$40 Million For A Function. We Built The Loop On One Desktop.

Verdict: a decision model that never learns what happened next is a rented function. Jev is a rung. We built the ladder — in a weekend, on one desktop, and you can install it before you finish this paragraph.

claude mcp add awdecide -- uvx awdecide mcp

What that one line gives your agent, measured:

• 97.0% accuracy from a brain that is only right 70% of the time — on 52 model calls instead of 400. Same brain. Run awdecide bench and watch it.
• 35,177 ms → 9.0 ms on a repeated decision, live on an ARC-AGI-3 fork, with zero model calls in the second answer.
• A wrong answer is never given twice. Report it wrong once and it is retired at that fork for good. A stateless model will hand it to you again tomorrow, at full price, with the same confidence.
• Probabilities that were earned. A 60/40 coin reads 0.60. Every answer is graded against the base rate on a Brier ledger you own — a hosted decision model cannot show you that ledger, because it never learns how its answers turned out.
• No account. No waitlist. No data leaves your machine. Standard library only, one SQLite file, Apache-2.0.

On September 15, TypeSafe AI launched Jev with a $40 million seed round led by DCVC at a reported valuation near $200 million. Jev takes a piece of state and a typed question and returns a choice, a score, or a yes/no with a probability. It is fast and it is cheap. It is also stateless: one endpoint, no way to tell it how the decision turned out, no memory of the last time you asked. The ten-thousandth identical question costs what the first one did and is exactly as likely to be wrong.

That is the architecture forty million dollars was raised on, and it is the wrong one. The decision is a function. The loop it sits in is the product.

What independent testers measured

TypeSafe's headline is "193.6x faster, 444.6x cheaper". Its own launch post says to expect those "on the higher end of real world gains", and the reference answers in that benchmark are the average of GPT-6 Astra and Claude Fable 5.1 — agreement with two other models, with no answer key. So look at the people who brought one.

Against human-labelled items, one independent bench ran Jev 1.13 and five frontier models over 200 typed decisions from four public datasets:

Last in accuracy, last in calibration — from the model whose entire pitch is a probability you can trust. (The author notes 50 items per task, so gaps under ten points are within noise. Jev's gap to the leader is 11.5.)

On a task it could not have seen, a second independent test generated 900 support tickets from rules. Jev's calibration error came in at 0.107, which the author puts at 4.4 times the noise floor. On the priority question — whose answer depended on a company policy that was not in the ticket — Jev scored 44.7% and reported its picks at an average probability of 0.74. The information did not exist, and it was three-quarters sure. The direction of the error also flips by question type: over-confident on choices and scores, under-confident on yes/no. One threshold cannot be safe for all three.

Its speed held up: a 0.43 second median at $0.025 per thousand decisions. That part is real, and it is the only part.

What the loop does instead

Ask, act, report what happened. The next identical decision is answered from evidence instead of from a model. On our full stack the same brain goes from 72.0% to 97.5% on a tenth of the model calls, and a judge scores 100% at zero calls where an LLM judge needed 24 calls and 523 seconds for the same 24 labelled items.

And the property a static model cannot have: it knows where its competence ends. Before a learned head is allowed to answer, it has to clear four gates, including a lower confidence bound and an out-of-distribution probe. Outside what it was taught it says so, and the call goes to the brain. It answers 0.9965 of what it is competent for and hands back the rest. Three-quarters sure about a policy it has never read is precisely the behaviour those gates exist to stop.

These are different benches, and we will say so plainly: nobody has yet run Jev and the loop on the same items. We will, the day we are off the waitlist — and the loop makes that easy, because Jev can simply be the cold rung.

Use it today, in the harness you already have

You do not need our platform, an account, or a waitlist. The loop ships as awdecide — Aither World Decide — a standalone package with zero dependencies.

Claude Code:

claude mcp add awdecide -- uvx awdecide mcp

Codex, in ~/.codex/config.toml:

[mcp_servers.awdecide]
command = "uvx"
args = ["awdecide", "mcp"]

Your agent gets decide before a bounded choice it makes over and over, and decide_outcome after it sees what happened. The harness's own model is the brain on the first sighting. The ledger is the memory on every sighting after. Want the loop to have its own brain? Point AWDECIDE_LLM_URL at any OpenAI-compatible endpoint — Ollama, vLLM, a frontier API, or Jev.

Then check our arithmetic on your own machine:

pip install awdecide && awdecide bench

Forty situations, ten sightings each, one deliberately imperfect brain that is right 70% of the time:

Same brain. More accurate because a wrong answer gets reported and retired; cheaper because a right one is never asked for again. Change --seed and rerun it — it lands at 95–97% on 52–59 calls every time. That is the difference between a benchmark you are told about and one you can run.

What to do with this

A fast static decider is a component. It belongs on the cold path of a ladder that remembers, abstains and improves — the rung you fall to when nothing has been learned yet. Plug in Jev, a local 8B, or a frontier model; the ladder is the product, and the ladder is what gets cheaper every day you run it.

Forty million dollars bought a function that charges you again for the answer it gave you yesterday. The loop runs on one desktop, and yesterday's answer is free.

The full numbers: 97.5% Accuracy On A Tenth Of The Model Calls.

Sources

• TypeSafe AI, Introducing System One Models & Jev
• jev-frontier-bench — Jev 1.13 against five frontier models, 200 typed decisions
• jev-ood-calibration — 900 rule-generated tickets plus three public benchmarks
• The Register, TypeSafe AI debuts model for machines
• TS2, TypeSafe AI raises $40 million for Jev, but its 445x cost claim is still self-tested
• Forkast, TypeSafe AI's Jev is not an LLM