typed/recipes
Directory / Claims ledger

Claims ledger

Every claim about Jev and Laya we track, in one table: who made it, our verdict on the evidence, and whether we have reproduced it.

Updated 24 Sep 2026. 18 entries: 4 misleading, 1 hype risk, 2 unverified, 4 plausible, 6 verified, 1 toy. Reproduced by us so far: 0.

A verdict rates the public evidence, not the model. How verdicts work explains the five questions behind each label. When we reproduce a claim, its row changes to Tested and links the method and raw results.

Performance and capability claims

Vendor claims, community claims and benchmarks, most doubtful first.
ClaimMade byVerdictTested by usEvidence
Laya's 0.766 benchmark accuracy
Laya reaches 0.766 on the typed-decisions benchmark.
Convai Innovations
Vendor claim, 18 Sep 2026
MisleadingNot yetSource
Laya: “6–8× faster than Jev”
At 32.8 ms per question on one GPU, Laya is 6 to 8 times faster than Jev.
Convai Innovations
Vendor claim, 18 Sep 2026
MisleadingNot yetSource
“Jev can't hallucinate”
Because outputs are typed, Jev cannot hallucinate.
Launch marketing
Vendor claim, 15 Sep 2026
MisleadingNot yetSource
TypeSafe: 193.6× faster, 444.6× cheaper
Jev is 40–200× faster and 40–400× cheaper than frontier LLMs, peaking at 193.6× and 444.6× on TypeSafe's workflows.
TypeSafe AI
Vendor claim, 15 Sep 2026
MisleadingNot yetSource
One 50 ms pass vs. 23 LLM turns
A 706K-parameter model fills a whole form in one 50 ms pass while an LLM agent needs 23 turns and 39.6 seconds.
X post
Claim, 18 Sep 2026
UnverifiedNot yetSource
Jev as an agent safety monitor
Checking each agent action with Jev first catches most attacks with almost no false blocks, at much lower latency.
X test report
Claim, 18 Sep 2026
UnverifiedNot yetSource
Jev + fallback vs. GPT-5.6 Luna on passage support
Same accuracy as GPT-5.6 Luna and faster, but the fallback route cost more overall.
Community benchmark
Benchmark, 20 Sep 2026
PlausibleNot yetSource

Builds, tools and replicas

Projects and tools, rated on whether they do what their authors say.
ClaimMade byVerdictTested by usEvidence
jev-trader: a trade decision every block
Asks Jev for a buy or sell decision on every Monad block on a live market.
jarrodwatts
Project, 16 Sep 2026
Hype riskNot yetSource
Kev: Jev-like models on Qwen
Open Jev-like decision models built on Qwen, with a TypeSafe-compatible API.
Community
Open replica, 21 Sep 2026
PlausibleNot yetSource
Model router: Jev picks which model answers
Jev decides which LLM should serve each request before it's forwarded.
X demo (948 likes)
Demo, 17 Sep 2026
PlausibleNot yetSource
Jev Ultrafast for Browser Use
Jev picks both the browser operation and the page element in one request; a small LLM is only called when text needs typing.
browser-use
Project, 16 Sep 2026
PlausibleNot yetSource
llm-typesafe plugin for the LLM CLI
Adds Jev to the llm command-line tool and Python library.
Simon Willison
Tool, 22 Sep 2026
VerifiedNot yetSource
jev-mcp: an MCP server for Jev
Lets Claude Code, Claude Desktop, and Codex call Jev and branch on the probabilities.
jkudish
Tool, 17 Sep 2026
VerifiedNot yetSource
typesafe-ai/skills: official agent skill
An MIT-licensed skill that teaches Claude Code, Codex, and other agents to call Jev.
TypeSafe AI
Tool, 17 Sep 2026
VerifiedNot yetSource
jevlike: an open reimplementation
Trains a small model to score a changing list of options in one forward pass, with Doom, chess, and Wikispeedia demos.
vinnylarouge
Open replica, 16 Sep 2026
VerifiedNot yetSource
openjev (now SemIf): a Jev-style model on one RTX 3090
Runs a Jev-style decision model locally on a single consumer GPU.
TheoLeeCJ
Open replica, 16 Sep 2026
VerifiedNot yetSource
upweight: a personal Hacker News front page
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.
Vishesh Baghel
Project, 16 Sep 2026
VerifiedNot yetSource
jevchat, jev-leftpad, and jev-2048
A chat model that picks one symbol at a time, left-pad as a Choice question, and Jev playing 2048.
Kyle Pena, Fatih Kadir Akın, Andy Gayton
Demo, 20 Sep 2026
ToyNot yetSource

Use this data

The full ledger is available as JSON, one object per entry with the claim, source, verdict, test status and dates. Each entry page also carries ClaimReview or Review markup.

To cite it: Typed Recipes Claims Ledger, typedrecipes.com/claims, retrieved [date].

Know a claim we are missing? Submit it for review.

Continue learning