Post
CommunityJev Practices & Patterns
Laya in shadow, next to Jev on every request
Kitsuno's engineering write-up on training an open-weight Laya model to take over parts of its Jev usage: eight days of work, run in shadow beside Jev on every job-side request, without training on Jev's outputs.
- Format
- Post
- Reading time
- 3 min
- Published by
- Community
- Author
- Kitsuno
- Last verified
- Oct 2026
Highlights
- Since 23 September, Jev answered 35,812 logged calls with 76 million input tokens for USD 3.68 in total.
- On a same-day test set, untrained Laya answered 44% of questions right against Jev's 94%; staging had 26% country and 50% work-mode hits.
- Rule zero: no Jev output became a training label; the agreement bars distillation and the Laya worker cannot read Jev's answers.
- Training used about 4,400 job ads and 44,000 questions, with a frontier model as teacher and human review for unclear cases.
- After the first Jev gate, model calls per day fell from about 4,400 to about 2,000 while new jobs per day almost doubled.
Watch out
Grounded in the truncated markdown copy of the post published 2026-10-03; the capture ends before the switch criteria and conclusions, and Kitsuno's traffic numbers are its own mixed-traffic counts, which the piece itself says to read as a direction rather than a controlled test.
Questions about this entry
Short answers drawn from the linked project or page.
Kitsuno's engineering story of training an open-weight decision model, Laya, to take over parts of its Jev usage: teaching it Kitsuno's own questions and running it in shadow beside Jev on every job-side request in production.
The post says money is not the reason: at its volume Jev costs a few dollars a week. The drivers are Kitsuno's principles, open-source models for AI and data staying in European data centers, since Jev is closed and runs in the United States.
No, by the team's reading of TypeSafe's customer agreement, which bars using Jev output to distill or train a similar or competing model. The post calls this rule zero and says the Laya worker's database role cannot read Jev's answers.
On a small staging test it found the right country for 26% of job ads and the right work mode for 50%, where Jev had 89. On a test set built the same day, untrained Laya answered 44% of questions correctly against Jev's 94%.
A local 9-billion-parameter model first, but it was not good enough at 91% on the test questions, so the team switched to a frontier model as teacher with human review of unclear cases: about 4,400 job ads and 44,000 questions.
The post reports roughly 4,400 model calls a day before the first Jev gate and about 2,000 after, while new jobs per day almost doubled, with tokens falling from about 11 million to about 5 million a day, a direction it does not call a controlled test.
The post reports 35,812 logged Jev calls since 23 September with 76 million input tokens for USD 3.68 in total, and 30,738 calls in the last seven days, with most checks in the job pipeline and the Be Found product.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsMore Jev resources like this
ABTdomain's workflow design for brand-protection review: three layers of typed Jev questions turn each new domain that carries a brand word into Release, Watch or Flag, with 0.3 and 0.7 probability lines deciding what happens next.
Jev Practices & PatternsArticle#community#catalog-import#yibie
Community
Mapika's open decision model is a 1.9B Qwen3.5-2B fine-tune that returns a probability for every option of every typed question in one forward pass, with no generated text. v11 adds a rank-64 LoRA stage and reports 0.577 on the JevBench public hard tier, up from v10's 0.459.
Jev Models & Reimplementations#community#catalog-import#madewithjev
Community
Local unofficial /v1/systemone adapter that serves Choice, Score, and Noul probabilities via first-token LLM judgments or a fine-tuned ModernBERT scorer, including Japanese docs.
Jev Practices & Patterns#community#awesomejev#catalog-import
Community
Argos1111/jev_localPython
Argos1111/jev_localPythonCLI that calibrates Jev questions against your own labels: write questions, label examples in JSONL, run check, and get a verdict per question — gate, gate-above-confidence, ranker, unusable, or too-few-examples.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
smkrv/jev-calibrateTypeScript
smkrv/jev-calibrateTypeScriptUnofficial laptop study of parallel constrained decoding: prefill once, share the KV cache across fields, and decide all schema fields in one batched forward pass without writing JSON.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
rorshopping/jev-on-a-laptopPython
rorshopping/jev-on-a-laptopPythonLive race rig that labels 1,000 real app reviews with Jev typed questions against Gemini JSON output on identical batches, then replays recordings and re-ranks rows locally without extra model calls.
Jev Practices & Patterns#community#jevbest#catalog-import
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.