{
 "name": "Typed Recipes Claims Ledger",
 "url": "https://typedrecipes.com/claims",
 "updated": "2026-09-24",
 "tested_by_us": 0,
 "verdict_scale": {
  "Verified": 5,
  "Plausible": 4,
  "Unverified": 3,
  "Hype risk": 2,
  "Misleading": 1,
  "Toy": null
 },
 "entries": [
  {
   "id": "laya-accuracy",
   "url": "https://typedrecipes.com/directory/laya-accuracy",
   "title": "Laya's 0.766 benchmark accuracy",
   "type": "Vendor claim",
   "made_by": "Convai Innovations",
   "added": "2026-09-18",
   "models": [
    "Laya"
   ],
   "use_case": "Claim vetting",
   "claim": "Laya reaches 0.766 on the typed-decisions benchmark.",
   "our_read": "That score comes from a checkpoint fine-tuned on the benchmark's own training split. The zero-shot base scores 0.362. Convai's Jev comparisons also use Jev numbers from other people's runs.",
   "verdict": "Misleading",
   "tested_by_us": false,
   "source": "https://huggingface.co/convaiinnovations/laya"
  },
  {
   "id": "laya-faster",
   "url": "https://typedrecipes.com/directory/laya-faster",
   "title": "Laya: “6–8× faster than Jev”",
   "type": "Vendor claim",
   "made_by": "Convai Innovations",
   "added": "2026-09-18",
   "models": [
    "Laya",
    "Jev"
   ],
   "use_case": "Claim vetting",
   "claim": "At 32.8 ms per question on one GPU, Laya is 6 to 8 times faster than Jev.",
   "our_read": "Compares local inference on your GPU with a hosted API's network round trip. Both numbers can be true and the comparison still isn't like-for-like.",
   "verdict": "Misleading",
   "tested_by_us": false,
   "source": "https://laya.convaiinnovations.com/"
  },
  {
   "id": "cant-hallucinate",
   "url": "https://typedrecipes.com/directory/cant-hallucinate",
   "title": "“Jev can't hallucinate”",
   "type": "Vendor claim",
   "made_by": "Launch marketing",
   "added": "2026-09-15",
   "models": [
    "Jev"
   ],
   "use_case": "Claim vetting",
   "claim": "Because outputs are typed, Jev cannot hallucinate.",
   "our_read": "It can't return an answer outside the schema you define. It can still pick the wrong answer with high confidence. This was the main argument in the 1,800+ point Hacker News launch thread.",
   "verdict": "Misleading",
   "tested_by_us": false,
   "source": "https://news.ycombinator.com/"
  },
  {
   "id": "typesafe-speed",
   "url": "https://typedrecipes.com/directory/typesafe-speed",
   "title": "TypeSafe: 193.6× faster, 444.6× cheaper",
   "type": "Vendor claim",
   "made_by": "TypeSafe AI",
   "added": "2026-09-15",
   "models": [
    "Jev"
   ],
   "use_case": "Claim vetting",
   "claim": "Jev is 40–200× faster and 40–400× cheaper than frontier LLMs, peaking at 193.6× and 444.6× on TypeSafe's workflows.",
   "our_read": "The workflows were built by TypeSafe's own team, and TypeSafe says so and calls the gains likely high-end. The multipliers are honest about their origin but shouldn't be quoted as typical.",
   "verdict": "Misleading",
   "tested_by_us": false,
   "source": "https://en.wikipedia.org/wiki/Jev_(AI_model)"
  },
  {
   "id": "jev-trader",
   "url": "https://typedrecipes.com/directory/jev-trader",
   "title": "jev-trader: a trade decision every block",
   "type": "Project",
   "made_by": "jarrodwatts",
   "added": "2026-09-16",
   "models": [
    "Jev"
   ],
   "use_case": "Agents and real-time",
   "claim": "Asks Jev for a buy or sell decision on every Monad block on a live market.",
   "our_read": "The code is real. What's missing is any evidence it makes money, and a fast classifier is not a trading edge. Treat it as a latency demo, not a strategy.",
   "verdict": "Hype risk",
   "tested_by_us": false,
   "source": "https://github.com/jarrodwatts/jev-trader"
  },
  {
   "id": "form-fill",
   "url": "https://typedrecipes.com/directory/form-fill",
   "title": "One 50 ms pass vs. 23 LLM turns",
   "type": "Claim",
   "made_by": "X post",
   "added": "2026-09-18",
   "models": [],
   "use_case": "Agents and real-time",
