awesome-typesafe-jev
Awesome Jev: a source-backed field guide to TypeSafe's System One model, with SDKs, live demos, agent tools, and independent evaluations.
https://github.com/AbdelStark/awesome-typesafe-jev
Last synced: 29 days ago
JSON representation
Community projects
Agent and developer tooling
- Augustus - train` agent skills for placing and building application-specific decision models: primitive, base-model, and method selection, data assembly, fitting, export/reload, bounded improvement, and offline evaluation. TypeSafe Jev is the default hosted exemplar, not a model this project reproduces, and rules-only or no-training outcomes are valid results.
-
Beacon
- harness agent memory tool with an explicit
beacon memory evaluations runcommand: Jev judges bounded, redacted trace projections for reusable lessons, which a person reviews before adding to project memory or installing as a skill. Hooks and dry runs do not call Jev; normal capture can retain sensitive session text locally, so inspect retention before enabling it. - Bicameral
- Building with TypeSafe Jev - 09-25 snapshot of the docs and tells the agent the live docs win on any conflict.
- Canny - only ledger, flag a "done" claim without a passing check after the last code edit, and use optional Jev Noul judgments for advisory rule checks or to recognize a non-completion message; its default gate allows a repeated stop after warning, the author has not measured project-wide quality gains, and enabling Jev sends clipped diffs, project rules, and final messages to TypeSafe.
- Codex Jev Router - driven Codex subagent router that asks Jev Choice and Noul questions about a short task summary, applies confidence thresholds in code, and falls back to Sol when the decision fails; hosted Jev receives the summary, while local compatible backends are optional. Its installer changes local Codex configuration with a backup, but routing is not enforced at the tool boundary.
- DGP - based agent protocol with a Jev adapter, immutable evidence frames, typed assessments, and application-guarded commits; the local reference app simulates domain effects, and opt-in live mode sends decision evidence to TypeSafe.
- Distill - confidence decisions leave the normal path in place. Jev does not decide tool permissions, and routing can send the user's request and recent steps to the configured endpoint.
- Every
- evoke
- fast-jev-compaction - hook plugin and npm library that asks Jev which older tool calls and results to keep, truncate, or remove without rewriting user and assistant messages; it sends a bounded version of the conversation and tool inputs to TypeSafe, and its relevance probabilities do not guarantee safe deletion.
- Foreman
-
fx
- labs/fx/blob/main/src/builtins/gateway/typesafe_permission_reviewer.zig): setting
review_modeltotypesafeai/jevsends the composed policy, context, and pending action to TypeSafe directly or through Vercel AI Gateway, then maps Jev's Choice to a permission decision; recorded probabilities and confidence are not threshold gates. - Hermes Jev Skills - feature data disclosure before use.
- Hermes JIT Context OS - agent context runtime with an optional Jev Noul ranker for candidate facts: it sends a bounded query and fact excerpts to OpenRouter, caches scores, and labels deterministic token-overlap fallback when the remote path is unavailable. Its published ten-task live-call record uses small SWE-bench-style fixtures built from inline files, not full repository checkouts or a SWE-bench score.
- Hippo Memory - in Jev reranker that batches Noul judgments over the top 40 recalled memories and falls back to a local cross-encoder on errors. The author's published study found better ranking on two corpora but no demonstrated answer-quality gain over the cross-encoder; enabling Jev sends the query and candidate memories to TypeSafe.
- hush - more-info, and possible-duplicate in one call, each applied only above a threshold the maintainer sets, and nothing at all below it.
- is-malicious
- jeff
- Jev Clean
- Jev Codex Router
- Jev Cookbook
- Jev MCP - compatible agents.
- Jev MCP by jkudish
- Jev Review - review workflow and local dashboard that follows structured signals through focused Jev calls.
- jev-align (Sutro) - learning CLI that evaluates CSV, Parquet, and JSONL data with Jev, asks people to label uncertain and randomly audited examples, and uses GEPA to propose improved definitions while keeping labels and proposal acceptance under human control.
- Jev-assisted compaction - aware compaction in the kamchatka agent.
-
Jev-assisted shell
- -assisted-shell
and ran with--advise, thekamchatka` agent classifies shell commands, providing the user with a quick, color-coded safety rating for each command that a model wants to run. - jev-axi - ergonomic CLI following the AXI conventions that gives coding agents Jev judgments for blocking risky tool calls, screening fetched content for prompt injection, triaging build logs, flagging risky diffs, and filtering or ranking many items; its own benchmark found agents using it read fewer files but cost the same, so it is meant for judgments rather than as a substitute for reading code.
- jev-belay - question Jev call only when files changed with no passing check since, and failing open on every error path.
-
jev-cli
- g jevctl
) that turns Jev judgments into pipeable, exit-code-gated shell commands:verifyclaims against evidence,screentext for prompt injection before an agent reads it,classify,extract,match,route,findandrerankup to 250 candidates,compactagent transcripts by dropping stale tool calls verbatim, andbatchany of them over JSONL with a concurrency pool; thresholds and--fail-on` policy live in code, not prompts, it works over TypeSafe, OpenRouter, or Cloudflare Workers AI, and it ships as a Claude Code plugin with a compaction hook. - jev-commit - commit hook where one Jev call judges whether the commit message matches the staged diff, flags debug leftovers and unmentioned work, and blocks only when it detects a credential.
