Jev for Trading Bots: Checking a TypeSafe Market-Making Demo
By Matt Cullerton, founder of Mavric and Apso
Jev is a decision model from TypeSafe AI that answers yes/no, multiple-choice and score questions about a block of text, with a published end-to-end latency of 70 to 500 milliseconds. A TikTok from @ray_fu shows it classifying crypto order flow and feeding a market-making quote engine, claiming Jev answers "under 100 milliseconds while normal LLMs take 3-10 seconds." The quote engine math is correct, but the video's own dashboard shows Jev taking 1.50 seconds, and its SDK code does not match TypeSafe's published SDK.
What does the demo show?
The video walks through a three-step build:
- A
feed.pypanel streaming BTC-USD one-minute candles, with tiles for buyer demand, order depth and price gaps. - A
market_snapshotJSON blob sent to Jev with three questions: a yes/no ("Is the incoming buy flow informed?"), a choice ("Which flow type is this?") and a score ("How strong is the directional pressure?"). Jev returns 87% yes, and "informed" at 81%. - A
quote_engine.pyloop that sets bid and ask prices using the Avellaneda-Stoikov model, with account limits such as position of at most ±4 lots and order notional of at most $50k.
How fast is Jev?
The voiceover says "under 100 milliseconds." The Jev answer panel in the same video reads "Real Jev output · 1.50 s," which is 15 times slower.
TypeSafe's launch post says "End-to-end response time is 70ms-500ms." Under 100 ms is the fast end of that range, not a typical result.
No source backs the "3-10 seconds" figure for other LLMs. TypeSafe itself claims Jev is "40x-200x faster" on its target tasks. InfoQ cites an OpenChamber analysis finding a median 7x speedup and 30x cost savings, and a Vercel engineer reporting a safety classifier running "five to 18 times faster" than the LLM it replaced.
Does the Jev SDK code in the video work?
No. The flow.ts snippet differs from the TypeSafe JavaScript SDK docs in four places:
| On screen | TypeSafe docs |
|---|---|
from "@typesafe/sdk" |
npm install @typesafe-ai/sdk |
new Jev({ apiKey: process.env.JEV_KEY }) |
new TypeSafeClient() |
jev.run({ text, questions }) |
client.systemOne({ state, questions }) |
yesNo("Is the buy flow informed?") |
noul(...) |
Jev is the model and TypeSafeClient is the client. The three primitives are Choice, Score and Noul.
One claim holds. The voiceover says Jev evaluates multiple questions in parallel, and the model docs say "Jev ingests the state once and evaluates every question against it in parallel."
Does Jev return a confidence score for yes/no questions?
No. The demo's yes/no card shows a "100% sure" badge. The Noul reference says: "There is no separate confidence value for a Noul, unlike a Choice or a Score." The single yes probability describes it completely. Anyone building a trading gate on that badge would be reading a field the API never sends.
The demo also caps input at 8,000 tokens. Jev's documented limit is "64k tokens per request; 32k tokens for state plus the longest question."
Is the Avellaneda-Stoikov math correct?
Yes, and it is the strongest part of the video. The on-screen spread formula, δ^a + δ^b = γσ²(T − t) + (2/γ)ln(1 + γ/κ), matches equation (3.18) of Marco Avellaneda and Sasha Stoikov's "High-frequency trading in a limit order book" (2006 preprint, published in Quantitative Finance in 2008).
The code matches too. r = s - q*γ*σ**2*t is the paper's reservation price, r = s − qγσ²(T − t). Setting bid, ask = r - d/2, r + d/2 centres the spread on the reservation price instead of the mid, which is the paper's "inventory" strategy.
The numbers on screen reconcile. In cycle 4, with mid 184.61, inventory +2 lots and T − t of 0.94, the engine posts bid 184.41 and ask 184.68. The gap is exactly the 0.27 spread, and the midpoint of 184.545 matches the reservation price of 184.55. In cycle 1, inventory is +3 and T − t is 0.98. Scaling cycle 4's skew of −0.058 by the formula gives 0.0907, against 0.091 displayed.
Is the demo running on live data?
It does not look like it. The video shows three unrelated prices: BTC-USD at 64,702.3 on the feed, BINANCE:BTCUSDT at 63,194.02 on a TradingView card, and a mid of 184.61 in the quote engine. The tiles also drift from the JSON they claim to summarise, such as "3.4 bps vs venue B" on a tile while the JSON says vs_venue_bps: 3.2. These look like animated mockups.
What is RLCD?
The video says Jev "is trained with RLCD instead of human preference." TypeSafe's launch post defines it as "Reinforcement Learning for Calibrated Decisions."
The acronym was already taken. RLCD: Reinforcement Learning from Contrastive Distillation, by Kevin Yang, Dan Klein, Asli Celikyilmaz, Nanyun Peng and Yuandong Tian, was submitted on 24 July 2023 and published at ICLR 2024. It is an unrelated preference-data technique, so searching the acronym can land on the wrong paper.
The calibration panel in the video compares Jev's confidence with market prices (Jev r = 0.99 against 0.97 for a reference model). That measures agreement with prices, not with outcomes, and its own table shows a miss: Jev gave 0.82 to "Unemployment rate ticks up," which resolved NO.
Bottom line
Jev is a real, fast decision model, and the Avellaneda-Stoikov quote engine in this demo is mathematically sound. The latency claim contradicts the video's own 1.50-second readout, the SDK code will not run as shown, and the yes/no confidence badge does not exist in the API. Use TypeSafe's docs, not this video, as the reference for building with Jev.
Sources
- Introducing System One Models & Jev, TypeSafe AI - launch post, latency range and RLCD definition
- TypeSafe docs: Models - context limits and parallel questions
- TypeSafe docs: Noul primitive - yes/no output format
- TypeSafe docs: JavaScript SDK - package, client and method names
- TypeSafe docs index - the three primitives
- Avellaneda and Stoikov, "High-frequency trading in a limit order book" - reservation price and spread formulas
- arXiv:2307.12950, RLCD - the earlier, unrelated RLCD paper
- InfoQ: TypeSafe AI Releases Jev - independent speedup figures
Published October 2, 2026. Based on a TikTok by @ray_fu. Full video notes (markdown)