rajivkuriakose/typesafe-jev-examples
User repository on GitHub
Worked examples for TypeSafe's Jev System One decision model, runnable today through OpenRouter
- Forks
- 1
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Demos & Experiments
TypeSafe Jev Examples
Two worked Jev examples you can run today through OpenRouter without an early-access key: seven-question ticket triage batched into one request, and a re-ranker that fans one question out per candidate and sorts on the returned probability.
- Stars
- 2
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- rajivkuriakose
- Last verified
- Oct 2026
About this project
Worked examples for Jev, TypeSafe's first System One model — runnable today through OpenRouter, without an early-access key. System One models do not generate prose. You send state and a map of named questions; you get back typed answers with calibrated probabilities. There is no parsing step and no free-text-to-struct failure mode, so the interesting work moves out of prompt engineering and into ordinary code. TypeSafe's direct API is behind an early-access waitlist. OpenRouter serves the same model today:
Quoted from the project's README. Read the full README.
- Languages
- Python
Highlights
- Example 01 sends seven Choice, Score and Noul questions in one parallel request over a single ticket.
- Thresholds scale to stakes: 0.60 to route a ticket, 0.85 for the billing branch that can move money.
- Example 02 fans one Noul out per article across a thread pool, because each candidate is different state.
- make test runs offline with no key and no network, treating the routing policy as pure code.
Watch out
Everything here is verified against the OpenRouter path. The direct api.typesafe.ai path is written against the SDK's documented surface but has not been run, because it needs an early-access key, so treat that branch as untested code. OpenRouter serves Jev on a dedicated decisions endpoint and /chat/completions returns an explicit error, which is the first thing that will bite you wiring this yourself. The headline re-ranking figures - top-1 and top-3 at 100% against a keyword baseline of 67% to 83%, over 156 calls, 69,474 input tokens, roughly $0.0029 and 15.2s - come from 26 help-center articles and 6 queries, and four of the six were already correct. The keyword baseline does no length normalisation, so longer articles match more, and all six gold passages sit at or above the mean length; the README calls that conservative rather than neutral. The queries and passages were authored before any run and not changed after, which the README flags as a deliberate choice rather than a neutral fixture. This README copy is cut off mid-request-example, so the sections after the OpenRouter endpoint detail were not visible. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
Yes, that is the point of the repo. OpenRouter serves the same model today as typesafe/jev-1.13 at $0.042 per million input tokens with output free, and the client picks a provider from your environment: TYPESAFE_API_KEY set wins and routes to api.typesafe.ai, otherwise OPENROUTER_API_KEY routes to OpenRouter. Both speak the same protocol, so nothing above the client knows which answered. The README is explicit that only the OpenRouter path is verified; the direct path has never been run.
OpenRouter serves Jev on a dedicated decisions endpoint, not chat/completions, and returns a message saying typesafe/jev-1.13 is a decisions model and cannot be used with the chat/completions endpoint, pointing at /api/alpha/decisions instead. This is not a text model: it never generates prose, so there is no completion to parse, which is also why the whole example set is typed choices, scores and nouls.
The README claims identical answers and attributes the saving to the state dominating the request either way: asking seven roughly costs what asking one costs in wall-clock time, so sending the state once instead of seven times is where the gain is. It draws the line explicitly - questions batch only when they share the same state, which is why Example 02 fans out one call per candidate instead, since every candidate is different state.
No, and the README says so. It is 26 articles against 6 queries - 156 calls, roughly $0.0029, 15.2s - where four of the six queries were already answered correctly by keyword search, the baseline does no length normalisation and all six gold passages are at or above the mean length, and both sides are reported as ranges because ties meant the baseline expressed no opinion at all. For measured numbers against an established baseline the README points at TypeSafe's re-ranking cookbook instead.
Clone the repo, run make setup, copy .env.example to .env and paste a key, then make test runs offline with no key or network and make run triages three sample tickets against Jev.
By what the answer means: one of a fixed set is a Choice, whether a condition holds is a Noul, and a degree along a described dimension is a Score. The README stresses these are not three flavours of classifier.
No, the README calls a single global cutoff the common mistake. Routing to technical is cheap to get wrong and uses 0.60, while billing can move money and uses 0.85, and escalation is expressed as separate code conditions rather than an averaged score.
More Jev resources like this
46GitHub stars
Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.
Jev Demos & Experiments#community#jevbest#catalog-import
38GitHub stars
Research recipe that fuses two frozen open readers into Jev-style yes/no judgments with a fit-free router, confidence gate, and candidate sets.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
FeiLiuEM/open-medical-jevPython
FeiLiuEM/open-medical-jevPython62GitHub stars
Simulation study that drives one xArm7 apple-to-plate task with Jev, a large reasoning model, and a small model through identical incremental controls.
Jev Demos & Experiments#community#jevbest#catalog-import
33GitHub stars
Local HTTP server aiming at Jev API shape using a BERT-family encoder, working with the official SDK by swapping base URL and key while stating it is not Jev-compatible yet.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
hawkymisc/typed-decision-bertPython
hawkymisc/typed-decision-bertPython32GitHub stars
Jev-protocol random baseline that answers noul, choice, and score questions from fast pseudorandom numbers without reading the prompt.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
karminski/Jev-QuantumRust
karminski/Jev-QuantumRust36GitHub stars
Small decision model that returns per-answer probabilities for yes/no, 2-255 choices, or 2-10 level scores in one prefill over the SystemOne format.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
akivet/Mica-v0.1-4BPython
akivet/Mica-v0.1-4BPythonFrom the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.