24 Ways To Put Jev To Work On AI Decision Problems

🔍 Read the full analysis: 24 Ways To Put Jev To Work On AI Decision Problems 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 uses for Jev, a tool he describes as answering typed questions about text or JSON so software can make narrow decisions. Meyer says three uses are already running in his publishing operation, 12 meet his test for a strong fit, seven need measurement and two are poor fits. The examples matter to teams considering automated checks because the guide argues that Jev should handle only high-volume, low-cost decisions, with uncertain cases sent elsewhere.

Meyer reports that his three live uses cover a relevance gate for matching stories to a site, an English-language check and a fallback topic classifier. He says a scan of 78,889 articles cost $2.01 and found 1,576 non-English articles, of which 1,553 were fixed. For the classifier fallback, Meyer reports 89% agreement with a frontier large language model overall, rising to 97% to 99% when Jev’s confidence was at least 0.8. These are results he attributes to his own operation, rather than an independent evaluation.

The guide proposes three answer types: a yes-or-no probability, a choice among options with probabilities and confidence, or a score on ordered levels. Meyer says one call can carry the material to assess and multiple questions, and estimates that it takes about 0.3 to 0.9 seconds and costs about $0.04 per million input tokens. In his 31-topic classification measurement, he reports 97% to 99% agreement with the frontier model at confidence of 0.8 or higher, compared with 42% below 0.5.

For teams evaluating a use, Meyer recommends replaying 300 to 500 past decisions, comparing results by confidence band and reviewing 20 disagreements. He says to wire a use into production only if the high-confidence band reaches 95%, then enable it behind a separate flag, test it on 5% to 10% of units and expand gradually. His suggested rule is to act on clear answers and route uncertain ones to a person or a more capable system.

Where Automated Checks May Help

The guide’s central point is that a low-cost model call can make it practical to check many routine items, while a confidence threshold can limit automatic action to cases the system handles well. In Meyer’s publishing examples, that means checking every article for language or relevance and sending borderline cases through the existing process. The claimed cost and throughput could matter to organizations that handle large volumes of repetitive decisions, but the figures are specific to Meyer’s stated setup.

Meyer also distinguishes technical convenience from evidence of a useful application. A question can be narrow and cheap to answer while still solving no observed problem. He classifies headline-pair deduplication as a poor fit because a canary test found no duplicates. For disclosure checks, by contrast, he calls the use a strong fit because a missed affiliate or free-product disclosure can pose a compliance risk; he recommends routing potential misses for human review rather than allowing automatic publication.

The method also puts responsibility for action in the surrounding software. Jev returns an answer and confidence, while the application decides whether to publish, route, hide or request review. That division makes the operating rule and escalation path as important as the model’s reported accuracy.

Meyer’s Four Conditions for Fit

Meyer says a good Jev application must meet four conditions: high volume, a narrow question that does not require multi-step reasoning, errors that are inexpensive or can be escalated, and a heuristic that has been shown to fail. He advises keeping a simple keyword rule when it works, rather than adding an AI check without evidence of a gap.

The guide groups its 24 examples across publishing, commerce, software, business operations and household tasks. In the publishing set described in the source material, proposed checks include thin-source detection, product relevance in roundups, disclosure presence, headline quality and comment moderation. Their status varies: for example, Meyer’s disclosure and comment checks are tagged strong fits, while source sufficiency and headline quality require measurement first.

For the first live relevance gate, Meyer says about 10,000 story-and-site pairings were judged over three days, with only 22% clearly on-topic. He reports that the system drops a pairing only when fit is clearly low and confidence is high, leaving uncertain cases on the prior publishing path. He presents that limited-action design as a way to reduce the risk of automating borderline judgments.

“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 Operation

The reported scan, costs, timing and agreement rates are attributed to Meyer; the source material does not provide an independent audit, dataset or detailed evaluation method for those results. It is not clear how the measurements would transfer to different content, question wording, language mixes or model configurations.

The guide says seven use cases need measurement because a visibly failing heuristic has not been established, and two fail at least one of the four fit conditions. The supplied material gives detail on some publishing examples but not the full list of 24 applications or all results. It also does not identify a release plan, product changes or external validation tied to the article.

Measure Before Wider Rollout

Meyer recommends that anyone testing a proposed use replay real past decisions, compare Jev’s answers with the outcomes and review disagreements. If the high-confidence cases reach his stated 95% threshold, his proposed next steps are to place the feature behind a flag, start with a 5% to 10% canary and expand only after reviewing performance. The guide does not announce a specific future milestone or say whether Meyer plans to publish additional evaluation data.

Key Questions

What is Jev, according to the guide?

Meyer describes Jev as a tool that takes text or JSON plus typed questions and returns answers, probabilities or scores that software can use to branch. He says it does not write, summarize or extract prose.

How many applications does Meyer say are ready or running?

Meyer says three are live in his publishing operation and 12 more meet his four-condition test for a strong fit. He classifies seven as requiring measurement and two as poor fits.

What does Meyer recommend before deploying a use?

He recommends replaying 300 to 500 prior decisions, comparing results across confidence levels and reviewing 20 disagreements. He says to proceed only when the high-confidence band reaches 95%, then test through a small canary rollout.

Are the performance figures independently verified?

The source material attributes the figures to Meyer’s own publishing operation and describes one classification measurement. It does not provide an independent audit or enough methodology to establish how the results would generalize.

Source: ThorstenMeyerAI.com