Research and benchmarks — page 13

Benchmarks Jev against other evaluation models in games with explicit states, legal actions, and measurable outcomes.

发明 RLHF 的人，这次做了个不会说话的模型：Jev 独立研究报告。52 页 PDF + 50 条中文实测复现包 + 143 条可回溯数据表

A small second eval for shadcn-ui/lint that uses TypeSafe's Jev to judge the linter's own output.

TypeSafe's Jev model plays Snake live in the browser, with a headless eval and a self-improvement loop

Application of TypeSafe Jev (noul judgment primitive) on the collusion.wiki corpus: agent vs human page authorship, head-to-head vs local Qwen3.8-Flash-Next

Ongoing Japanese research deck on Jev and System One models, maintained as Markdown slides.

TypeSafe ranking, classification, retrieval and structured-decision tools for Pi coding agents

Evaluating TypeSafe's Jev as a fast monitor and action gate for agent sabotage in SHADE-Arena, compared with Gemini 2.5 Flash/Pro.

Charts: TypeSafe Jev evaluated on Thai standardized exams vs 110 other models

Does Jev predict stock returns from news? It reads the news well; there is no tradeable alpha. Three arms separate reading from recall.

Jev (TypeSafe) vs. Gemini 3.8 Flash vs. GPT-5.6 Luna na anotação estruturada de sentenças do TJSP: qualidade, tempo e custo

Position paper: the Hidden-Markov and fuzzy primitives missing from TypeSafe AI's Jev and System-One decision models. Two lemmas, one principle (Deferred Crispification), one architecture (BSF-S1).

Can Jev pick the winner of a real headline A/B test? 64.5% across 10,984 Upworthy randomized experiments, 74.7% when the difference was decisive.

TypeScript experiments, evaluations, and latency benchmarks for TypeSafe's Jev model

Rust port of TypeSafe system-one-adapter (LLM-backed system_one evaluations)

Daily LLM benchmark on Poople (poople.io): one-shot word ladders to POOP, scored against provable optimal, accuracy vs cost via OpenRouter

"Newbee" isn't "LLM writes code for you"—it's an LLM inside a long-running, versioned, self-modifying Elixir environment. It uses Elixir as hands and brain; the environment evolves smarter and cheaper via real usage evidence—every change evaluable, revertible, attributable.

Benchmark for silent omissions in AI coding-agent patches: 310 labeled instances from 8 Python libraries, a deterministic AST detector, 4 LLM/probabilistic judges, pre-registered rules and paired cluster-bootstrap CIs. Includes the results that went against the thesis.

Evaluation System for DG Fundraising Cause

Open-source adaptive evaluation for LLM, RAG and agentic systems. System One judges first, LLMs only when necessary.

Claude thinks, jev clicks: a two-way loop between Claude Code and jev-ultrafast. 5-6x faster browser tasks, with before/after benchmarks.

Jev (TypeSafe AI) as a System One decision layer for a multilingual EN/UZ/RU RAG helpdesk, with an independent benchmark

Local-first hybrid RAG: traditional vs Jev-style (System One) pipelines over your own documents — with a built-in benchmark lab (6 scenarios, 48 questions, independent LLM judge). Hybrid won +8.4pp correctness at identical cost.

Model Context Protocol (MCP) server & OpenCode skill for San Francisco Early Learning For All (ELFA) preschool search, voucher calculations, and TypeSafe Jev System One evaluation