A 24-Part Guide To Using Jev In AI Decision-Making

🔍 Read the full analysis: A 24-Part Guide To Using Jev In AI Decision-Making 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 published a guide on September 29 mapping 24 ways to use Jev for narrow, repeated decisions in areas including publishing, commerce and software. Meyer says three applications are running in his publishing operation, while 12 other use cases meet his criteria for a strong fit; the guide is his account, and its performance figures have not been independently verified.

Jev takes a text or JSON state and a set of typed questions, then returns structured answers that software can use to branch. The guide describes three answer types: a yes-or-no probability, a choice with probabilities and confidence, and a score on ordered levels. Meyer says it does not write or summarize content, and that a call takes about 0.3 to 0.9 seconds. He estimates the price at about $0.04 per million input tokens.

Meyer reports three live publishing uses. A relevance gate 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, he says, finding 1,576 non-English articles and fixing 1,553 in place. A fallback classifier covering 31 topics achieved 89% agreement with a frontier large language model, according to the guide.

For that classifier, Meyer reports agreement of 97% to 99% when Jev’s confidence was at least 0.8, compared with 42% below 0.5. These are results from his measurement, not a general benchmark. His recommended pattern is to let software act on clear, high-confidence results and send uncertain cases to a person or another system. The guide also lists checks for disclosure language and comment moderation as strong fits, while labeling thin-source detection, product matching and headline quality as requiring measurement.

Reported Uses and Confidence Thresholds

Meyer describes Jev as a way to apply low-cost checks to routine decisions at scale, with uncertain cases routed for review. His publishing examples include language checks, comment routing and affiliate disclosure checks. The effects of these uses depend on answer accuracy and how organizations handle errors.

Meyer says his language check processed nearly 79,000 articles for $2.01. That estimate and the reported fixes describe his workflow; they do not establish costs or performance for other organizations or material. For disclosure checks, the guide recommends human review of likely misses rather than automatic publication.

The guide sets out four conditions for considering a use case: high volume, a narrow question, low-cost errors or a route for uncertain cases, and evidence that an existing heuristic fails. Meyer advises keeping a keyword rule when it works.

Testing the Four Fit Conditions

The guide proposes testing a use case before connecting it to a live workflow. Meyer recommends replaying 300 to 500 past decisions, comparing results overall and across confidence bands, then reviewing 20 disagreements to determine which system was right. He says a use case should be wired in only if the high-confidence band reaches 95% in that evaluation.

For rollout, the guide calls for a dedicated feature flag that is off by default, followed by a canary covering 5% to 10% of units before a wider release. These are Meyer’s suggested thresholds and process, not reported industry standards. The guide assigns 12 use cases to “strong fit,” seven to “measure first,” three to “live,” and two to “poor fit.”

One example marked poor fit is same-event deduplication: Meyer says a canary found no duplicates, so the use case lacked evidence of a problem to solve. He says it could be reconsidered if a duplicate issue is measured later. In publishing, he also reports that 88% of news items he processes begin from a bare headline, motivating a proposed thin-source detector; that detector remains in the “measure first” category.

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

— Thorsten Meyer, guide author

Independent Results Remain Unreported

The guide provides Meyer’s own measurements, but it does not identify an independent evaluation of Jev’s accuracy, latency or cost. It is also unclear how the reported agreement changes across different datasets, languages, topics or operating conditions. The comparison with a frontier model measures agreement; the guide does not establish that either system was correct in every disputed case.

Several proposed applications remain unvalidated by the guide’s own criteria. Meyer says seven need measurement because a failing heuristic has not been demonstrated, and that two fail at least one fit condition. The article excerpt describes publishing use cases and begins a section on commerce and customer operations, but does not provide the full details of all 24 applications.

Measure Before Wider Deployment

The next step proposed in the guide is to test candidate uses against real past decisions, inspect disagreements and confirm performance in the high-confidence band. Teams that meet the suggested threshold would then introduce a feature flag and a small canary before expanding use. The guide does not announce a product launch, independent study or deployment schedule for the seven cases marked “measure first.”

For readers evaluating the method, the open practical question is whether their own workflow has enough volume, sufficiently narrow decisions and a visible error in its current rule. Meyer’s reported results offer examples from one publishing operation; each proposed use still requires validation against the organization’s own data and error costs.

Key Questions

What is Jev, as described in the guide?

Jev accepts text or JSON plus typed questions and returns structured answers, including probabilities, choices or scores with confidence. Meyer says software decides what action to take from those answers.

How many use cases does the guide identify?

Meyer maps 24 use cases: three he says are live, 12 strong fits, seven that need measurement and two he considers poor fits.

What results does Meyer report from live use?

He reports a language scan of 78,889 articles for $2.01 and 1,553 fixes among 1,576 non-English items found. He also reports 89% agreement for a 31-topic fallback classifier, rising to 97% to 99% at confidence of 0.8 or higher. These are figures from his operation, not independently verified benchmarks.

How does the guide recommend checking a use case?

Meyer recommends replaying 300 to 500 past decisions, comparing accuracy across confidence levels and reviewing 20 disagreements. He proposes proceeding when the high-confidence band reaches 95%, then using a feature flag and a 5% to 10% canary.

Source: ThorstenMeyerAI.com