- jev-engineering
- jev-graphify - Labs/graphify) code-graph communities by having Jev pick from candidate names built from the graph, then routes a developer question to the community and symbol to read first; sends symbol names, file paths and short code comments (not source bodies) to TypeSafe, and has only been tested on one codebase.
- jev-lint - licensed fuzzy linter that uses Jev to flag team-rule violations as Claude Code and Codex edit, with configurable rules and local thresholds; added code goes to TypeSafe and API errors fail open.
- jev-logtriage
- jev-mobile - step choices over prevalidated UI actions, with confidence gates, pagination, escalation, and optional LLM planning; currently a proof of concept tested mainly against Android Settings.
- jev-pref - defined semantic preferences into Jev Noul or Choice checks over code changes, then map results to advisory or blocking outcomes in code; supports tuning on labelled diffs, and sends reviewed change text to TypeSafe.
- jev-pruner - in Codex wrapper that use Jev to trim long Bash output before the main model sees it, while preserving diagnostics and archiving the full output locally for recovery; scoring sends command output and session history to TypeSafe, and its retention rules are heuristics rather than guarantees.
- jev-router
- jev-skill-router - suggestion cookbook to a UserPromptSubmit hook over user, plugin, and project skills; it sends the prompt text and skill descriptions to TypeSafe, starts in a log-only shadow mode, and ships thresholds that are not yet calibrated on its own data.
- jev-use - in PreToolUse gate that can only deny or ask, and a typed escalation contract (writing, open_ended, oversized, unsure, unreachable) returns every other step to the LLM rather than guessing — an unreachable backend escalates instead of allowing, so a gate that cannot be judged never waves a command through. Interchangeable TypeSafe, OpenRouter, and Vercel AI Gateway backends, a routing skill, and a native pi extension; its own live benchmarks, including the runs where Jev did worse, are published in the repo.
- jev.nvim - language question with Jev, and ranks answers by probability in the quickfix window.
- jevcal - question confidence threshold to a target accuracy on your own labeled data, verifies it on a held-out split, estimates how much traffic still needs a fallback model, and re-checks the locked thresholds in CI; publishes no Jev results of its own, and thresholds fitted on fewer than about 100 labeled rows should not be trusted.
- JevDroid - run budgets; goals and visible UI text are sent to the selected TypeSafe or Vercel provider.
-
jevgrep (allebee)
- by-meaning CLI (
jevgrep-clion PyPI) for logs and other text, includingtail -f: it batches one Jev Noul per line and prints lines above a code-set threshold. Its hand-labelled, 195-line synthetic-log benchmark compares Jev with Claude; every judged line goes to TypeSafe or OpenRouter, and optional--explainsends selected lines to Anthropic through OpenRouter. - Jeview
- JevLoop (Python) - confidence decisions can trigger LLM arbitration, and a shared guarded kernel executes file and shell operations in Docker sandboxes; task and tool context are sent to the configured model providers.
-
Jevonian
- compatible APIs. The
jevonian/autoroute asks Jev to choose a model and thinking level after deterministic compatibility and quota filters; explicit routes skip Jev. The ledger records the serving model, route reason, token usage, and estimated cost. Jev receives recent messages and tool results, and the optionalfullPromptsetting can send the whole conversation. Low confidence is flagged rather than automatically rerouted; costs and cache savings are estimates. -
jevr
- style
path:lineresults; the query and selected source excerpts go to TypeSafe in requests containing up to 24 KB of source text. - JevRouter - only decision receipts.
-
jgrep
- g jevgrep
) that splits files or git diff hunks into 5–60 line chunks, packs several chunks into one request with a Noul question per chunk, and printsfile:linehits above a probability threshold with grep-style exit codes, so a diff can be linted in CI against rules written in English; ships an interactivejgrep init` and an opt-in Claude Code and Codex skill; chunks are judged in isolation so cross-file questions do not match, and every chunk's text is sent to TypeSafe's API. - langchain-skill-router - turn skill routing for LangChain deepagents: a Jev adapter or another configured judge ranks and verifies candidates before middleware loads one skill or offers a shortlist; the request, recent conversation, and skill text go to that judge, and the published evaluation uses one generated testbed and one agent model.
- Mobile Jev - theme demo checks the resulting switch; it needs a device plus Mobilerun and TypeSafe keys, while CI tests do not control a live phone.
- Oko - function excerpts; sends the question and up to 90 code chunks per search to TypeSafe, and a keyword-only mode works without a key.
- oxlint-plugin-jev
- patdown - swappable Jev backend and configurable probability threshold; it sends matched file content to TypeSafe, has no request budget or result cache yet, and its semantic verdicts need human review for consequential gates.
- perch - -since` scopes CI runs. It sends method source to TypeSafe and reads every method in scope, so review request cost and findings before using it as a gate.
- pi-heed - first, tests-before-push) and checks side-effecting tool calls against it before they run; rules handle side effects and paths while Jev only classifies how each message changes the policy, re-checks exceptions and skips tools a free-text rule cannot concern; ships a replayable benchmark and an experiment log on Jev calibration and question design; experimental, shadow mode by default, fails open, and the benchmark is scripted rather than drawn from real sessions.
