Jev and Laya: a low-cost decision layer for AI workflows
Update, 20 October 2026: Clairic separates choices from extracted values
Cloudflare introduced Clairic in Workers AI preview on 19 October. A decision step selects from known options; extraction returns values such as an amount or date. This extends the article’s scope without implying that Jev and Clairic are interchangeable.
The model documentation lists typed results, uncertainty and a rate of $0.10 per million input tokens, with no output-token fee. Cloudflare warns that a valid type does not establish a correct value and recommends code for complex calculations. Open weights are still planned.
For an invoice pilot, separate what the document says, what we calculate and who accepts the result. The model can propose an amount and due date; code checks totals and permitted formats, and an accountant handles ambiguity. A predicted payment date must never be presented as a date printed on the invoice.
In n8n or Make, I would start with an HTTP step over anonymized examples, without accounting-system writes. Store the source document, schema, proposed value, uncertainty and later correction. Missing data belongs in a review queue rather than being filled with a plausible guess.
This is a proposed pilot, not a Clairic benchmark. Compare field-level errors, review time and cost per accepted invoice with your existing extraction. Classification and open-ended extraction need separate test sets.
October 11, 2026 update: Decisions API adds another routing option
Marek Bartoš highlighted Decisions API on October 10. OpenAI released the beta on October 6, before the video.
The official guide documents POST /v1/decisions, gpt-6-luna, text and image inputs, and predicate, choice and score questions. Extracting a custom JSON schema belongs in Responses API with Structured Outputs.
I would add photo-backed complaints to the routing pilot alongside Jev. My proposed n8n or Make HTTP step would receive minimized text and an image, choosing only between billing, damage, shipping and other. Application code would validate the category and hand API failures to a person. This is an integration proposal, not a tested connector.
Start with 200 human-labelled cases. Measure category confusion, cost including retries, and time to correct assignment. Compare the text-only subset separately: giving one system images changes the experiment. Calibrate each model's acceptance threshold independently. The classifier proposes a queue; a separate rule or person approves refunds.
The tests and prices below retain their stated September dates. They are not a current benchmark of all three options.
September 29, 2026 update: a concrete Czech review test
AI s rozumem published a practical Jev test on September 28. The video description mentions roughly 350 reviews; the public results application uses 275 reviews containing text, comparing Jev 1.13, Haiku 4.5 and Sonnet 5. Those sample counts should not be treated as interchangeable.
With Czech questions selected, the application reports median latency of 321 ms for Jev, 2.90 seconds for Haiku and 4.89 seconds for Sonnet. Latency comes from 25 sequential reviews sent from one Mac. Cost and throughput use the full set with eight concurrent requests. Estimated cost per 1,000 reviews is about $0.09 / $3.76 / $9.47. These are the author’s results for that setup, not our measurements or a service guarantee.
88% agreement between Jev and Sonnet is not 88% accuracy. Without reference labels, the application cannot establish which model is right. Agreement falls to 53% on whether a review is actionable for management. That task-level breakdown matters more for deployment than a single average.
For a company pilot, I would copy the method rather than a threshold: have a domain expert label a separate validation set, compare individual categories and route uncertain results to review. n8n records the proposed label and subsequent correction. Until false negatives are understood, the model proposes priorities but closes nothing automatically.
The application says names are anonymized for display only; models received the original text. In a company support pipeline, I would minimize data before the API call. This is a text-feedback analysis example, not an assessment of patients or quality of medical care.
The official TypeSafe skill supplies API context to coding agents. One useful documentation recommendation is to keep questions and thresholds together so a person can review them. Installing the skill does not establish classification accuracy.
September 26, 2026 update: Jev inside SQL and agent history
Jirka Herník’s September 25 video adds two concrete projects to the routing discussion. These are community implementations, not a new model release. The pg-jev changelog dates its initial version to September 17.
PostgreSQL: decisions near the data, inference outside it
pg-jev lets SQL filter and classify rows by meaning, but sends the data being evaluated to the TypeSafe API. It requires PostgreSQL 14–17, plpython3u and superuser access. According to its documentation, hosting that withholds those capabilities cannot run the extension.
For a support pilot, I would create a restricted view containing IDs and anonymised text from selected tickets. Jev would propose categories; n8n would save them to a separate table for staff to compare with actual resolutions. I would enable production routing only after evaluating the mistakes. This is a proposed integration, not a test performed here.
Watch the number of rows processed: an ordinary query now triggers paid network requests. The project provides per-statement row and character limits; set them before the pilot. Measure cost and latency on your own workload rather than treating the author's benchmark as a service guarantee.
