Jared Palmer’s Apache-2.0 Kev family exposes Jev-compatible typed decisions from 0.8B to 27B parameters, with inspectable training code, frozen suites, calibration reports, and unusually candid limits.
Korallis’s multi-agent development stack wraps Jev in explicit schemas, candidate limits, confidence bands, validation, fallbacks, and an exact-commit merge gate.
Ian Langworth’s open NetHack bot turns each game state into legal actions for Jev, exposing both the promise of typed choices and the amount of deterministic machinery needed around them.
Vercel’s Python team tried Jev on a useful live classifier and a deliberately awkward code generator; the failures make the model boundary clearer than another launch explainer.
A 15-task serving benchmark finds Jev strongest on content judgments, weak on predicting other models, and useful for length-aware vLLM scheduling—with important reproducibility limits.
Red Hat compared Jev, open decision models, small classifiers, and LLM judges on prompt injection and content safety; no single paradigm won on quality, latency, and portability.
A 575,442-call audit found Jev well calibrated on familiar closed-choice tasks but confidently wrong when answer-relevant information was absent or events fell beyond its observed knowledge boundary.
Luigi maps four Pi extensions that use Jev before and after tool calls, for lazy tool loading, compaction, and workflow gates—including a counterintuitive threshold failure.
WebSpeaker’s small bilingual pilot found Jev faster and slightly more accurate on isolated routing decisions, yet every end-to-end variant made the final chatbot answer slower.
Rishi Raj Jain’s demo uses broad candidate generation, one batched Jev classification call, and a Postgres function that applies role-based redaction—while leaving important production gaps visible.
Anthus tested GPT-6 Luna on 3,600 logic problems and found that its confidence stopped separating right from wrong as reasoning depth increased—but this is a preview of OpenAI’s Decisions API, not a test of the unreleased endpoint.
huemorgan2 turns Jev into a deliberately awkward chatbot—one decision per key—and uses the failures to show why option design and intermediate structure matter.
Amazon’s Strands Labs released a 1.9B-parameter Jev-style model with weights, training data, scripts, evaluations, failed experiments, and unusually explicit limits—not merely an API clone.
Omar’s Apache-2.0 Gutsy packages a Qwen3.5-0.8B fine-tune, calibrated option probabilities, explicit rejection, state caching, and a Jev-style local API—with unusually candid limits.
Silas Liu derives failure modes from Laya, Lev, and CLM’s architectures, tests them under controlled changes, then uses the same probes and API token counts to infer Jev’s likely design.
LangChain’s e-discovery example uses Jev for bounded page decisions, ordinary Python for routing, an LLM only for redaction, and human interruption for uncertain privilege calls.
Baki GUL’s open-source guardrail turns Jev scores into allow, review, or block decisions while preserving missing evidence—and publishes the false positives its development suite still exposes.
Vercel’s comparison is useful because it avoids a winner: Jev offers decision-native distributions, while Astra keeps the broader generation, vision, and tool-use job.
Chris Mungall’s Jevotron turns structured files into field-level review queues and backs the workflow with an unusually transparent, if small, agent-trace evaluation.
djev-run packages DiffusionGemma-Jev behind a TypeSafe-compatible decision endpoint, with Cloud Run GPU deployment, fixed answer slots, and public benchmark claims that still need independent reproduction.
Halv reports equal aggregate verifier passes at 57.1% lower recorded model cost, but its 42-pair checkpoint changes the whole workflow, uses 3.49× as many tokens, and is not an isolated test of Jev.
John Lindquist demonstrates Jev as a low-latency decision layer for voice commands, function routing, record matching, app navigation, search, games, agent coordination, and live coaching.
John Kutay reports testing Jev in Rippling's GrowthOS: a large reported win on scoring outreach, a loss inside a text-to-SQL loop that needs business context, and a plan to use Jev as the relevance function for ad-hoc retrieval.
Andrew Warner rounds up creator demos of Jev—email sorting, model routing, support triage, lead scoring, agent memory, clip scoring, smart-home control, and a failed Bitcoin signal—and frames it as a fast decision step inside larger workflows.
A last30days research brief on TypeSafe's decision model: where developers are trying it, what the speed claims mean, and why type safety is not correctness.
Laurie Voss makes the architectural case for cheap typed decisions, calibrated probabilities, and a hybrid Jev-plus-LLM evaluation loop—while separating launch claims from early tests.
A hands-on walkthrough compares Jev with Opus, Haiku, and GPT-5.4 Mini on bounded decisions; the speed is striking, but the labels and tiny sample do not establish general accuracy.
A browsable starting list of Jev projects: routing, guardrails, scoring, agents, data tools, and games. Leads to investigate, not tested endorsements.
A concrete Jev integration: batched typed judgments, a deterministic playbook, and a human making the final call.
Sam Witteveen builds a local model router that combines task classification, difficulty, privacy, and web-access decisions, then swaps cloud Jev for local Semif to keep the routing check on-device.
Nate B Jones defines Jev-shaped problems—complicated input, constrained decisions—and maps four ways a cheap general-purpose classifier can fit into everyday software.
Cole Medin tests Jev as a game controller, pull-request triage step, and model router, arguing that fast, general-purpose classification complements rather than replaces generative AI.
Sam Witteveen tests seven open Jev-style projects, comparing their classification techniques, local demos, and limits—and why harder decisions still need a reasoning model.
Jeremy Chone turns a raw Jev API wrapper into sysone 0.0.3, with typed questions, typed responses, mixed string/index keys, examples, tests, and an explicit early-release warning.
Ray Amjad explores Jev as a fast screening layer for coding agents: skill selection, browser verification, comment audits, qualitative linters, and code-review triage.
Sam Witteveen tests Jev's choice, score, and yes/no decisions, explores practical classification workflows, and separates schema reliability from correctness and undisclosed architecture.
Usutaku demonstrates Jev-based ad filtering, synthetic email routing and browser control on mostly self-made pages, with limits on what the demos establish.
Diogo Almeida argues that pleasing a human and running reliable software are different objectives. His case for moving beyond AI assistance—and why he says Jev's approach is not simply RLVR.