-
pi-jev
- mode tool-call gate, output judge, and a general typed
jev_asktool. - pi-jev-compaction - output excerpts that may contain private data, and tests do not establish live relevance quality or cost savings.
-
pi-jev-context
- ranked blocks survive, kept lines stay verbatim, and a
context_recalltool returns the original; ships experiment reports and a findings log with pre-registered synthetic sets and weakly labelled replays of real sessions, including a negative result (Jev-judged pruning of old context dropped information needed later, so that part stays shadow-only); experimental, shadow mode by default, and the real-session replays come from one user's sessions. -
pi-verdict
- path prompts) for clear decisions, then routes gray-zone cases to a fail-closed enforcing classifier (configurable via
classifierModelto point at Jev: allow/ask/deny); the Jev backend is experimental — served through OpenRouter or TypeSafe's direct API, it ignores protected-path hints and can be swayed by adversarial transcript content; transcripts are sent to whichever classifier backend is configured. - pi-warden - typesafe that return Jev's verdict to the agent as a held tool result or a short steer instead of a dialog, check writes against a project rules file, and grade their own holds against the user's next message on recorded sessions; the action guard is calibrated on one user's 17k calls, the other guards on synthetic cases only.
- pytest-jev
- Second Thought - ticket example run.
- Skillbox - hosted agent skill library with opt-in Jev recommendations over task text and authorized active skill descriptions, using an owner-provided TypeSafe, OpenRouter, or Vercel AI Gateway key. Failed, oversized, or rate-limited evaluations fall back to deterministic search rather than returning a partial model ranking.
- SkillRanker - in network disclosure; fresh ranking sends redacted session context and skill excerpts to TypeSafe, and its license includes an OpenAI/Anthropic rider rather than plain MIT.
- slop-grader - fix violations.
- Sniff Test - in Jev Noul judgments over individual paragraphs, available as a CLI, pre-commit hook, and GitHub Action; its author-published comparison uses a small seeded corpus, with some judgment rules tuned on those same seeds.
- Supercov
- Switchboard - ups, tool calls, and resume to avoid unnecessary prompt-cache disruption. It preserves native authentication and sends task text to the configured TypeSafe, Vercel, or OpenRouter Jev endpoint; raw prompt history is stored only when explicitly enabled.
- toolgate - call firewall for Claude Code (PreToolUse hook) and MCP servers (stdio proxy): static rules first, then seven Jev Noul risk questions per action mapped to allow, ask, or deny by policy thresholds, with an action ledger, key-aware redaction, and a local JSONL audit log. Sends redacted tool input, cwd, and recent prompts to TypeSafe, OpenRouter, or Vercel AI Gateway (owner's key); in one author's 1,369-decision window on 0.9.1, blind second labeling of 150 sampled allows found no permissive misses, while 21% of decisions required an ask, mostly from one axis. Version 0.11.0 addresses that axis but has not been re-measured. Heuristic redaction and model judgments are defense in depth, not a sandbox.
- Vercel Eve
-
VexJoy Agent
- harness coding-agent toolkit whose
/dpath uses up to two Jev evaluations to select agents, skills, and a pipeline, then checks the proposed intent before dispatch; deterministic force-route rules bypass Jev. Request text goes to TypeSafe or Vercel AI Gateway, and the installer changes local agent configuration and hooks. - wakegate - running agents on Workers, Durable Objects, and Node: before a sleeping agent's LLM is resumed on a timer or incoming event, Jev answers one Choice (wake, not yet, unrelated) against the agent's own sleep note, and code skips the wakeup only below 0.2 on wake while always waking on user messages, bare timers, a skip limit, errors, and timeouts; its eval is 21 hand-written scenarios, not a benchmark.
Applications and workflows
- BTK audit studies - distance triage: 1,204 pages judged per run, 4,816 typed judgments in under 3 minutes, $0.0048 per 12-query batch (jev-1.13.0).
- discoprint
- DocJev - document, eight-packet comparison includes raw results but no human label review; normalized page text goes to the decision provider.
-
Formanator
- -yolo
and the MCPcreate_claim` tool can submit without that prompt. -
GeekLink Jev Subtitle Translator
- source SRT translator and quality-control tool that sends each source–translation pair to Jev as a
Nouldecision, then flags suspected omissions or meaning changes for human review without automatically rewriting subtitles. - HA-Jev
- Jev Chat Assistant - reported, Lark capture is partial, and chat content goes to the configured model APIs.
- Jev Search
- Jev Trade - run path; a configured live key sends real testnet or mainnet orders. Live demo.
- Jev Trader - USDC bot that packages order-book and trade data for an opt-in Jev buy/sell Choice while code places post-only quotes; the default model and public demo use a mock dry-run, and configuring a private key can place real orders and consume gas. The stated block-time latency is measured with the mock model, not Jev.
- Jev Wrapped - ray of a public Telegram channel: code reads up to 1,500 posts of the last twelve months from Telegram's public web preview, sampled evenly across the months when there are more, Jev answers a Choice over ten kinds of post and three Noul questions (paid ad, clickbait, emotional pressure) about each, and code applies fixed thresholds and draws the shares month by month on a shareable card that links the highest-scoring posts for a manual check; only public post texts are sent to Jev, the shares are model judgments that can misread partner promotions as ads, no sign-in, open source (MIT).
