of 0.93 or higher and catching all four planted failures.

Every’s independent writing check asked 21 questions across 37 documents and returned 777 judgments in under 0.7 seconds for an estimated quarter of a cent, then expanded to 1,709 judgments for under one cent across eleven scenarios. In a separate pass-ship test of 12 synthetic passages, Jev caught six of seven planted defects at a median 0.35 seconds per passage against 8.83 seconds for Fable 5.1 at high effort, which frames the trade clearly: close to frontier accuracy on bounded checks at two orders of magnitude less time.

6. Large-scale document classification

When the job is label every row, cost per row decides whether the project runs. Classic pipelines used for research corpora, archives, and listings can move to Jev where one call carries the row and returns the labels.

The state is one item, such as a paper title and summary, a support transcript, or a product title and description plus seller attributes. Questions are a Choice over the taxonomy plus Nouls for cross-cutting attributes, and Scores where rank matters such as evidence strength or listing quality.

Hierarchies with many leaves need staging. Jev’s Choice caps at 255 options, so a taxonomy like patents with roughly 27,000 nodes is walked with beam search, with a beam of three in TypeSafe’s example, with each level run in parallel. Deterministic code owns ordering, grouping, and counting, while Jev supplies the semantic signal per level.

The throughput reports come from early user runs. They are independent measurements, not vendor dashboard figures. One user classified 1,018 summarized research papers across 24 topics for $0.08 with a median 256 milliseconds per paper, while the summaries from a generative model cost $3.99 in the same pipeline.

Another run classified 98,000 listings in ten minutes. A third reported half a million input tokens for about two cents, which matches published input pricing. Each run batched one item at a time and treated accuracy as something to evaluate on owned labels before replacing the existing classifier.

7. Real-time game and interactive decisions

The game examples are not product pitches for games. They show that a 100 millisecond decision can live inside a loop that a multi-second model cannot. Jev receives structured state that describes the current situation in text and numbers, such as positions, inventory, health, active links, or UI element lists, and returns a Choice over the next action.

Thresholds here are action gates, with review gates sitting elsewhere in the system. High probability executes immediately.

Low confidence or a close split across two top actions falls through to a planner model or to a human-provided default. Because the state is a constructed data object built from text and numbers, the result depends on how well the application translates the world into fields Jev can read.

Two launch demos illustrate the operating point. A Doom bot fed structured game state ran at about ten decisions per second, which TypeSafe costed at roughly $7 per hour, with Jev answering in 0.114 seconds against 8.566 seconds for GPT-5.6 Terra on the same single-query comparison. A Wikiracing bot chose among hundreds to thousands of links per step and never selected a link that did not exist, which highlights the type guarantee: a Choice returns only from the set you supplied, so hallucinated options are out of scope even when the judgment itself may be wrong.

The shared pattern across all seven

Each build separates what should be modeled from what should be coded. Jev handles the fuzzy semantic judgment with a probability attached, while code handles translation, enrichment, arithmetic, policy, permissions, and side effects. Questions are decomposed so they are independent, batched so one state is paid once, and written with criteria a reviewer can audit.

Confidence turns the workflow from a single label into three exits. High probability acts automatically with logging. Mid probability routes to a stronger model or a person.

Low probability or a close race across top options is held and flagged. The value of that structure depends on your own labeled data. Before any threshold ships, run Jev in shadow beside the current process, compare to human judgments, and fit bands that keep accuracy stable as input distribution shifts.

Sources

• TypeSafe AI, Introducing System One Models and Jev, September 15, 2026.
• TypeSafe AI, Workflow evals overview and per-workflow pages at evals.typesafe.ai, including Security Incidents, Agent Trace Observability, Invoice Processing, and Customer Service.
• TypeSafe AI Docs, Quickstart, API reference, Jev 1.13 model notes, and cookbooks for citation checking and retrieval scoring.
• Hassan El Mghari via Hermes AI, classifying 1,018 research papers with Jev for $0.08 at 256 ms median per paper, September 17, 2026.
• Every, Mini-Vibe Check: TypeSafe's Jev Judged Everything I've Written in 0.7 Seconds, September 15, 2026, reporting 777 judgments in under 0.7 seconds and 1,709 judgments for under one cent.
• Flavio Copes, A deep dive into Jev, September 17, 2026, including 98,000 listings in ten minutes and half a million tokens for about two cents.
• LangChain Blog, Can Jev Be a Better Agent Evaluator, September 20, 2026, reporting 433x to 913x lower variance and $0.00035 per Jev judgment.
• The Register, TypeSafe AI debuts model for machines that plays Doom, September 16, 2026, reporting 0.114 s versus 8.566 s.
• Arize AI, TypeSafe Jev: Can Decision Models Replace LLM Judges, September 18, 2026, including spam calibration and aggregate workflow numbers.
• Oflight, TypeSafe Jev Tutorial: API, Code Examples and 12 Use Cases, September 19, 2026, including 12.2x cheaper batching on a 13-question request.