24 Ways To Use Jev For Decisions Across AI Projects

🔍 Read the full analysis: 24 Ways To Use Jev For Decisions Across AI Projects on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Thorsten Meyer published a map of 24 ways to use Jev for small decisions across AI projects, reporting that three publishing workflows are already running in his operation. He says 12 use cases meet his four-condition fit test, seven require measurement and two are poor fits; the material available here describes the publishing cases but does not show the remaining examples.

Meyer describes Jev as a decision tool that takes a text or JSON state and typed questions, then returns answers that software can use to choose a next step. He says it does not write, summarize or extract information. The stated answer types are a yes probability, a choice among options with probabilities and confidence, or a score on ordered levels. Meyer reports that one call takes about 0.3 to 0.9 seconds and costs about $0.04 per million input tokens.

The central design principle is to act on clear answers and send uncertain cases to a more capable system or a person. Meyer reports that, in a 31-topic classification measurement, Jev agreed with a frontier LLM 97% to 99% of the time at confidence of 0.8 or higher, compared with 42% below 0.5. Those figures are his reported measurement; the source does not provide the dataset, evaluation method or independent replication.

For the three live publishing uses, Meyer reports scanning 78,889 articles for $2.01 with the language check. It found 1,576 non-English articles, of which 1,553 were fixed by rewriting in place at the same URL. A relevance gate assessed about 10,000 site-and-story pairings over three days, with 22% deemed clearly on-topic. A classifier fallback reportedly agreed with a frontier LLM 89% of the time overall, and 97% to 99% in its high-confidence band.

Where Cheap Decisions May Help

Meyer says Jev could be used to apply low-cost checks to large volumes of small decisions that might otherwise rely on keyword rules or go unchecked. In the publishing examples he describes, the language scan identified articles for correction, while the relevance screen was designed to remove matches assessed as clearly off-topic and leave uncertain cases on the existing publishing path.

The confidence threshold and fallback rule determine how results are used. A probability from the tool does not itself determine whether content is published, hidden or corrected; Meyer says application code makes that decision. He says publishers should measure performance on real items in the confidence band that triggers automatic action and assess whether the cost of errors is acceptable.

Meyer also says low cost alone is not sufficient reason to deploy the tool. A check that does not identify a meaningful problem would not improve the workflow. The proposed use cases span settings including publishing, commerce and household decisions, where error costs and fallback options may differ.

Meyer’s Four Conditions

Meyer says Jev should be used only when four conditions hold: high volume, a narrow question, low-cost errors or a route for uncertain cases, and a heuristic that demonstrably fails. He recommends testing against 300 to 500 past decisions, comparing results overall and by confidence band, then reviewing 20 disagreements to determine which answer was right. His stated deployment bar is at least 95% accuracy in the high-confidence band.

His proposed rollout is to put the feature behind a separate flag that is off by default, try it on 5% to 10% of units, and then expand. The source presents this as Meyer’s recommended process, not a report of a controlled study or a guarantee that the same threshold will suit every use case.

The publishing examples show how the tags are meant to work. The thin-source detector is marked “measure first”: it would judge whether a source contains enough verifiable facts to support a report, but its current 300-character rule has not yet been shown to fail through a measured comparison. Product matching for roundups and headline quality also need baseline error measurements. By contrast, disclosure checks and comment moderation are listed as strong fits. Same-event deduplication is a poor fit because Meyer’s canary found no duplicates, so he says a present problem has not been demonstrated.

“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 Behind the Use Map

The supplied material does not include the complete list of 24 cases. It begins a section on commerce and customer operations but cuts off before detailing those examples; the announced software, business operations and home cases are also absent. The stated totals—three live, 12 strong fits, seven needing measurement and two poor fits—therefore cannot be checked against a full case-by-case list here.

The source gives Meyer’s results and estimates but no underlying records, benchmark procedure, cost calculation details or independent assessment. It is also unclear how the 31-topic comparison was assembled, what counted as agreement, and whether the reported accuracy holds outside that measurement. The figures should be read as claims about his own tests and operation, not as established performance across other systems.

Measure Before Wider Rollout

Meyer’s stated next step for unproven applications is to replay several hundred real past decisions, examine disagreements and verify performance by confidence band. A use case should move toward deployment only if the existing heuristic is shown to fail and the high-confidence results meet the proposed threshold. Any rollout would then begin behind a disabled-by-default flag and with a small canary group.

The source does not give a publication date for further results, a schedule for rolling out the other proposed cases or details of the remaining 15 workflows. Whether the seven “measure first” cases become strong fits will depend on those measurements; the article’s available figures do not settle that question.

Key Questions

What is Jev, according to Meyer?

Meyer describes Jev as a tool that receives a text or JSON state and typed questions, then returns probabilities, choices or scores that code can act on. He says it does not generate prose.

How many of the proposed uses are live?

Meyer says three are running in his publishing operation. He classifies 12 as strong fits, seven as needing measurement and two as poor fits.

What publishing results does the source report?

Meyer reports that a language check scanned 78,889 articles for $2.01, found 1,576 non-English articles and fixed 1,553. He also reports results for a relevance gate and a classifier fallback; these are figures from his account.

What should be tested before adopting Jev?

Meyer recommends replaying 300 to 500 past decisions, checking performance by confidence band and reviewing disagreements. He says a failing heuristic should be demonstrated before deployment.

Are all 24 use cases available in the source?

No. The supplied text details the live publishing cases and several proposed publishing checks, then cuts off as the commerce section begins. The other announced examples are not provided.

Source: ThorstenMeyerAI.com