- jev-fit - confidence "not sure" state; closed source, and the pasted text is sent to TypeSafe's API.
- jev-research-pipeline - centric Obsidian note only for what code keeps; source text goes to TypeSafe and kept text to DashScope, thresholds are still the cookbook starting values (pilot: 5 of 7 goal conditions met), MIT.
- jevmeter - sentence authored evaluation does not establish accuracy on real video.
- JevNoiseGate - filter for verification codes that never reaches the API and a fail-open default on every uncertain path; experimental and developed against a single device, message bodies are sent to the API except when that local gate matches, and credentials are stored unencrypted in app-private storage.
- Jevometry - like output distributions, parameter sensitivity, and Fisher geometry, with an optional TypeSafe SDK adapter that sends experiment states and questions to the provider. System-information and Cramér–Rao results require explicit statistical assumptions; its first live smoke test captured responses but did not yield stable Fisher estimates.
- JevPDF - owned threshold light up ranked by probability; only extracted text is sent to TypeSafe, never the PDF, and the API key stays behind a server-side proxy.
- JevSpan - shot named entity recognition for Chinese and English text: code splits the text at punctuation and enumerates candidate windows with exact character offsets, Jev answers Choice questions that nominate, verify, and fix the boundary of each entity, and every result keeps its probability and a decision trace (CLI, Python library, and local web UI; MIT). Its authors report 73.7 average strict F1 on 200-sentence samples of 12 public NER benchmarks with jev-1.13.0, with aggregate results committed; input text is sent to TypeSafe, and an entity that no candidate window covers cannot be found.
- Jevtown - set price ladder. Personas are computed from their id rather than written by a model, and each reaction is sampled from the returned probabilities with a fixed seed, so the result is a simulation of a typed audience and not a forecast of real behaviour; open source (MIT), no sign-in, runs on a TypeSafe or OpenRouter key.
- Lossless Rewrite
- Paper Radar - type judgments, then applies code-owned thresholds to publish a page and RSS feed; a committed 50-paper run reports 5 seconds and $0.001957 with jev-1.13.0, while calibration requires the reader's own labels and the judgments do not inspect full papers.
- Paper Trellis Citation Verifier - reviewed manuscript citation checker: code verifies retrieved passages, Claude proposes evidence, and Jev scores whether the cited passage supports the claim; citation text and paper content go to the selected providers, and the model verdicts have no labelled biomedical validation set yet.
- QuantDinger - hosted quantitative trading platform that uses Jev System One as an optional, auditable PASS / REJECT gate for strategy and Quick Trade entry orders, with LLM fallback and deterministic bypasses for exits and protective orders; enabling the filter sends its prepared decision context to the configured AI provider.
- Ring Zero Security - 2.0 userspace and GPL-2.0 kernel runtime security for AI coding agents: deterministic Linux BPF LSM policy enforces protected-file boundaries while optional Jev checks can only raise a finding from scored tool calls or selected instruction-bearing files. Checks are off by default; enabling the hosted provider sends redacted tool arguments or file content to TypeSafe, and the repository does not measure whether model scoring improves detection.
- Slop Filter for LinkedIn - post labelled set, which is too small to characterise precision beyond that sample; English and text only, and the LinkedIn DOM selector will need updating when LinkedIn rebuilds its feed.
- Tax Document Classifier - licensed, text-only classifier for 261 federal tax forms that sends PDF text to Jev; the author reports no wrong labels on two test corpora but 38 low-confidence pages, without committed page-level results.
- TypeSafe Typewriter
Browser agents
- Cline Jev Browser
-
fastbrowse
- alpha browser agent where Jev picks actions from observed controls, an LLM plans and reads, and code requires page quotes for answer claims; a classifier stops before consequential actions unless the operator passes
--authorize, but that gate is not a guarantee. Browser Use Cloud is the default browser, with local Chrome optional. Page observations go to the selected Jev provider and planner model; its author publishes versioned per-run comparisons, not an independent benchmark. - Jev Browser - licensed Playwright navigator with MCP, CLI, and library interfaces: Jev chooses DOM actions and judges goal/stuck state from page excerpts sent to TypeSafe, OpenRouter, Cloudflare, or Vercel; an optional separate model generates typed text. Code bounds the loop and returns a trace, final page, and screenshot, but does not require human approval for each action; iframe, shadow-DOM, and file-input controls are unsupported.
- Jev Browser Use - allowed clicks and scrolls while Codex handles typing and verifies outcomes; browser execution is validated only in Codex, page text leaves the browser, and the author's approximate speedup is not an independent benchmark.
- Jev for Chrome - action human approval; includes a 17-task headless-Chromium suite with recorded traces.
- Jev Social - gated choices over observed, bounded socai CLI operations. The goal and captured result excerpts go to OpenRouter; local code validates choices and stops at access gates. Browser capture is read-only except for an optional TikTok media download when explicitly requested; there is no per-step human approval.
- Jev Ultrafast
- jev-agent-browser - browser performs them, and ambiguous or blocked flows escalate back to the parent.