   "claim": "A 706K-parameter model fills a whole form in one 50 ms pass while an LLM agent needs 23 turns and 39.6 seconds.",
   "our_read": "Precise numbers from a single post. The model is a small task-specific specialist, not Jev, so it's evidence for the category argument, not for any product.",
   "verdict": "Unverified",
   "tested_by_us": false,
   "source": "https://github.com/yibie/awesome-jev"
  },
  {
   "id": "safety-monitor",
   "url": "https://typedrecipes.com/directory/safety-monitor",
   "title": "Jev as an agent safety monitor",
   "type": "Claim",
   "made_by": "X test report",
   "added": "2026-09-18",
   "models": [
    "Jev"
   ],
   "use_case": "Guardrails and safety",
   "claim": "Checking each agent action with Jev first catches most attacks with almost no false blocks, at much lower latency.",
   "our_read": "“Most” and “almost no” are doing all the work. No attack set, no counts, no false-positive rate. The pattern is sound; the result isn't checkable.",
   "verdict": "Unverified",
   "tested_by_us": false,
   "source": "https://github.com/yibie/awesome-jev"
  },
  {
   "id": "kev",
   "url": "https://typedrecipes.com/directory/kev",
   "title": "Kev: Jev-like models on Qwen",
   "type": "Open replica",
   "made_by": "Community",
   "added": "2026-09-21",
   "models": [],
   "use_case": "Tooling and open replicas",
   "claim": "Open Jev-like decision models built on Qwen, with a TypeSafe-compatible API.",
   "our_read": "Exists and is runnable. We haven't compared its answers with Jev's on the same questions.",
   "verdict": "Plausible",
   "tested_by_us": false,
   "source": "https://simonwillison.net/2026/Sep/21/jev/"
  },
  {
   "id": "fallback-bench",
   "url": "https://typedrecipes.com/directory/fallback-bench",
   "title": "Jev + fallback vs. GPT-5.6 Luna on passage support",
   "type": "Benchmark",
   "made_by": "Community benchmark",
   "added": "2026-09-20",
   "models": [
    "Jev"
   ],
   "use_case": "Retrieval checks",
   "claim": "Same accuracy as GPT-5.6 Luna and faster, but the fallback route cost more overall.",
   "our_read": "Reporting a downside is a credibility signal. A useful reminder that a cascade only saves money when the fallback fires rarely.",
   "verdict": "Plausible",
   "tested_by_us": false,
   "source": "https://www.ayautomate.com/jev-builds"
  },
  {
   "id": "model-router",
   "url": "https://typedrecipes.com/directory/model-router",
   "title": "Model router: Jev picks which model answers",
   "type": "Demo",
   "made_by": "X demo (948 likes)",
   "added": "2026-09-17",
   "models": [
    "Jev"
   ],
   "use_case": "Routing and triage",
   "claim": "Jev decides which LLM should serve each request before it's forwarded.",
   "our_read": "The most common pattern in launch week, and LangChain shipped an integration for it. Savings depend on how often the cheap route is right, which the demo doesn't report.",
   "verdict": "Plausible",
   "tested_by_us": false,
   "source": "https://github.com/yibie/awesome-jev"
  },
  {
   "id": "browser-use-ultrafast",
   "url": "https://typedrecipes.com/directory/browser-use-ultrafast",
   "title": "Jev Ultrafast for Browser Use",
   "type": "Project",
   "made_by": "browser-use",
   "added": "2026-09-16",
   "models": [
    "Jev"
   ],
   "use_case": "Agents and real-time",
   "claim": "Jev picks both the browser operation and the page element in one request; a small LLM is only called when text needs typing.",
   "our_read": "Real repo from a well-known agent framework, demoed on a flight search. We haven't measured end-to-end task success against the standard Browser Use agent.",
   "verdict": "Plausible",
   "tested_by_us": false,
   "source": "https://github.com/browser-use/jev-ultrafast"
  },
  {
   "id": "llm-typesafe",
   "url": "https://typedrecipes.com/directory/llm-typesafe",
   "title": "llm-typesafe plugin for the LLM CLI",
   "type": "Tool",
   "made_by": "Simon Willison",
   "added": "2026-09-22",
   "models": [
    "Jev"
   ],
