Jev In Your AI Toolkit: 24 Ways To Model Decisions

🔍 Read the full analysis: Jev In Your AI Toolkit: 24 Ways To Model Decisions on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

• 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 uses for Jev, a tool that returns structured answers to narrow questions so software can make routine decisions. In the Sept. 29 article, Meyer says three uses are live in his publishing operation, while 12 more meet his criteria for a strong fit; the figures and results are based on his own account.

Meyer describes Jev as a system for sending text or JSON state alongside typed questions and receiving answers that code can use. Its response types include a yes-or-no probability, a choice with probabilities and confidence, or a score on ordered levels. It does not write or summarize text, according to the article; the application’s code decides what to do with each answer. Meyer says a call takes about 0.3 to 0.9 seconds and costs about $0.04 per million input tokens.

Three publishing checks are described as running in Meyer’s operation: a relevance gate for matching stories to a site, an English-language check, and a fallback topic classifier. Meyer reports that a scan of 78,889 articles cost $2.01; the language check found 1,576 non-English articles and fixed 1,553. He also reports about 10,000 story-site pairings assessed in three days and 89% agreement between the fallback classifier and a frontier large language model, rising to 97%–99% when Jev’s confidence was at least 0.8. These are figures supplied by Meyer, and the article does not provide independent verification or full evaluation methods.

The article assigns the 24 proposed uses to four categories: live, strong fit, measure first, and poor fit. In publishing, examples include disclosure checks and comment moderation as strong fits, while a thin-source detector and headline-quality scoring need measurement first. Meyer says same-event deduplication is a poor fit for now because a canary found no duplicates to address. The supplied article excerpt also describes the framework and some publishing examples, but ends as it begins its commerce and customer-operations section.

A Test Before Adding Automation

The practical question raised by Meyer’s list is whether a task occurs often enough, is narrow enough, and has low enough error costs to justify automated decisions. His proposed approach is to act on high-confidence cases and send uncertain ones to a person or a more capable system. That can make routine checks cheaper to run broadly while keeping a fallback for ambiguous cases.

The framework also puts measurement ahead of implementation. Meyer says teams should replay 300 to 500 past decisions, compare results overall and across confidence bands, and review 20 disagreements. He recommends wiring a use into production only if the high-confidence group reaches 95%, then starting with a feature flag and a 5%–10% canary. Those thresholds are his recommendations, not an independently established standard.

For readers considering similar tools, the distinction between a plausible use and a measured problem matters. A classifier may be inexpensive and fast, but if existing rules work or the canary finds no errors to fix, automation may add complexity without a demonstrated benefit. Meyer’s own deduplication example illustrates that constraint.

How Jev Routes Routine Judgments

Meyer’s central distinction is between generating content and classifying a state. A user supplies a story, page, or other state together with questions; Jev returns typed outputs such as a probability that text is English or a choice among topics. The application then applies a rule. For example, a low-confidence answer can be sent to human review rather than used automatically.

For a 31-topic classification, Meyer reports that Jev agreed with a frontier model 97%–99% of the time at confidence of 0.8 or higher, and 42% of the time below 0.5. The article presents this as his measurement. It does not specify the dataset, sample size, model identity, or how agreement was calculated, so the figures should be read as reported results rather than a general performance guarantee.

The uses described span publishing, commerce, software, business operations, and the home. The available source details the publishing examples and begins the commerce section, but does not include the remaining use-case tables. It therefore supports the overall count and category breakdown as Meyer’s claims, but not a full account of all 24 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

Results Need More Evaluation Detail

The published account does not provide independent testing of Jev’s cost, speed, or accuracy figures. It also leaves out details needed to assess the reported model comparison, including the test sample and evaluation method. The reported results cannot establish performance across other datasets or organizations.

Meyer says seven uses need measurement because a failing heuristic has not been demonstrated, and two are poor fits. The supplied material does not identify the two poor-fit cases or give the full list of 24 uses. It is also unclear how the live checks perform over time beyond the specific counts Meyer reports.

Measure Before Wider Rollout

Meyer’s recommended next step for a candidate use is to replay a sample of real past decisions, review disagreements, and check accuracy by confidence level. If the high-confidence group meets his 95% threshold, he advises testing behind a feature flag on 5%–10% of units before expanding deployment. Whether teams follow those steps or obtain similar results remains unknown.

The source describes three uses already running in Meyer’s publishing operation, but it gives no schedule for further deployments or independent evaluation. Further details on the remaining use cases, their measurement results, and any wider rollout were not available in the supplied material.

Key Questions

What is Jev?

Meyer describes Jev as a tool that takes text or JSON state and typed questions, then returns structured answers such as probabilities, category choices, or scores. Software applies the resulting rules.

How many Jev uses does Meyer say are ready or running?

Meyer says three uses are live and 12 additional uses are strong fits. He classifies seven as needing measurement first and two as poor fits. These are his assessments.

What results does Meyer report from the publishing checks?

He reports a scan of 78,889 articles for $2.01, finding 1,576 non-English articles and fixing 1,553. The article does not include independent verification or the full evaluation details.

What test does Meyer recommend before deployment?

He recommends replaying 300 to 500 past decisions, reviewing disagreements, and checking accuracy by confidence band. His suggested deployment threshold is 95% accuracy in the high-confidence band, followed by a limited canary.

Does the article provide all 24 use cases?

No. The supplied source gives the overall category counts and several publishing examples, then cuts off as the commerce section begins. The remaining use-case details are not included.

Source: ThorstenMeyerAI.com