- jev-skip - segment sponsor probability on the seek bar before the intro ends, with no crowd database; reports catching 77% of SponsorBlock's sponsor seconds across 23 videos at $0.0008 a video.
- PlotVeil - rejected check leaves the comment covered rather than revealed. Requests go through the author's Cloudflare Worker, which forwards comment text, video title and channel but not the anonymous install ID; the committed evaluation is a 10-sample hand-written regression set across English, Chinese, Japanese and prompt injection, not a production accuracy measurement. Install PlotVeil.
- unclutter - element snippets, then stores reusable local hiding rules by page template. Paid analysis is manual by default; optional on-visit analysis sends snippets to the selected provider. The API key stays in unencrypted local extension storage.
- voice-browser - controlled Playwright browser: Jev chooses a typed intent and target from speech and a bounded page snapshot, while code applies confidence gates and asks for confirmation before actions it classifies as destructive. The Web Speech API sends audio to Google, and transcripts, page context, and recent actions go to TypeSafe; the author's 34-case integration result uses captured fixtures, not arbitrary websites.
Client libraries and integrations
- Advocaat
- AnyDecisionModel - backed choices, and ordered scores over either TypeSafe's Jev API or a local MLX language model on Apple silicon; the local backend reads answer-token probabilities without generating text, and its calibration is caller-configured rather than established for every task.
- Ax TypeSafe integration - form text and numeric signature fields are not native Jev outputs, so use its explicit native interface for scoring rather than treating a bounded number as a Score rubric.
- DuckDB Jev - query budgets, caching, and request telemetry; published live throughput uses a repeated synthetic ticket corpus and does not measure classification accuracy, and binaries must match the DuckDB version and platform.
- Hunch - ts) offers the same interface; not affiliated with TypeSafe AI.
- Hunch for Python - jev`) that maps Jev Choice, Score, and Noul to functions over strings, lists, and DataFrames, deduplicates per-row requests, and keeps thresholds in caller code; selected rows and context go to TypeSafe or a configured gateway, and optional LLM escalation sends uncertain rows to that provider. A TypeScript port offers the same interface.
- Jev for Apple Foundation Models - transport tests and two demos, but requires iOS, macOS, or visionOS 27 and an API key kept on a backend or trusted machine rather than in a mobile app.
- jev-acp
-
Jev-Switch
- shaped
/v1/systemoneendpoint and routes each request through an editable graph of upstream routes with failover; v0.1.0 adapters cover only Vercel AI Gateway and Laya, not TypeSafe's own API, request state goes to the selected upstream, and the README is in Chinese. - jev.zig - time typed answers, retries with per-attempt timeouts, and diagnostics. Questions are fixed at compile time.
-
Jevelry
- stage TypeScript CLI, library, and coding-agent skill for reusable Jev question templates: it maps typed answers to act, mark, or fall back, and keeps a local decision/outcome log for review. Question state goes to TypeSafe; the log keeps a state hash by default but can opt into full state, and configured
runcommands may execute locally without per-action confirmation when confidence is high. - jevframe - question evaluation to pandas and eager Polars, with full Choice and Score probability distributions, bounded row concurrency, and opt-in memory caching; each uncached row sends selected context to TypeSafe, and LazyFrame expressions are outside v0.
-
jevql
- shaped CLI and Go/TypeScript/Python SDKs that let you write
WHERE jev(alias, 'condition'),jev_prob,jev_choice, andjev_scoreagainst a vanilla Postgres with no extension: it runs the plain SQL on the server, judges the surviving rows with Jev in batches (cached in local SQLite), and applies filter, sort, and group in the client; every row that survives the SQL filters is sent to TypeSafe and judged, so put cheap predicates in SQL first. - jevsearch - orders or drops them; with no embeddings or vector database it only re-ranks what the keyword pass retrieves, and keyword order stands if TypeSafe is slow or down.
- json-render - supplied candidates through Vercel AI Gateway. The Jev API is unreleased and requires a source build; application code owns components, actions, and side effects.
- Laya for Node.js - compatible Laya Choice, Score, and Noul model locally through ONNX Runtime, batching a question set in one call; first use downloads about 1.7 GB of weights, and its reference-output test runs only when that model bundle is present.
- LlamaIndex Jev - engine selector on the official Python SDK: Jev scores retrieved passages and chooses which tool handles a query; score mode is a 0–3 rubric, not cosine similarity.
- Milvus Model - specific, requires a TypeSafe API key, and is confirmed in source but not yet in a verified package release.
-
mysql-ailike
- language row filters in
WHEREand a three-argumentailike(left, right, prompt)for semanticJOIN ... ONconditions, powered by Jev; it is a preview release, each uncached evaluation is an API request, and evaluated values and prompts are sent to TypeSafe. -
NeuroLink
- weights Laya, declared per provider through an
inferenceKindsfield rather than inferred from behaviour, and consumes it internally for model routing, context compaction, MCP tool selection, and RAG planning; every internal caller uses a fail-open wrapper that returnsnullon any failure, so with no key configured the library behaves exactly as it did before, and it is not affiliated with TypeSafe AI or Convai Innovations. - OCaml SDK - based client.
