User repository on GitHub
One client for every decision model: Choice, Score and Noul over Jev, OpenRouter, laya, MLX, CrossEncoders and LLM fallback
- Forks
- 0
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Practices & Patterns
Decide — gopaljigaur
pydecide is one Python client for typed System One decisions — Choice, Score and Noul — over hosted Jev, OpenRouter's Decisions endpoint, local laya on PyTorch or MLX, a sentence-transformers CrossEncoder, or a JSON-prompted LLM fallback.
- Stars
- 1
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- gopaljigaur
- Last verified
- Oct 2026
This is one of several Jev listings for the same project. The canonical entry is Decide — vsekhar.
About this project
decide is one Python client for typed decisions (Choice, Score and Noul) over any "System One" decision model: TypeSafe's hosted Jev, OpenRouter's Decisions endpoint, the open-weight laya family (PyTorch and MLX), any sentence-transformers CrossEncoder, and a JSON-prompted LLM fallback. The core of the library is the fallback chain: a Client takes an ordered list of these backends and a confidence policy, and if a backend errors or answers with low confidence, the next backend in the chain is tried, so a cheap…
Quoted from the project's README. Read the full README.
- Languages
- Python
Highlights
- A Client takes an ordered backend list plus a confidence policy, so a cheap local model backs off to a hosted one only when it is unsure.
- Client.from_env() needs no variables once a local backend is importable; it tries laya_mlx or laya first, then typesafe, openrouter and llm as keys appear.
- Gate thresholds can be global, per question or per type. Score answers are never gated, whatever a threshold says about "score".
- meta.route records how the chain went: "laya_mlx:ok", "<name>:error", or "<name>:low_confidence:<question>=<value><threshold>".
Watch out
Local backends are not free: the all extra pulls PyTorch and MLX, about 2 GB, and the mlx extra needs Python 3.11+ — on 3.10 it installs nothing and the laya_mlx backend is simply unavailable. The crossencoder backend cannot be selected from the environment at all; you construct it by hand and pass it into the Client. laya now ships its own laya-serve, so the bundled HTTP server only earns its place when you want several backends, local and hosted, behind one TypeSafe-compatible endpoint. The README calls the JSON-prompted LLM fallback the least trustworthy of the six backends, and the quickstart outputs are from a single local run with no environment variables set. 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.
The endpoint is one POST with bearer auth, and the README says as much. What the client adds is the chain: you hand Client an ordered list of backends and a Gate, and a backend that errors or comes back under threshold is recorded in meta.route before the next one is tried. You also get a decide ask CLI and an HTTP server that speaks the TypeSafe wire protocol if you want local models behind a hosted-shaped endpoint.
Only if you configure it. from_env puts an importable local backend first, and a hosted backend is only selected when its key is already in the environment, so with just pydecide[mlx] installed nothing leaves the machine. Once a hosted key is present it sits in the chain and takes over when the local one fails the gate, and route on the result tells you which backend answered and how the ones before it failed. DECIDE_BACKENDS overrides the detection entirely.
The all extra pulls PyTorch and MLX, which the README puts at about 2 GB. The mlx extra requires Python 3.11 or newer, and on 3.10 it installs nothing, leaving laya_mlx unavailable. You do not need any of it to start: the base package plus a hosted key is enough, FastAPI and uvicorn already ship in the base install, and the server extra is deliberately empty so existing install commands keep working.
The README claims no calibration and reports no measurement of it. The mechanism is documented, though: for a Choice the gate reads the top probability, for a Noul it reads max(noul, 1-noul), and per_question or per_type thresholds take precedence over the global min_confidence. Score answers are excluded from gating. Weigh that against the README's own verdict on the JSON-prompted LLM backend before you put one at the end of a chain.
The import is the swap. TypeSafeClient subclasses the SDK response, so code reading answers, choices or usage is untouched, and result.decide adds backend, model, latency and route. Errors keep the vendor types, so TypeSafeInternalServerError for a 503 and TypeSafeAPIConnectionError for a transport failure still catch. With a fallback and every tier failing you get an error that is both a TypeSafeError and a decide.AllBackendsFailed.
The backend table covers hosted TypeSafe at api.typesafe.ai, OpenRouter's alpha/decisions endpoint, OpenAI's Decisions beta with gpt-6-luna, local laya on PyTorch or MLX, a sentence-transformers CrossEncoder, and any OpenAI-compatible chat model, which the README calls the least trustworthy of them. The CrossEncoder cannot be selected from the environment; you construct it by hand and pass it into the Client.
decide ask takes the text and the questions as flags, for example decide ask "I had a rough day" --noul "Is the writer doing well?" --choice good,bad, and prints a table of answers with probabilities plus a footer line reporting backend, route and latency. The README's shown output is a real run against laya_mlx with no environment variables set at all.
More Jev resources like this
4GitHub stars
Framework-agnostic guard layer that puts Jev in front of an agent's tool calls: a Choice risk tier, a Noul destructiveness check and a Score blast radius decide allow, deny or ask.
Jev Practices & Patterns#community#awesomejev#catalog-import
Community
lgy1027/jevshieldPython
lgy1027/jevshieldPython1GitHub stars
Runnable benchmark putting Jev, jev-preview and Claude Sonnet 5 through one real agent task — classifying a tool call as readonly, destructive, privileged or exfiltration — on the same 60 labelled cases, with raw per-call results committed.
Jev Practices & Patterns#community#awesomejev#catalog-import
Community
themsquared/jev-benchmarkPython
themsquared/jev-benchmarkPython16GitHub stars
Vendor-neutral harness that tests whether Jev reranking improves a small RAG stack, using frozen hybrid BM25 plus BGE-M3 candidates and isolating retrieval, rerank, and generation stages.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
erendikmenn/jev-rag-benchmarkPython
erendikmenn/jev-rag-benchmarkPython17GitHub stars
Local offline System One server speaking the hosted /v1/systemone wire format, so existing SDK code runs unchanged against small on-machine models.
Jev Practices & Patterns#community#awesomejev#catalog-import
Community
amithgc/local-jevPython
amithgc/local-jevPython8GitHub stars
Line-oriented semantic search that adds a Jev meaning pass over your own shortlisted files next to exact search. Real search is the default with a key, while dry-run previews the payload offline with no network or credentials.
Jev Practices & Patterns#community#awesomejev#catalog-import
Community
larguesa/jev-searchPython
larguesa/jev-searchPython7GitHub stars
Four read-only MCP tools that rank files, cut a file down to the relevant part and pick from long lists, measured on a frozen SWE-bench Lite split.
Jev Practices & Patterns#community#awesomejev#catalog-import
Community
Bentlybro/siftrPython
Bentlybro/siftrPythonFrom the Jev guides
Original write-ups on what this entry is for.
Most Jev accuracy problems are state problems. What belongs in a state, what is a pre-made judgment to leave out, and why naming the slice you ask about matters more than the wording.
Jev Practices & Patterns#state#design#primitives
Read article
A Noul returns one number and no separate confidence. How to read the value, pin its boundary with criteria, split compound questions, and set a per-action threshold that matches the cost of being wrong.
Jev Practices & Patterns#noul#primitives#confidence
Read article
A Score is a weighted position, not a ruler. Why the number can land between levels, why levels are judged separately, and how to combine several atomic scores in code instead of one crowded rubric.
Jev Practices & Patterns#score#primitives#scoring
Read article
From the Jev community
Posts from builders shipping with Jev right now.