Post

LLMs can write essays, generate code, and reason through tasks. But what if an AI system doesn’t need to write anything, just make a fast, structured decision? Meet Jev, TypeSafe AI’s “System One” model. 🧵👇

2/17 Jev’s core idea: send a piece of state (ticket, incident, document) plus typed questions. It returns machine-consumable answers and probability/confidence signals not a paragraph your app must parse. Think decision API, not chatbot.

3/17 Three primitives: • Choice = pick from options • Score = rate against a rubric • Noul = estimate whether a statement is true (0–1) Ask several questions in one request; Jev evaluates them in parallel. docs.typesafe.ai/introduction

4/17 The key difference: an LLM generates a flexible sequence of tokens—great for language, but apps often need to validate/parse the output. Jev constrains answers to a predeclared type/options, reducing glue code for bounded decisions.

5/17 Important caveat: “can’t hallucinate” is too broad. Jev can’t return an out-of-schema label, but it can still choose the wrong allowed answer or mis-score something. Confidence helps manage uncertainty; it doesn’t make a decision true.

6/17 Jev isn’t simply “an LLM with JSON mode.” TypeSafe describes a different model stack: a parallel sampler and RLCD (Reinforcement Learning for Calibrated Decisions). Its full architecture/training recipe isn’t public, so test the claims yourself.

7/17 Best mental model: LLM writes. Jev decides. Code acts. Let an LLM handle open-ended language and synthesis, Jev route/classify/score/check bounded choices, and deterministic code enforce permissions, thresholds, logging and fallbacks.

8/17 Example: an SRE copilot receives an alert. The LLM summarizes the incident and drafts a runbook step. Jev scores severity, selects the owning team and checks whether escalation criteria are met. Code gates paging; humans approve risky changes.

9/17 Where Jev may fit: high-volume, latency-sensitive decisions with a known answer space. Think ticket triage, queue routing, policy checks, content moderation, ranking, evaluations and agent/tool routing. Benchmark against your data; no model wins every task.

10/17 Limitations: Jev doesn’t write prose, code or explanations. It works best with narrow, well-scoped questions and known choices. For ambiguous problems or multi-factor reasoning, decompose the task or keep the reasoning step with an LLM.

11/17 Also: Jev is available via TypeSafe’s hosted API; its model weights aren’t public. Validate probability calibration on your own domain, log model versions, and use human review for uncertain or high-impact decisions. Treat vendor speed/cost claims as claims to verify.

12/17 Current status: Jev 1.14.0 was announced on Oct 14. TypeSafe says the direction includes small quality/stability improvements and better multilingual support. A detailed public delivery roadmap hasn’t been announced. typesafe.ai/blog/jev-1-14

14/17 Project #1: build a support-ticket router. Use Choice for team, Score for urgency/frustration and Noul for “Is this an outage?” Send low-confidence cases to a human. Runnable Python/JS quickstart: docs.typesafe.ai/introduction...

15/17 Project #2: build a cloud/SRE change gate. Jev checks whether a change includes rollback, monitoring and approval evidence; your code enforces policy. Never let a model score bypass IAM/RBAC, least privilege or a required human approval.

16/17 Try Kev for self-hosted experiments: it provides a Jev-compatible decision API, but it’s an independent model—not Jev’s weights/training. Compare accuracy, calibration, latency and cost on your own eval set: github.com/jaredpalmer/...