-
pg-jev
- powered
WHEREpredicates, probabilities, Choice, and Score over table rows; it batches rows, caches answers per session, and offers optional query spend caps. Every judged row goes to TypeSafe, and installation needsplpython3uand superuser access, which many managed Postgres hosts do not provide. -
pi-typesafe
- managed TypeSafe client with a batched
typesafe_evaluatetool and offline-testable transport; requests are billable and opt-in per user. - RubyLLM TypeSafe
- s1-rs - free testing.
-
scala-jev-sdk
- effect, or ZIO, retries that honour
Retry-After, and local validation that rejects a malformed question set before it costs a round trip; every call returns anEitherrather than throwing, effects other thanFutureand the blockingIdentitymust supply a one-line sleeper for the retry timer, Scala 2.13 is not supported, and it is not affiliated with TypeSafe AI. - SemanticPolicy - 2.0) that writes application and agent decisions as Boolean, Choice or Score rules Jev answers, turns the probability into a verdict with thresholds measured on labelled examples by an evals CLI run from a clone, and adds checks before the model call, before a tool runs and after it returns in Microsoft Agent Framework agents; the evaluated text goes to Jev through OpenRouter or TypeSafe's endpoint and the library does not log it, and a rule is a signal, not a security boundary.
- Swift SDK
- TypeSafe AI for Rust
- TypeSafe AI Swift SDK - free Swift 6 client for Choice, Score, and Noul questions with strict concurrency, configurable retries, and network-free transport tests; production Apple apps should proxy requests through a backend.
-
TypeSafe SDK for Go
- over-environment configuration, retries that honor
Retry-After, anerrors.Is-matchable error tree, andlog/sloglogging that redacts credential headers but logs request bodies at debug level; answer types the client does not model are dropped with a warning rather than failing, and the module has no tagged release yet, sogo getresolves a pseudo-version. - TypeSafe SDK for Java - style builders for nested criteria, retries matching the official SDKs, status-specific exceptions, and a Spring Boot starter; depends only on Jackson and is not affiliated with TypeSafe AI.
- TypeSafe SDK for Kotlin - specific exceptions, HTTP and SOCKS5 proxy support, and runtime checks that each answer matches the question that produced it; targets Android and the JVM only, is distributed through JitPack rather than Maven Central, and is not affiliated with TypeSafe AI.
- TypeSafe SDK for PHP - call switch between TypeSafe and OpenRouter's decisions endpoint, retries and per-call overrides, any PSR-18 transport, and a Laravel service provider; calls are synchronous, model listing works only on TypeSafe, and it is not affiliated with TypeSafe AI.
- typesafe-ai-rails - sdk, with Rails configuration, persisted usage and cost telemetry, and opt-in confidence policies for Choice and Score answers.
- typesafe-rs - focused Rust transport SDK designed around behavioral parity with the official clients.
- typesafe-sdk - safe pooled HTTP connections; requires Ruby 3.1 or newer and has no async client.
- typesafe_sdk
- TypeSafeAI.Net
- Vercel AI Gateway - party hosted gateway entry for calling Jev through Vercel's AI SDK and gateway.
-
Vercel AI SDK for Python
- beta Python SDK includes an experimental
evaluateoperation that asks typed Choice, Score, and Boolean questions through AI Gateway usingtypesafe-ai/jev; the evaluation API is still experimental and requires Gateway access. -
vgi-typesafe
- row token usage, plus an
is_true()scalar forWHEREclauses; several questions share one request per row and repeated values are asked once per batch, but every other non-null row is a billable request that sends its content to TypeSafe's API.
Community and updates
- Client libraries and integrations - typesafe-jev/categories/agent-and-developer-tooling/) · Browser agents · Applications and workflows · Games and robotics · Evaluations and independent research · Showcases and field notes.
Evaluations and independent research
- Awesome Jev Robustness
- blink - hosted typed-decision models with Apache-2.0 code that return Choice, Noul, and Score option probabilities in one forward pass per batch through a TypeSafe-compatible API; the weights are non-commercial research and evaluation only, and there is no hosted TypeSafe-compatible endpoint.
-
Bosun v3.1
- 2.0 open-weight Qwen3 decision model family with released 0.6B and 1.7B weights and shipped
choice,score, andnoulinference; a separate Jev-style alternative, not hosted Jev. - Convex Decision Evals - run leaderboard comparing Jev with structured-output LLMs on 108 four-option questions about the Convex platform; its open runner sends questions and shuffled options to OpenRouter, uses native Jev Choice, and publishes per-answer evidence. This tests one platform's API knowledge, not whether a model can write working code.
- DecisionBench - grounded decision models, with pinned tasks and a native OpenRouter adapter for Jev.
- Janus - decide`) that measures a threshold on your own data and ships none by default; the protocol was frozen before any result and the raw JSONL and figures are committed, and no routing parameter transferred between the two datasets, as the optimal threshold, the sign of the accuracy gap between the two models, and whether routing paid for itself all changed; the Web of Science labels come from publication metadata rather than per-document annotation, so part of the error measured there is label ambiguity.
- jeff - hosted GLiFormer 400M server for Choice, Score, and Noul through a Jev-compatible API, with public benchmark code and per-item results. Its documented JevBench comparison finds weaker accuracy than Jev on reasoning-heavy items; hosted cost figures are estimates and depend on deployment throughput.
