24 Ways Jev Can Support Your AI Decision Models

🔍 Read the full analysis: 24 Ways Jev Can Support Your AI Decision Models on ThorstenMeyerAI.com

Prime Big Deal Days · Oct 6–7Offer from Amazon

Get hardware and tech essentials delivered free — and shop member deals

• Fast, free delivery on millions of items
• Access to Prime Big Deal Days deals on October 6–7
• Prime Video, Amazon Music and more included

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

Thorsten Meyer says he has mapped 24 possible uses for Jev, a tool that supplies structured answers to software, and has three checks running in his publishing operation. His account describes using the tool for high volumes of narrow decisions and routing uncertain cases elsewhere.

Meyer describes Jev as a system that takes text or JSON alongside typed questions and returns answers that code can use directly. It does not, he says, write or summarize material. The listed answer types include a yes-or-no probability, a choice among options with probabilities and confidence, and a score on an ordered scale. Meyer estimates a call takes about 0.3 to 0.9 seconds and costs about $0.04 per million input tokens; those performance and cost figures are his claims.

For his three live publishing applications, Meyer reports that a relevance check assessed about 10,000 story-and-site pairings in three days, with 22% judged clearly on-topic. A language check scanned 78,889 articles for $2.01, identified 1,576 as non-English, and led to 1,553 being fixed, according to his account. A third use classifies headlines into 31 topics as a fallback when a primary language model makes an error. Meyer reports 89% agreement with a frontier model overall and 97% to 99% agreement when Jev confidence was at least 0.8.

The source also describes proposed publishing checks, but does not provide the complete list of 24 uses. Among the cases detailed are disclosure detection and comment moderation, which Meyer labels strong fits; a thin-source detector, product matching for roundups and headline-quality scoring, which he says need measurement; and same-event deduplication, which he calls a poor fit after a canary test found no duplicates. The excerpt ends during a later section, so it does not establish the remaining cases or the full breakdown across other industries.

Where Structured Checks May Help

Meyer’s guide describes using structured answers in software workflows for routine checks. A program can act on a result and send low-confidence answers to a person or a more capable model. Meyer says this arrangement could help teams apply checks to many items while accounting for uncertainty.

His examples distinguish among cases by their reported results. He labels disclosure detection and comment moderation strong fits, while saying a thin-source detector, product matching for roundups and headline-quality scoring need measurement. His deduplication canary found no cases to fix, and he describes that use as a poor fit. The categories combine reported production use, proposed applications and cases that Meyer says need further measurement.

The reported scan figures describe Meyer’s experience in his own operation. The material supplied does not include the underlying dataset, detailed error rates for each application, or an outside assessment. It therefore does not establish general performance, savings or safety outcomes.

Meyer’s Four-Part Fit Test

Meyer says a proposed application should meet four conditions: high volume, a narrow question that does not require multi-step reasoning, errors that are cheap or can be routed to a more capable reviewer, and a heuristic that has been shown to fail. He advises keeping a working keyword rule when it works, rather than replacing it simply because a model is available.

His suggested evaluation begins in “shadow” mode: replay 300 to 500 past decisions, compare results overall and across confidence bands, and review 20 disagreements to determine which answer was right. Meyer says teams should connect the system only where the high-confidence band reaches 95%. He also recommends giving the check a separate flag that is off by default, testing it on 5% to 10% of units, and then expanding its use. These are recommendations in his guide, not independently verified standards.

The production examples illustrate the proposed approach. In the relevance gate, Meyer says the system drops a story only when it is clearly a poor fit and confidence is high; uncertain items keep their previous path. For the language check, his stated rule rewrites an article in place when the probability that it is English falls below 0.1. The exact implementation details and the evaluation method behind the reported outcomes are not included in the provided material.

“Jev is the right tool wherever a system needs thousands of small judgements and can hand the unclear ones to something smarter.”

— Thorsten Meyer

Performance Outside Meyer’s Operation

The supplied source is an excerpt and stops partway through its commerce and customer-operations section. It does not detail all 24 mapped uses, substantiate the claimed overall split of three live uses, 12 strong fits, seven cases needing measurement and two poor fits, or show how the cases outside publishing were evaluated. The claim that 15 uses are ready to build or already running is Meyer’s characterization.

It is also unclear how the reported agreement with a frontier model translates into correctness against independently checked answers. Agreement between two systems is not, on its own, a measure of accuracy. The source does not specify the sample size or review process for the 31-topic comparison, nor give confidence-band results for the other applications. No independent cost comparison or account of mistakes that caused harm is provided.

Measure Before Wider Rollout

Meyer’s proposed next step for teams considering a use is to test it against real historical decisions, inspect disagreements and check results by confidence band before connecting it to live workflows. His suggested rollout starts with a separate feature flag and a small canary. The source does not announce a product launch, a release date or a planned expansion of his own deployment.

For readers seeking the full map, the complete article would be needed to review the remaining use cases and their questions, decision rules and fit labels. Until those details and evaluation data are available, the cases described here support a narrower conclusion: Meyer reports several active publishing checks and presents a measurement-first framework for deciding where additional checks might fit.

Key Questions

What is Jev, according to Meyer?

Meyer describes Jev as a tool that receives text or JSON and typed questions, then returns structured answers such as probabilities, classifications or scores for software to use.

Which Jev uses does Meyer say are live?

He lists a relevance gate for stories and sites, an English-language check, and a fallback classifier for headlines across 31 topics. The reported results come from his own publishing operation.

How does Meyer recommend evaluating a use case?

He recommends replaying 300 to 500 past decisions, comparing outcomes by confidence band and reviewing disagreements. He says to connect a check only where its high-confidence band reaches 95%, then begin with a small canary.

Are all 24 use cases documented in the supplied material?

No. The source excerpt provides details for several publishing cases and stops during the next section. The remaining use cases and the full industry breakdown cannot be confirmed from the material provided.

Source: ThorstenMeyerAI.com