24 Ways To Use Jev To Frame AI Decision Challenges

🔍 Read the full analysis: 24 Ways To Use Jev To Frame AI Decision Challenges on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

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 24-use-case map for Jev on Sept. 29, describing how the tool could handle narrow, high-volume decisions across publishing, commerce, software, business operations and home tasks. Meyer says three applications are already running in his publishing operation, with about 90,000 decisions made so far; his classification of the other ideas is based on a four-condition fit test he outlines in the article.

Jev takes text or JSON state and a set of typed questions, then returns answers that software can use to make a decision. The article says it does not write, summarize or extract information. Its answer types include a yes-or-no probability, a choice among options with probabilities and confidence, and a score on ordered levels. Meyer says a call takes about 0.3 to 0.9 seconds and costs about $0.04 per million input tokens.

Meyer reports three live publishing uses: checking whether a story fits a site, checking whether an article is in English, and classifying a headline when a primary language model fails. For the language check, he says a scan of 78,889 articles cost $2.01 and found 1,576 non-English articles, of which 1,553 were fixed. He reports 89% agreement with a frontier model for the classifier fallback, rising to 97% to 99% when Jev’s confidence was at least 0.8. These figures come from Meyer’s account; the article does not provide an independent evaluation.

Of the remaining ideas, Meyer labels 12 strong fits, seven “measure first” and two poor fits. Examples include disclosure checks and comment moderation as strong fits, a thin-source detector and headline-quality scoring as uses needing measurement, and same-event deduplication as a poor fit after a canary found no duplicates. The supplied article excerpt begins a section on commerce and customer operations but does not include its full list of use cases.

A Confidence Gate for Routine Decisions

The proposal matters to teams that handle large volumes of small classification or filtering tasks. If a system can reliably resolve straightforward cases and send uncertain ones to a person or a more capable model, teams may be able to apply a check across more items without treating every model output as authoritative. In Meyer’s account, confidence scores make that division possible: software sets the thresholds and determines what action follows.

The article also argues for testing whether a decision problem exists before adding another tool. Meyer’s same-event deduplication canary found no duplicates, so he calls it a poor fit for his operation. That example highlights a practical constraint: low cost and narrow questions do not establish value on their own. A use case needs a measurable error in the existing process, and the cost of a wrong answer must be manageable.

How Meyer Tests Jev’s Fit

Meyer’s four conditions are high volume, a narrow question, cheap errors and a visibly failing heuristic. He recommends replaying 300 to 500 past decisions in shadow mode, comparing results overall and by confidence band, and reviewing 20 disagreements to judge which answer was correct. He says a team should wire in a use only where the high-confidence band reaches 95%, then test it behind a separate flag on 5% to 10% of units before wider rollout.

The three reported live uses illustrate that approach. For site relevance, Meyer says the system judged about 10,000 story-and-site pairings over three days and found only 22% clearly on-topic. The rule drops a story only when its fit score is clearly low and Jev is confident; uncertain cases continue along the previous publishing path. For language detection, the article describes rewriting non-English articles in place at the same URL when the model’s yes-or-no probability falls below 0.1.

Meyer also reports a 31-topic classification measurement in which Jev agreed with a frontier language model 97% to 99% of the time at confidence of 0.8 or higher, and 42% of the time below 0.5. Agreement with another model is a comparison, not proof that either answer is correct. The article’s recommended review of disagreements is intended to test that distinction on a team’s own examples.

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

— Thorsten Meyer

Independent Validation Is Missing

The performance, cost and volume figures in the article are Meyer’s reported results. The supplied material does not include independent validation, detailed evaluation methods, or the underlying data needed to check the reported accuracy and savings. Agreement with a frontier model does not establish accuracy against a verified answer key.

The source excerpt also ends partway through the commerce and customer-operations section. It therefore does not show all 24 use cases or enough detail to assess every proposed application. For the seven ideas marked “measure first,” Meyer says the existing heuristic’s failure has not yet been proven. The article does not provide results from broader deployments of those proposals.

Measure Before Wider Rollout

Meyer’s next step for teams considering a use is to test it against past decisions in shadow mode, review disagreements and confirm that high-confidence answers reach the stated 95% threshold. He recommends enabling a successful application behind a separate flag, starting with a 5% to 10% canary, then expanding its use if the results hold. The article does not announce a specific rollout date or an independent benchmark.

Key Questions

What does Jev do?

Jev returns typed, confidence-scored answers to questions about supplied text or JSON. According to Meyer, software can use those answers to branch into actions such as filtering, routing or review.

How many Jev use cases does Meyer say are ready?

Meyer says three are live in his publishing operation and 12 more meet his criteria for strong fits. He classifies seven as needing measurement first and two as poor fits.

What evidence does the article give for live use?

Meyer reports a scan of 78,889 articles costing $2.01, with 1,576 non-English articles found and 1,553 fixed. He also reports results for site relevance and a classifier fallback; the figures are attributed to his account.

How does Meyer recommend testing a proposed use?

He recommends replaying 300 to 500 past decisions in shadow mode, comparing confidence bands and reviewing 20 disagreements. He says to proceed where the high-confidence band reaches 95%, then start with a 5% to 10% canary.

Are all 24 use cases detailed in the supplied article?

No. The source material provided here stops partway through the commerce and customer-operations section, so the remaining use cases are not available to assess.

Source: ThorstenMeyerAI.com