- Jev Capability Atlas - response API suites from cited third-party results and TypeSafe claims; its case studies help frame task-fit questions, but they do not establish population-level accuracy or a reusable confidence threshold.
- Jev Enterprise Decision Fabric - case Jev-versus-Claude agent-action evaluation, public labels and raw JSONL, report-rebuild tooling, and a decision inspector; one annotator revised labels after reviewing a Jev pilot.
- Jev IDS - licensed intrusion-detection prototype that asks Jev one Noul and one five-way Choice per network flow, comparing it with GPT-5.6 Luna and a Random Forest on NSL-KDD; it publishes run records and evidence limits, but the reported results are a 300-flow pilot and the gateway masks Jev's exact version.
- Jev in Search: Three Practical Evaluations - memory reranking, and multi-hop relationship selection in DeepSearcher, MemSearch, and Vector Graph RAG, linking implementation and evaluation artifacts; some datasets are private, some comparisons use different sample counts, and the animated speed illustration is simulated.
- Jev in the Wild - driven survey of 2,170 public GitHub Jev projects, mapping early ecosystem growth, application domains, and decision-use patterns; it measures repository activity rather than production use.
- Jev Judge vs Dimension Scores - scored dimensions with locally fitted weights, 5,477 test rows and 34.1M input tokens for $1.43; decomposition reached 0.9076 against 0.8373 on Japanese NLI but flagged about 25× more hard benign rows as attacks, and four repair attempts failed, on dimensions the author wrote himself.
- Jev Rerank Bench - level results, uncertainty intervals, and documented limitations.
- Jev Spam Eval - shot spam study against trained TF-IDF baselines, including results and explicit post-hoc-tuning caveats.
- jev-calibration-audit - 1.13.0 audit with reproducible code and per-call JSONL: tests abstention options, matched Korean and English items, question-shape interference, and option order; its strong abstention finding is on the KoBBQ dataset, not a universal calibration guarantee.
- jev-certify - powered audits over Jev 1.13 through OpenRouter, with request plans, 2,412 journalled answers and usage records, analysis code, and tests; its per-incoming-query risk bound depends on matching calibration and deployment traffic, and the study shows the scope gate missing its target after out-of-scope prevalence shifts while fallback accuracy remains unmeasured.
- jev-does-not-play-dice - way spinners) and on synthetic forecast documents, with raw responses and offline recomputation; Choice put 82.9% on face 1 across all 400 rolls at 19.0% accuracy and turned a stated 30% risk into 5.3%, while Noul stayed near the truth for 2 to 4 options and reported 15 to 17% for 8 to 20 options; no comparison with other models and no realistic task set, so it shows how far the probabilities can drift when the input carries no evidence rather than how Jev ranks against alternatives, run in launch week against jev-1.13.0.
- jev-fanout-bench - question-per-call Jev requests through OpenRouter's TypeSafe-compatible endpoint, with synthetic English tickets sent to the provider, raw responses and charge records, and repeated requests to measure answer stability; the tickets are unlabelled, and separate-call latency was measured serially.
- jev-measured - to-head on 27 author-written support tickets; the author publishes raw data and corrections to earlier comparison errors.
- jev-ood-calibration - generated support tickets; probabilities were near calibrated on the public sets but overconfident on an unseen priority rule, while the Boolean question showed a different error direction. The synthetic task is one family and the Gateway did not expose a fixed model version.
-
jev-orderby-bench
- registered gate on pairwise inversion, Score ordinality against a graded target, calibration, and negation and paraphrase invariants; jev-1.13.0 passes on 20 Newsgroups topic membership and fails four of six conditions on Amazon ESCI human-graded product relevance, the DuckDB integrations' request shapes are shown to change the numbers (a 40-row batched state fails the ranking gate that one row per request passes), and two-decimal output leaves 53 of 360 rows tied at the top so
LIMIT kcuts inside a tie; one seed and 30 ESCI queries, aggregates committed, corpus and cached responses regenerated locally for about a cent. - jev-sec-bench - 1.13.0 security evaluation with Go runner, raw per-sample results, and a TUI: 662 public prompt-injection messages and 200 matched vulnerable-code pairs; its reported classification scores use the study's stated context and a fixed 0.5 threshold, so deployment policy and corpus labels matter.
- Jevals.com - decision logs](https://github.com/Jevals/jevals-data) and the methodology are public, while the run harness is not published, LLM probabilities come from a prompt adapter, and the Gateway does not expose Jev's exact version.
- JevBench - model benchmark for typed decisions with 534 frozen cases per complete entrant, source adapters, scoring code, public per-task outcomes, and aggregate result artifacts. Its four-axis score includes accuracy, calibration, latency, and cost; some self-hosted latency is adjusted by assumption, hosting costs can be estimates, and withheld cases are exposed to the services being measured.
- jevmlx - licensed local decision layer for Apple Silicon that scores constrained Boolean, enum, and multi-select fields from MLX model logits in one prefill, returns schema-valid JSON, and offers a System One-compatible endpoint and benchmark runner; probabilities are over the supplied options and need calibration against labeled task data.
