Jev For AI Decisions: 24 Methods, Workflows, And Use Cases

🔍 Read the full analysis: Jev For AI Decisions: 24 Methods, Workflows, And Use Cases on ThorstenMeyerAI.com

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

Get office and shipping supplies 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 map of **24 potential Jev uses** on Sept. 29, reporting that **three are already running** in his publishing operation and that 12 more meet his stated fit test. Jev returns typed answers that software can use to make narrow decisions. The article describes his criteria for assessing where the system might be used in high-volume workflows.Meyer divides the 24 cases into **three live uses**, **12 strong fits**, seven that need measurement and two poor fits. He says the live workflows have processed about 90,000 decisions so far. The available source describes the publishing examples and the method for assessing fit, but its account of other sectors ends partway through a commerce section; it does not provide the full list of 24 cases.

Meyer describes Jev as a system that receives text or JSON state alongside typed questions and returns answers that code can branch on, rather than prose to interpret. The answer types include a yes-or-no probability, a choice with probabilities and confidence, or a score on ordered levels. He reports that one call takes about **0.3 to 0.9 seconds** and costs about **$0.04 per million input tokens**.

In his publishing operation, Meyer says a relevance gate assessed about **10,000 story-site pairings in three days**, with 22% judged clearly on-topic. A language check scanned 78,889 articles for $2.01, found 1,576 non-English items and fixed 1,553. A classifier fallback agreed with a frontier LLM 89% of the time overall, according to Meyer; in a 31-topic measurement, he says agreement was 97% to 99% at confidence of 0.8 or higher and 42% below 0.5.

A Test Before Adding Automation

Meyer’s proposed test for automation requires high volume, a narrow decision, errors that are cheap or routed to a stronger reviewer, and evidence that an existing heuristic visibly fails. He recommends retaining a working keyword rule unless measurement shows a gap.

Meyer recommends replaying **300 to 500 past decisions**, comparing results overall and by confidence band, then reviewing 20 disagreements to determine which system was right. He says to use a workflow only where the high-confidence band reaches 95%, with a separate flag, disabled by default, and a canary on 5% to 10% of units.

In the relevance gate, Meyer says low-fit cases are dropped only when Jev is confident; uncertain cases continue to publish as before. The article reports these figures as Meyer’s results and does not provide an independent evaluation.

How Jev Fits Publishing Checks

Meyer presents the three running workflows as examples of checks made practical by low-cost, repeated decisions. The **relevance gate** combines two yes-or-no questions and a fit score to compare a story with a site profile. The **language check** asks whether an article title and opening paragraph are in English; Meyer says items below a 0.1 threshold are rewritten in place at the same URL. The **classifier fallback** chooses among 31 topic descriptions when the primary LLM errs, with a keyword rule as a last resort.

His fit labels distinguish deployed uses from potential ones. “Live” means running in his fleet; “strong fit” means all four conditions are met; “measure first” means the failure of the existing heuristic has not been proven; and “poor fit” means at least one condition fails. In publishing, he lists **disclosure detection** and **comment moderation** as strong fits. Product matching in roundups, headline quality and thin-source detection need measurement first. Same-event deduplication is a poor fit in his canary because he found zero duplicates.

The examples show how the test applies to existing rules. Meyer says a 300-character rule cannot distinguish a fact-dense wire item from a teaser, while a regex may miss paraphrased disclosures about free products or affiliate links. For disclosure checks, his suggested rule sends misses to human review and bars automatic publication. For moderation, he proposes auto-approving acceptable comments or hiding spam only above a 0.9 confidence threshold, with other cases queued.

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

— Thorsten Meyer, in the Sept. 29 article

Evidence Still Needed Across Uses

The reported performance figures come from Meyer’s own measurements; the source does not describe an independent audit, detailed evaluation method or comparison dataset for the production workflows. The reported agreement with a frontier LLM is a comparison with that model, not a measure of verified correctness against ground truth. The article also does not establish how results vary across different publishers, content types or operating conditions.

Seven of the mapped uses still need evidence that the existing heuristic fails, by Meyer’s own classification. The source says his deduplication canary found zero duplicates, but gives no sample size or time window. It also ends during the commerce and customer operations section, leaving the remaining use cases and their specific rules unavailable in the supplied material. The stated total of 24 can be reported, but the full inventory cannot be reconstructed from this excerpt.

Measure Before Wider Rollout

Meyer recommends that prospective users replay **300 to 500 historical decisions** and examine 20 disagreements before enabling a workflow. If the high-confidence band meets his 95% threshold, he recommends putting the integration behind its own flag, beginning with a 5% to 10% canary and expanding from there. These are recommendations in his article; the source does not announce a broader product launch or a scheduled follow-up milestone.

For the seven “measure first” cases, the next evidence would be a measured error rate for the current rule or matcher. Until that is available, Meyer’s framework classifies the cases as promising but unproven. The supplied article excerpt does not say when he will publish the complete list or provide further results.

Key Questions

What did Thorsten Meyer publish?

He published a guide mapping **24 potential Jev use cases** across publishing, commerce, software, business operations and the home. The provided source excerpt details publishing cases and only begins its commerce section.

How many use cases does Meyer say are ready or running?

He says **three are live** in his publishing operation and 12 additional cases are strong fits. Seven need measurement, while two are poor fits. These classifications are Meyer’s assessment.

What does Jev return?

According to Meyer, Jev returns typed answers such as yes-or-no probabilities, choices with confidence, or scores on ordered levels. Software can use those outputs to branch without parsing a prose response.

What evidence does Meyer recommend before deployment?

He recommends replaying **300 to 500 real past decisions**, checking results by confidence band and reviewing 20 disagreements. His proposed rollout threshold is 95% in the high-confidence band, followed by a limited canary.

Source: ThorstenMeyerAI.com

×HALLOWEEN

Halloween Picks

halloween

halloween

View Latest Price

As an affiliate, we earn on qualifying purchases.