   "use_case": "Tooling and open replicas",
   "claim": "Adds Jev to the llm command-line tool and Python library.",
   "our_read": "The quickest way to try a Noul from a terminal.",
   "verdict": "Verified",
   "tested_by_us": false,
   "source": "https://simonwillison.net/2026/Sep/21/jev/"
  },
  {
   "id": "jev-mcp",
   "url": "https://typedrecipes.com/directory/jev-mcp",
   "title": "jev-mcp: an MCP server for Jev",
   "type": "Tool",
   "made_by": "jkudish",
   "added": "2026-09-17",
   "models": [
    "Jev"
   ],
   "use_case": "Tooling and open replicas",
   "claim": "Lets Claude Code, Claude Desktop, and Codex call Jev and branch on the probabilities.",
   "our_read": "Labeled a proof of concept by its author. Fine for experiments.",
   "verdict": "Verified",
   "tested_by_us": false,
   "source": "https://github.com/jkudish/jev-mcp"
  },
  {
   "id": "typesafe-skills",
   "url": "https://typedrecipes.com/directory/typesafe-skills",
   "title": "typesafe-ai/skills: official agent skill",
   "type": "Tool",
   "made_by": "TypeSafe AI",
   "added": "2026-09-17",
   "models": [
    "Jev"
   ],
   "use_case": "Tooling and open replicas",
   "claim": "An MIT-licensed skill that teaches Claude Code, Codex, and other agents to call Jev.",
   "our_read": "Official and installable. The fastest way to let a coding agent use typed decisions.",
   "verdict": "Verified",
   "tested_by_us": false,
   "source": "https://github.com/typesafe-ai/skills"
  },
  {
   "id": "jevlike",
   "url": "https://typedrecipes.com/directory/jevlike",
   "title": "jevlike: an open reimplementation",
   "type": "Open replica",
   "made_by": "vinnylarouge",
   "added": "2026-09-16",
   "models": [],
   "use_case": "Tooling and open replicas",
   "claim": "Trains a small model to score a changing list of options in one forward pass, with Doom, chess, and Wikispeedia demos.",
   "our_read": "A good read for understanding how the architecture works.",
   "verdict": "Verified",
   "tested_by_us": false,
   "source": "https://github.com/vinnylarouge/jevlike"
  },
  {
   "id": "openjev",
   "url": "https://typedrecipes.com/directory/openjev",
   "title": "openjev (now SemIf): a Jev-style model on one RTX 3090",
   "type": "Open replica",
   "made_by": "TheoLeeCJ",
   "added": "2026-09-16",
   "models": [],
   "use_case": "Tooling and open replicas",
   "claim": "Runs a Jev-style decision model locally on a single consumer GPU.",
   "our_read": "Real and runnable. It is not Jev, and its accuracy is its own.",
   "verdict": "Verified",
   "tested_by_us": false,
   "source": "https://github.com/TheoLeeCJ/SemIf"
  },
  {
   "id": "upweight",
   "url": "https://typedrecipes.com/directory/upweight",
   "title": "upweight: a personal Hacker News front page",
   "type": "Project",
   "made_by": "Vishesh Baghel",
   "added": "2026-09-16",
   "models": [
    "Jev"
   ],
   "use_case": "Scoring and ranking",
   "claim": "Six sliders (technical depth, drama, practical utility, AI slop, novelty, career relevance) re-rank the HN front page instantly, with each story's scores visible.",
   "our_read": "Live and working. Jev scores each story once; the re-sort happens in the browser, which is why sliders feel instant. A good model for any “rank a feed by my priorities” build.",
   "verdict": "Verified",
   "tested_by_us": false,
   "source": "https://upweight.vercel.app"
  },
  {
   "id": "toys",
   "url": "https://typedrecipes.com/directory/toys",
   "title": "jevchat, jev-leftpad, and jev-2048",
   "type": "Demo",
   "made_by": "Kyle Pena, Fatih Kadir Akın, Andy Gayton",
   "added": "2026-09-20",
   "models": [
    "Jev"
   ],
   "use_case": "Agents and real-time",
   "claim": "A chat model that picks one symbol at a time, left-pad as a Choice question, and Jev playing 2048.",
   "our_read": "Deliberately silly and a good way to see what the primitives can and can't do.",
   "verdict": "Toy",
   "tested_by_us": false,
   "source": "https://simonwillison.net/2026/Sep/21/jev/"
  }
 ]
}