Inference
Jev is a System One model from TypeSafe AI. It reads state and returns typed answers with probabilities. It does not write text, so there is nothing to parse and nothing to interpret.
Published specification
| fact | value | source |
|---|---|---|
| maker | TypeSafe AI | source |
| announced | 15 September 2026 | source |
| model class | System One model | source |
| input | text, JSON objects and arrays of text | source |
| output | typed answers with probabilities, no text | source |
| question types | choice, score, noul | source |
| request budget | about 32,000 tokens shared by state and questions | source |
| sampling | every question in a request answered in parallel | source |
| training method | Reinforcement Learning for Calibrated Decisions | source |
| endpoint | POST api.typesafe.ai/v1/systemone | source |
| model string | jev-latest | source |
| named after | William Stanley Jevons | source |
Measured on this site
| measure | value |
|---|---|
| requests sent | |
| questions answered | |
| fastest answer | |
| median answer | |
| nine in ten under | |
| ninety nine in a hundred under | |
| slowest answer | |
| input tokens used | |
| mean tokens per request | |
| mean confidence on the routing answer |
Latency distribution
| answered in | requests | share | |
|---|---|---|---|
| under 100 ms | |||
| 100 to 149 | |||
| 150 to 199 | |||
| 200 to 249 | |||
| 250 to 299 | |||
| 300 to 399 | |||
| 400 to 499 | |||
| 500 to 749 | |||
| 750 to 999 | |||
| 1000 and over |
Request schema
route
choice
Given `trade`, `wallet`, `pool`, `flow` and `routed_so_far`, what does this token's market most need the fee from this one trade to do?
| option | description |
|---|---|
| burn | Buy the token and destroy it. Best when the pool is deep enough for the trade sizes it sees, demand is present, and reducing supply helps holders more than anything else. |
| deepen | Add the fee to pool liquidity. Best when `trade.size_against_depth_pct` is high, `pool.liquidity_to_market_cap_pct` is low, or a single trade can move price too far. |
| holders | Pay the fee out to wallets that have kept their tokens. Best when selling pressure is high, when a wallet is leaving, or when supply is concentrated in few wallets and the ones who stayed are absorbing the sale. |
regime
choice
Which description fits `flow` right now?
| option | description |
|---|---|
| compressed | Buys and sells are balanced and price has barely moved over the last 50 trades. |
| directional | One side clearly leads and price has moved steadily in that direction. |
| dislocated | Price has moved sharply, led by a few large trades relative to pool depth. |
pool_stress
score
How hard does `trade` press on `pool`?
| option | description |
|---|---|
| The pool absorbs this trade with almost no price movement. | |
| The trade moves price noticeably. | |
| The trade moves price severely for a pool of this depth. |
concentration_risk
score
How concentrated is ownership according to `pool.top_ten_share_of_supply_pct` and `wallet.share_of_supply_pct`?
| option | description |
|---|---|
| Ownership is widely spread. | |
| A small group holds a meaningful share. | |
| A few wallets could move the market alone. |
exit
noul
The wallet in `wallet` is leaving its position entirely with this trade.
Why typed output
A chat model asked to split a fee would answer in a sentence, and the sentence would have to be parsed, checked and trusted. It could refuse, wander, or return numbers that do not add up. Jev cannot do any of that. It can only return a probability for each of the options it was given, and those probabilities always add up to 1. That makes its answer usable as money without anything in between. It also answers fast enough, and cheaply enough, to be asked about every single trade as it happens instead of once a day about all of them.