Agent history: selecting evidence instead of rewriting everything
fast-jev-compaction asks Jev which older tool-call/result pairs to keep, shorten or remove. It is available as a library and a Claude Code plugin. Selection can still discard important evidence.
Conversation text and tool inputs go to the decision API, so this is not purely local housekeeping. I would start with a non-sensitive repository and compare identical tasks using original and compacted histories: completion, lost constraints, repeated reads and total cost. The integration needs a safe fallback for classification or API failures. Preserving the exact wording of retained messages does not guarantee that all necessary information survived.
September 24, 2026 update: Laya adds a local option
Jirka Herník’s September 23 episode introduces Laya alongside Jev. That is a fresh comparison signal, not a verified first-release date for Laya.
The project repository provides local text decisions, typed choices and probabilities, a self-hosted HTTP server and an Apache‑2.0 license. It can serve a routing step; it does not replace a generative model writing customer replies.
The demo site reports 21 ms on an M1 Max and a 500-example Jev comparison. These are the authors’ measurements, not our benchmark or proof of a universal “200× faster AI” claim. They also flag language routing, an uncalibrated multilingual checkpoint and difficulties comparing numbers.
For Czech and Slovak support, I would build separate validation sets containing typos, short messages and mixed languages. Let the model suggest a queue, code verify financial amounts and people handle uncertain cases. Start in shadow mode, recording proposed labels without changing actual routing.
Measure incorrect automatic decisions, escalation rates and staff time spent fixing mistakes. Avoiding a per-call API fee still leaves hardware and operating costs. Laya is worth testing when you need your own processing boundary; local deployment alone guarantees neither correct decisions nor a secure workflow.
Many automation steps do not need a paragraph. They need a decision: whether an email is a complaint, which queue should receive a lead, or whether a document needs manual review. Using a large generative model for every such choice can be wasteful.
TypeSafe introduced Jev on September 15, 2026, as a model for typed, probabilistic decisions. The launch is early access. Jirka Herník also discussed it in AI mastermind 40 on September 17. Original announcement.
A decision has a different contract from a chat response
Jev constrains the output to defined decisions rather than arbitrary prose. TypeSafe emphasises type correctness and probabilities.
A valid type does not guarantee a correct decision. “Complaint” can be an allowed label and still be wrong for a particular message. An evaluation therefore needs real business examples, not just a count of successfully parsed responses.
For support triage, I would start with a small vocabulary: billing, technical issue, complaint and other. “Other” is a legitimate route, not an embarrassment to eliminate at any cost.
A proposed n8n or Make integration
Give each incoming email an internal ID, remove unnecessary attachments and send only the information required for classification. The decision step selects a category; application code validates the result.
Clear cases enter the relevant queue. Ambiguous cases go to a stronger model or a person. A generative model drafts a reply later, if the process needs one.
Log the classification schema version, model, decision and any correction made by the support team. Otherwise you can count processed messages without knowing whether the system is getting them right.
Derive the threshold for automatic processing from a validation set. A reported probability may not be well calibrated on your data. “Automatically approve everything above 90%” is not a defensible universal default.
Evaluate the whole cost
TypeSafe quotes $0.042 per million input tokens, with free output. Its headline savings multiplier comes from four vendor-built workflows. It is neither an independent guarantee for your workload nor evidence that Jev replaces a frontier model everywhere.
Count first-pass decisions, escalations, retries and the human time spent correcting mistakes. A cheap step that sends too many errors into an expensive branch may lose overall.
A missed sales lead has a different cost from a mislabelled internal note. In an invoice process, I would let a decision model select a review queue; payment approval or supplier bank-account changes would remain separately controlled.
A pilot that can support a decision
Use several hundred historical cases, including ambiguous examples and rare categories. Keep the data used to design the schema separate from the final evaluation set. Have a domain owner confirm reference labels instead of treating another model as ground truth.
Track category confusion, escalation rates, the error rate among automatically accepted decisions and the cost per correctly completed case. Repeat the evaluation when the schema changes.
Jev is an interesting candidate for a narrow decision layer. Open-ended writing, explanation and research still need other tools. The biggest gain may come from removing thousands of routine, limited-choice decisions from the large model's workload.
Sources: TypeSafe launch, September 15, 2026, TypeSafe early access, Jirka Herník: AI mastermind 40, September 17.
Added on September 22, 2026 and dated to the original announcement. The integration described here is a proposed pilot, not my own Jev benchmark.