Jev: understand, compare and test TypeSafe's model

Jev, launched in early access by TypeSafe on September 15, 2026, evaluates questions defined by the application and estimates the probabilities of possible answers.

Choice also provides a selected option; Score, a weighted score; Noul, the probability of yes.

Code combines these results, applies rules and thresholds, and decides whether to route a ticket, call a tool or request a review. A generative model can then write the response for the person.

TL;DR

My assessment, dated September 22, 2026 and still standing on October 1: Jev is worth testing, with claims that need qualification.

The negative POCs of September 27 below concern only Laya running locally, not the Jev API. I have not tested the API myself yet, so they neither confirm nor overturn this assessment.

Classification predates Jev: the value of its API depends on errors, cost and latency for a specific task, and the sources reviewed establish neither a scientific revolution nor the “scam” allegations.

The full report above brings together the analysis, sources and examples. Both the French and English editions include full-content search, videos linked to the cited moments and code examples you can copy.

Aristote pilot added September 27 (Laya, local only): keep the current rules. On 60 test tickets, Laya emits 11 area suggestions, 8 matching provisional annotations, compared with 17 and 12 for title-only rules.

It leaves 49 tickets without a suggestion, including 8 that exceed context. The annotations from 2 agents have no human validation. This result rejects the tested configuration; no Jev benchmark was performed.

The 3 other POCs have run (Laya, local only): stop the tested configurations. Across 120 held-out cases, Laya proposes 2 matching links out of 38 for Sentry, 1 out of 9 for reviews and 6 matching classifications out of 20 for CI. CI rules reach 5 out of 6.

Both matching Sentry links were already explicitly associated. Annotations remain provisional and 15 inputs exceed context. Read the results and limitations.

The September 26 synthetic experiments remain separate: triage at 14 correct answers out of 20, versus 15 for rules, and an 8.55 ms warm filter median.

Accepted Aristote tickets have a 131.2 ms engine median. Inputs, question and measurement scope differ; these timings are not interchangeable.

Where to start

• Understand how it works: Choice, Score and Noul, followed by the differences between probabilities, confidence and calibration.
• Assess the claims: classifiers, marketing and “scam” allegations, with arguments in favour, criticism and the limits of each.
• Try it in code: install the SDK and send a first ticket. The example covers API keys, errors and precautions before a billable call.
• Explore the projects: jgrep and semantic tools, then local Laya-MLX and Kev.
• Try it locally: our results and limitations, then install Laya on an Apple Silicon Mac, without an AI API after downloading.
• Read the test on real tickets: Aristote pilot, protocol and verdict, including all 3 comparators and annotation limits.

Read the numbers in context

• 70 to 500 ms per call: TypeSafe’s claimed range, dependent on the task and access conditions.
• $0.042 per million input tokens, with free outputs: price checked on September 21, 2026. At 1,000 tokens per request, 100,000 requests would cost $4.20 for Jev alone, excluding retries and other services.
• Roughly 200× faster and 445× cheaper: results from TypeSafe’s launch comparisons, not guaranteed gains for every application.

The report compares those claims with independent tests and explains how to count costs across the whole workflow.

A structured response can still be wrong; its format does not remove the need to measure quality.

Inside the report

• 32 chapters and 152 sources covering the API, publicly described architecture, training methods, benchmarks and open alternatives.
• Video analyses, including Theo and Fireship, with timestamped links to demonstrations and criticism.
• Proposed tests for GitHub, Sentry, release notes, YouTube Insights and Paper Insights.
• 17 repositories added on September 24: filtering, sorting and matching commands, context selection, review integrations and local models. Documented features remain separate from authors’ measurements.
• Diagrams, a glossary and a protocol for comparing Jev with existing rules, classifiers and LLMs with constrained output.
• Our September 26 Laya experiments: offline operation, quality across 60 synthetic predictions, timing, browser errors and a public walkthrough. The report includes the evidence data.
• The Aristote pilot: 30 development tickets, 60 test tickets, precision, coverage, abstentions and overflows. The aggregates and configuration are available without private issue text or identifiers.

An evolving report

Version 0.19.0 adds the Sentry, review and CI POCs, their comparators and collection limitations. For CI, Laya provides 1 more matching suggestion in total, with 13 additional errors. All 3 candidates are rejected.

The aggregate file contains metrics and selected questions without private per-case data. The reproduction folder’s evidence check was rerun without new model inference.

Version 0.18.0 adds the negative Aristote area-triage result. Laya’s 8 matching suggestions out of 11 do not compensate for low coverage.

The title-and-body lexical variant reaches 40 out of 51, with 11 errors; it is not deployed and needs human validation. Sentry, review and CI POCs were still in preparation in version 0.18.0; their results are included in version 0.19.0.

Version 0.17.0 added the M5 Max synthetic measurements and walkthrough with pinned software and weights. The 62 program tests and 6 browser workflows check operation, not classification quality. The lab’s JSON download remains unconfirmed in the integrated browser.

The September 24 collection keeps its own date: JevKit tools, jgrep and Kiln experiments, Laya-MLX, Kev, official recipes and the Python SDK 0.7.1 release notice.

It added 5 uses, 6 proposed trials and clarifications of the ByteByteGo graphic without executing those projects at that time.

Star counts retain their September 24 observation date.

The distinction between Jev evaluations and application decisions remains explicit. The 4 original infographics are available in English and French.

The tutorial remains pinned to SDK 0.7.0 and checked with a mock transport; this does not measure model performance. Earlier results retain their dates and limitations.

The report shows its last update at the top. Its revision history records additions and corrections.

Have a question, a test result or a source to contribute? Contact me on LinkedIn.