What Does Jev Bring To AI Decision Models? 24 Ideas

🔍 Read the full analysis: What Does Jev Bring To AI Decision Models? 24 Ideas on ThorstenMeyerAI.com

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

Get smart everyday buys 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 September 29 account of 24 ways to use Jev, a tool that answers typed questions about text or JSON so other software can make decisions. Meyer says three applications are already running in his publishing operation, 12 others meet his fit criteria, seven need measurement and two are poor fits. He recommends acting on high-confidence answers and routing uncertain cases for further review.

Meyer describes Jev as a decision tool rather than a writing or summarization system. A request contains a state and typed questions; the response provides answers such as a yes-or-no probability, a choice among options with probabilities, or a score on ordered levels. The application’s code determines what to do with those answers. Meyer reports that a call takes about 0.3 to 0.9 seconds and costs about $0.04 per million input tokens.

Three uses are live in Meyer’s publishing operation. A relevance gate assessed about 10,000 story-and-site pairings in three days, with 22% judged clearly on-topic; the rule drops a story only when Jev is confident it fits poorly. A language check scanned 78,889 articles for $2.01, Meyer says, identified 1,576 non-English articles and fixed 1,553 by rewriting them in place. A fallback classifier, used when a primary language model errs, achieved 89% agreement with a frontier model overall and 97% to 99% agreement at confidence of 0.8 or higher, according to Meyer.

The remaining examples span publishing, commerce, software, business operations and the home. Publishing ideas include detecting whether a source contains enough verifiable facts, checking product relevance in roundups, reviewing disclosures and moderating comments. Meyer marks disclosure checks and comment moderation as strong fits. He says thin-source detection, product matching and headline-quality review need measurement first; same-event deduplication is a poor fit because his canary test found no duplicates.

Where Confidence Can Save Review Time

Meyer proposes using a low-cost classifier for large volumes of narrow decisions, while retaining human or stronger-model review for uncertain cases. In publishing, this could allow routine checks such as language, disclosure wording or comment categories to be applied across many items. Jev supplies an answer; the system owner determines whether to publish, hold, rewrite or escalate.

Meyer reports that high-confidence classification aligned closely with a frontier model in one 31-topic measurement, while agreement was lower at low confidence. The result supports evaluating confidence-based routing in that test, but does not establish accuracy across other tasks or organizations. Meyer recommends measuring an existing process before replacing it, since automation may not improve a process that already works or lacks a demonstrated error to address.

Meyer’s Four Checks for Fit

Meyer says a use should pass four conditions before Jev is wired into a workflow: it must involve high volume, ask a narrow question, have cheap errors or a route for uncertain cases, and address a heuristic that is visibly failing. He advises replaying 300 to 500 past decisions, comparing results overall and by confidence band, and reviewing 20 disagreements to judge which system was right. He recommends integration only where the high-confidence band reaches 95% in that test.

For rollout, Meyer proposes a separate feature flag that is off by default, followed by a canary on 5% to 10% of units. His categories distinguish uses he says are live from strong fits that meet all four tests, cases that still need measurement, and poor fits that fail at least one condition. The article’s supplied examples detail publishing uses; it introduces additional areas but the available source text ends during the commerce section, so it does not substantiate all 24 proposals.

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

— Thorsten Meyer

Evidence Beyond Meyer’s Tests

The reported accuracy, cost, speed and production results come from Meyer’s own measurements. The supplied material does not identify an independent evaluator, describe the full testing methodology or establish that the same results would hold for other users and tasks. The 97% to 99% agreement figure applies to a specific 31-topic classification and confidence threshold, not every Jev use.

The source text provided for this article stops partway through its commerce section. It therefore does not give enough detail to verify the full set of 24 examples, the exact distribution across all five areas or the rules for every proposed use. Meyer also says seven cases need measurement and identifies a failed deduplication canary, underscoring that several suggestions remain hypotheses rather than demonstrated improvements.

Measure Before Wider Deployment

Meyer’s proposed next step for each unproven use is to replay 300 to 500 real past decisions, compare results by confidence band and inspect disagreements before integrating the tool. Where results meet his stated threshold, he recommends a feature flag and a 5% to 10% canary rollout. The article does not set a publication date for further results or describe a broader independent trial, so the status of the seven measure-first cases remains open.

Key Questions

What does Jev do?

According to Meyer, Jev takes text or JSON and typed questions, then returns answers such as probabilities, category choices or scores. The calling software uses those answers to decide what action to take.

How many Jev uses does Meyer say are in production?

Meyer says three uses are live in his publishing operation: a relevance gate, a language check and a fallback topic classifier.

What is Meyer’s test for whether a use fits?

He calls for high volume, a narrow question, low-cost errors or escalation for uncertain answers, and evidence that the existing heuristic is failing.

Are the reported accuracy figures independently verified?

The figures in the supplied article are Meyer’s reported measurements. The material provided does not describe independent verification or show that the results generalize to other systems.

Which proposed applications still need testing?

Meyer labels seven uses as requiring measurement, including thin-source detection, product matching and headline-quality review. He says a same-event deduplication canary found no duplicates, making that proposal a poor fit in his test.

Source: ThorstenMeyerAI.com