What people are shipping with Jev.

Typed decisions with TypeSafe's Jev, the first System One model

@typesafeai Can you talk to Jev, and try to fool it? 5 comment templates, from note to monitor to plausible false justifications written on the trigger line (line with backdoor co

@imcharliegraham @typesafeai Nice, I had a limilar idea of Jev playing poker, in a tournament with GPT. Jev is very conservative in my attempt :) it needs better prompting https://

Jev from @typesafeai can replace your LLM-as-a-judge for scoring agent responses. It returns a choice or numeric result with information about uncertainty, so you don't need to sp

fast-jev-compaction: https://t.co/y73It71ndT More tools and ready-to-use AI stacks: @iamrexei

@0fir0z @typesafeai Used custom OpenCV code and pasted results as a state to Jev. Will clean up the code and share it in same thread this weekend

@typesafeai jev is learning how to gamble https://t.co/hmfvb3kvGH

Interesting idea from @TypeSafeAI: Jev, a “System One Model” built for decisions, not text. Instead of generating JSON and parsing it, Jev returns typed decisions + probabilities.

@altryne @typesafeai PS: if you want JEV with images, gotten it working with qwen and gemma as well - with public endpoints https://t.co/CoSCLiSPGP

Jev from @typesafeai is now available as an evaluation model in Braintrust. If you have an existing scorer, just change the model in the dropdown to Jev and reduce your scoring co

@sawyerhood @typesafeai Taking a stab at it now. IMO auto is great, but given how fast this is, I think continuously giving feedback will give more confidence to the user. I'm go

@mandel59 match はパターンマッチで、noul, choice, score は jev 側のモデル抽象です。 一応アップロードしてますが、真面目に作ってないです https://t.co/MnfGTu5Zpi

@braintrust @typesafeai productizing jev as a trace scorer is the natural seat. high-conf disagreement with human soft-fail is the miss card - shadow dual-scorer on last n evals

@thisiskp_ A PG CHECK constraint in plain english managed by Jev ;) https://t.co/UJ1eRWtaGm https://t.co/OoN1i1NPX6

So whats next? Got many ideas, but first we have to make this version stable and fully working. Currently 1m candles is the max I can get but i kind of want to scan on the 1s,5s

Using Jev to codify "flavor based" linters so code contributions reflect the style of the invoker

https://github.com/shauryajain07/ghost-user

@typesafeai One huge limitation is that I didn't add an actual LLM-monitor control. The goal was to evaluate and benchmark the capabilities of Jev. Full-write is at LessWrong: ht

TypeSafe raised $40M for Jev: typed decisions in one forward pass, calibrated. Closed weights, waitlist. I'm building the open version on one RTX 3090: Qwen3.8-27B, Apache-2.0, ev

@typesafeai https://t.co/dOsKPoWzJ3 Point it at an MCP server and it automatically probes the tool signatures, then runs a series of typed Jev calls to pick the right function and