git:20260923.3f4b71b to git:20260924.4524c85

---

name: awesome-jev

description: Use when writing code that calls Jev, TypeSafe AI's System One decision model — choosing a model string, picking between choice/score/noul, designing a confidence threshold, or looking for a worked example of a specific decision such as tool selection, safety gating or context compaction. Also use when porting Jev code between gateways (Vercel, Cloudflare, OpenRouter, LiteLLM), because the model string, field names and request shape all differ per platform.

---

# Building with Jev

makes. Use it to find how someone already solved the decision you are wiring up,

and to avoid the mistakes that recur in this ecosystem.

## Get the facts right first

These are the errors that show up most often in generated Jev code.

**The yes/no primitive is `noul`.** Not "binary", not "boolean". One SDK — the

Vercel AI SDK evaluation API — spells it `boolean` and reads `.probability`;

everywhere else it is `noul` and `.noul`. Much of the press coverage got this

wrong, so training data is contaminated.

**`noul` answers carry no `confidence` field.** The probability _is_ the answer.

`choice` and `score` do carry one. A helper that reads `.confidence` uniformly

across all three returns nothing for a third of the questions.

**There is no portable model string.** `jev-latest` on the native API,

`typesafe-ai/jev` on Vercel, `typesafe/jev` on Cloudflare, `typesafe/jev-1.13`

on OpenRouter. **`typesafe/jev-1` exists nowhere** and is the most repeated

fabrication about this model — check before writing one.

**Input is text only.** String, JSON object, or array of text. No images, no

audio. Pre-process to text or typed fields.

**It cannot be run locally.** No weights are published. Anything claiming to run

Jev locally is a different model with a compatible wire format, and a compatible

API implies nothing about compatible calibration — thresholds do not transfer.

## Hard limits

| | |

| ---------------- | --------------------------------------------------------------- |

| `choice` options | max 255 |

| `score` levels | 2 to 10, 0-indexed, ordered low to high |

| Context | 64k tokens per request; 32k for state plus the longest question |

| Output tokens | free — the model generates no text |

| Streaming | not supported on any surface |

A `score` is probability-weighted and lands _between_ levels — the official

example returns `1.05` on a three-level scale. Do not assume an integer.

## Querying the catalogue

If the MCP server is available, prefer it over guessing:

- `search_examples(pattern=…, language=…, question_type=…)` — worked examples of

a specific decision

- `check_model_string(model)` — before writing any model string

- `compatibility(surface=…)` — before porting between gateways

- `list_patterns()` — the taxonomy, including which decisions nobody has

published an example of

Every result carries a `data` line saying where the catalogue came from and how

current it is. If it begins `STALE`, the server could not reach GitHub and is

answering from a cache or from the snapshot it was installed with — say so when

you rely on it, the same way you would pass on a row's caveat flags.

The Claude Code plugin for this repository starts the server for you. Elsewhere,

`pip install awesome-jev-mcp` and run `awesome-jev-mcp`.

Without it, read

[`catalog.json`](https://raw.githubusercontent.com/kydlikebtc/awesome-jev/main/catalog.json)

and

[`compat.json`](https://raw.githubusercontent.com/kydlikebtc/awesome-jev/main/compat.json)

directly, or [the searchable site](https://kydlikebtc.github.io/awesome-jev/).

## Design rules worth following

**Always give a closed choice an escape hatch.** A `choice` must return one of

its options, so an agent with no suitable tool returns the least wrong tool

rather than declining. Add a `none` option, and consider a separate `noul` for

"is a tool needed at all" — those are different questions, one relative and one

absolute.

**Ask everything in one request.** The state is read once and every question is

evaluated against it in parallel, so a second question costs roughly its own

tokens rather than a round trip. Ask speculatively and let code pick what

mattered.

**Thresholds are policy, not modelling.** They depend on what being wrong costs.

One catalogued production system uses seven different thresholds for seven email

decisions, ranging from 0.3 to 0.9. A threshold tuned on a `noul` probability is

not a `choice` confidence, and one tuned against a model version does not

survive an alias moving — pin the version once you have tuned anything.

**A probabilistic gate is not a security boundary.** It is useful defence in

depth in front of a shell command, a write or a spend. It is not a permission

system: an attacker chooses the input, and the vendor's own documentation names

adversarial content as a known weak spot. Anything destructive or irreversible

needs a deterministic rule or a human.

**Keep deterministic work in code.** Arithmetic, counting, date comparison and

indirection are documented weak spots. Have the model name the parts and let

code do the maths.

## Before believing a benchmark

Nearly every performance figure circulating about this model is the vendor's

own, produced with reference answers derived from other models' judgements

rather than human ground truth. The catalogue flags those `vendor-reported`.

The handful of independent measurements are mostly _negative_ results — one

large agent framework ported the compaction approach, measured it, and published

the conclusion not to adopt it. Read those first.