-
jevos
- licensed 1B yes/no model released as a q4_k_m GGUF that runs on CPU with llama.cpp behind a Jev-compatible
/v1/systemoneendpoint; it answers only Noul questions, and Choice and Score requests return 422. The authors report 0.815 accuracy against 0.927 for hosted Jev on 2,000 questions from unseen policies, but the training code and evaluation set are not published, so that result cannot be reproduced from the repository. - Kev - licensed, locally runnable Jev-style Choice, Score, and Noul models at 0.8B, 4B, and 9B, with released weights, training code, a System One-compatible server, frozen evaluation suites, and a playground. Its author reports a 0.822 new-source development accuracy for Kev-9B against 0.857 for hosted Jev on that suite, but Jev's training data is unknown, so this is not a controlled architecture comparison; test calibration and option order on your own data before setting a decision threshold.
- Laya - to-head, and raw calibration and some languages remain weak.
- LeJudge - licensed PushT study and cost module that turns LeWorldModel's imagined latents into closed-vocabulary facts, asks Jev typed questions about natural-language constraints, and adds the probabilities to a CEM planner's cost; ships oracle and keyword baselines, cached responses, and an offline paper rebuild; live runs send constraint text and probe facts to TypeSafe, and results cover one world model and twelve constraints, with near-miss false positives and rollout drift as material limitations.
-
LLM Memory Audit: Jev earnings test
- registered test of whether
jev-1.13-20260917remembers how 12,533 US earnings announcements turned out, comparing each company only with itself; it found no detectable memory (within-company AUC 0.506 on beats, against 0.587 pooled for Claude Sonnet 5 as a positive control) and publishes code, per-call probabilities and a freeze proof, but withholds the licensed FMP earnings and price data, and the verdict stays provisional until 757 prospective announcements are scored in January 2027. - LLM2Jev - licensed local toolkit that reads causal-model logits for Choice, Score, and Noul answers through Transformers or SGLang, with a System One-shaped HTTP endpoint, web and Snake demos, and shared-prefix cache measurements. Its published performance benchmark compares cache modes on one model and GPU, not decision quality against Jev.
- Love-Language Arena - reversed and negated re-asks to expose position bias, used to test three open models as fully crossed judges of Chinese and English rewrites, with self-preference correction, per-judge Platt calibration, dev/holdout prompt selection, and raw per-call JSONL. It does not call Jev, and its validation labels are Claude-generated rather than human, so the judge pass rates and model ranking show that the pipeline runs, not that the judges are valid.
- Luce - style decision models: a task description, an LLM teacher that writes the data, then LoRA plus a decision head on Qwen3-4B-Base returning calibrated choice, score, and boolean probabilities on a 12 GB GPU; the README reports accuracy and ECE against Jev on identical test items, including where training does not help, with a GPU-free replay demo.
- NanoJev - style model for one-pass action probabilities in Maze, Snake, and ViZDoom, with public checkpoint, training data, and replayable game comparisons; it is specialized to those games and is not a general Jev API replacement.
-
OneJev
- 2.0 open multimodal System One models in four sizes (0.8B to 27B, plus a 27B FP8 build) that answer
choice,score, andnoulabout screenshots, photos, video, and text in one forward pass through a/v1/systemone-compatible server; independent of hosted Jev. - Open Alternative to Jev - licensed Python library that reads option-token probabilities from open models through Transformers or vLLM, with packed and separate question modes, temperature scaling, benchmark scripts, and raw result files; its measurements compare modes of the same open model rather than Jev, packed answers can change with question order, and the package is not a drop-in System One API server.
- Open Medical Jev - licensed independent reproduction of the Jev decision loop with two frozen open-weight readers, a fit-free confidence gate, and split-conformal candidate sets; its four-reading configuration lands within 2 points of hosted Jev 1.13.0 on three 600-item medical licensing papers without training anything, running locally on one 24 GB GPU; only the China paper is a real exam (the US and India papers are fixed-seed draws), the exam items are not redistributed, and its coverage–accuracy curve still trails hosted Jev at the highest precision tiers.
- OpenJev (DiffusionGemma) - licensed decision server that reads Choice, Score, and Noul probabilities from DiffusionGemma 26B through a System One-shaped API, with NVIDIA and Apple silicon backends plus optional image input. Its NVIDIA path pins an unmerged vLLM branch, some limits differ from Jev, and answer quality needs evaluation on the reader's own tasks.
- poorjev - shot NLI models with temperature scaling and conformal abstention; ships a reproducible calibration eval (ECE 0.170 to 0.071 on its own small labelled set, cross-validated) and runs offline with no API key.
- Rizzo Flow - model decision server that reads answer-token probabilities instead of generating text, with a Jev-compatible Choice, Score, and Noul API plus its own numeric primitive. Its authors explicitly make no quality-parity claim with Jev; probabilities need calibration on the user's own data.
-
RSI-Jev
- like 2B open model trained by a recursively self-improving AI research system, with published version notes including failed experiments (MIT code, Apache-2.0 weights); it returns
choice,score, andnoulprobabilities through a/v1/systemone-compatible server. Its authors report 0.791 on typed-decisions after training on that benchmark's train split and 0.774 on 873 held-out scienthoon tickets; MMLU-Pro remains a weakness at 0.383 on a 1,000-question subset.