24 Practical Applications Of Jev For AI Decision Models

🔍 Read the full analysis: 24 Practical Applications Of Jev For AI Decision Models on ThorstenMeyerAI.com

Thorsten Meyer published a Sept. 29 article mapping 24 possible applications of Jev, a tool for answering narrow questions that software can use to make decisions. He says three uses are already live in his publishing operation, which has processed about 90,000 decisions, and identifies 12 additional strong fits, seven cases that need measurement and two poor fits.

Meyer describes Jev as a system that receives text or JSON and typed questions, then returns structured answers rather than prose. Its answer types include a yes-or-no probability, a choice among options with probabilities and confidence, or a score on ordered levels. The software around the tool decides what action to take. A single call containing the state and questions takes about 0.3 to 0.9 seconds, and Meyer gives a cost of about $0.04 per million input tokens.

The three live examples are a relevance gate for matching stories to a site, a language check, and a fallback topic classifier. Meyer reports that the relevance gate assessed about 10,000 story and site pairings in three days, with 22% judged clearly on-topic. In the language check, he says a scan of 78,889 articles cost $2.01 and found 1,576 non-English articles; 1,553 were fixed by rewriting them in place at the same URL. For the classifier fallback, he reports 89% agreement with a frontier LLM overall and 97% to 99% when Jev’s confidence was at least 0.8.

The supplied article excerpt also describes six publishing applications. They include checks for whether a source has enough verifiable facts, whether products fit a roundup, whether a disclosure is present, headline quality and comment moderation. Meyer labels the disclosure check and moderation strong fits. He says the thin-source detector, product-fit check and headline check need measurement first. Same-event deduplication is marked a poor fit because a canary test found no duplicates.

Where Confidence Changes the Workflow

Meyer describes Jev as a way to apply narrow checks across many items and route uncertain results to people or more capable models. The article gives language and disclosure checks as examples of issues that could be reviewed in publishing workflows. Meyer says the tool supports software decisions and does not write or verify reports by itself.

Meyer reports that Jev agreed with a frontier LLM 97% to 99% of the time at confidence of 0.8 or higher, compared with 42% agreement below 0.5, in a 31-topic classification. These figures measure agreement with that model, not accuracy against an independent ground truth. The excerpt does not establish whether either system’s classifications were correct.

Meyer’s fit categories distinguish live deployments from proposed uses and cases requiring further measurement. He recommends testing past decisions and first checking whether an existing heuristic fails visibly. The excerpt does not provide results for all proposed applications.

The Four Tests Behind Jev’s Uses

Meyer says an application should meet four conditions: high volume, a narrow question without multi-step reasoning, errors that are cheap or can be routed to a stronger system, and a visibly failing existing heuristic. The fourth condition is the one he says remains unproven for seven proposed uses. His list therefore distinguishes measured live deployments from strong fits and ideas that still need evidence.

For a pre-deployment check, Meyer recommends replaying 300 to 500 past decisions, comparing results overall and within confidence bands, and reviewing 20 disagreements to judge which system was right. He says to wire Jev in only where the high-confidence band reaches 95%. The proposed rollout uses a separate flag that is off by default, followed by a canary covering 5% to 10% of units.

The reported language scan and topic classifier offer different kinds of evidence. The scan has a stated number of articles, cost, and corrections. The classifier result is an agreement rate against another model. Those measures answer different questions, and the excerpt does not provide an independent audit of the corrections or labels.

“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 for Wider Use

The article excerpt supplied for this report ends partway through its section on commerce and customer operations. It states that the article maps 24 applications, but the available text gives details for only the three live uses and six publishing cases. The remaining applications, including the full commerce, software, business operations and home lists, are not available here, so their questions, rules and fit labels cannot be assessed.

The excerpt does not identify the full methods behind the reported measurements, provide the underlying decisions for independent review, or say how often the live systems send uncertain cases to people or stronger models. The classifier comparison is against a frontier LLM; the text does not report a human-reviewed accuracy baseline. The cost figure is stated per million input tokens, but the excerpt does not give enough usage detail to calculate a typical total cost per workflow.

For several proposed checks, the main unknown is whether an existing system actually makes enough mistakes to justify a change. Meyer gives same-event deduplication as an example: the canary found zero duplicates, so the article classifies it as a poor fit for now. The supplied material does not provide measured error rates for the product matcher, thin-source rule or headline assessment.

Measure Before Expanding the Rollout

Meyer’s proposed next step for unproven applications is to compare Jev with real past decisions before connecting it to a live workflow. Teams would review disagreements, check performance at different confidence levels and proceed only when the high-confidence results meet the article’s stated 95% threshold. The suggested rollout then starts with a feature flag and a small canary covering 5% to 10% of units.

For the seven cases tagged measure first, the key milestone is evidence that the current heuristic fails visibly and that the candidate check improves the process. The article’s examples imply that different checks will need different outcome measures: detection and correction counts for a language scan, for example, or reviewed labels for classification. No later deployment schedule or expansion plan is stated in the supplied excerpt.

Key Questions

What did Meyer report about Jev?

In a Sept. 29 article, Meyer mapped 24 possible applications and said three are live in his publishing operation. He grouped the other uses as strong fits, cases needing measurement, or poor fits.

What does Jev return?

According to Meyer, Jev returns typed answers such as a yes-or-no probability, a choice among options, or a score. The calling software uses those answers to decide what to do.

How did Meyer say teams should test an application?

He recommends replaying 300 to 500 past decisions, comparing results across confidence bands and reviewing 20 disagreements. His stated threshold for wiring a use into a workflow is 95% in the high-confidence band.

Are all 24 applications detailed in the available source?

No. The supplied excerpt describes the live publishing uses and six publishing proposals, then cuts off during the commerce section. Details for the remaining applications are not present in the material provided.

Source: ThorstenMeyerAI.com