Skip to main content
FitForJev

Jev 1.13 · fit-v1 rubric

Paste a fuzzy problem.
Get typed questions back.

Turn fuzzy AI problems into typed questions Jev answers. Code computes the fit verdict.

100ms vs 2,500ms (25x) · $0.004 vs $0.50 per 100k · bands ≥0.72 / 0.45–0.71 / 0.25–0.44 / <0.25

Try an example

Your verdict lands here

Paste one fuzzy problem above, press Go, or pick an example. You get a typed verdict and decision core.

The pipeline

Inbound, code prep, Jev, guard, route

Follow five steps.

  1. 01 · Inbound

    Problem in

    Paste your fuzzy idea: a bot, a router, a scorer, a guardrail.

    "AI support bot that triages tickets and drafts replies."

  2. 02 · Code prep

    Decompose

    Split it into atomic sub-decisions you answer in seconds. If it does not split further, narrow the scope.

    intent · severity · churn risk · escalation · reply draft

  3. 03 · Jev decisions

    Jev mapping

    Each sub-decision maps to one primitive. You route generation to an LLM. You keep math, dates, regex in code.

    Choice gives the category. Score gives urgency. Noul gives needs_human.

  4. 04 · Code guard

    Compose

    You compose rubric answers into the verdict in code. You grade the fit, Jev grades the questions.

    computeFit returns native, hybrid, weak, not_jev, or use_code.

  5. 05 · Route

    Architecture

    You keep control flow in code, send semantic judgment to Jev at 100ms, send prose to the LLM.

    You run code first and send judgment to Jev. A code guard checks the result. You send the LLM only prose.

The rubric

Five signals add to the score. Three penalties subtract.

You answer ten literal questions in parallel. Your code composes the answers into a fit score.

  • 0.30

    Atomicity

    You split it into independent checks you answer in seconds.

  • 0.25

    Decision-centric

    You get a category, rating, or yes/no your software acts on.

  • 0.20

    Bounded output

    You pick from a fixed set, ordered levels, or yes/no.

  • 0.15

    Text state

    You fit all needed facts in text. Images, audio, video need a different tool.

  • 0.10

    Latency pressure

    You run it many times per minute or inside a user-facing request.

Economics · 100k calls

100ms vs 2,500ms (25x)

$0.004 vs $0.50 per 100k

Jev versus frontier LLM latency and cost per 100k calls. Source: TypeSafe docs.
Jev latency100ms
Frontier LLM2,500ms
Jev cost / 100k$0.004
LLM cost / 100k$0.50

Source: TypeSafe docs (opens in new tab).

Estimated savings ≈ $5.95 per year at 100k per month

Code owns
<1ms · $0
Jev owns
about 100ms · $0.042 per M tokens
LLM owns
2,000 to 8,000ms · $5 to $15 per M tokens
Human owns
hours/days · high

$0.042 per 1M inputs · free output · 64k context / 32k state · 250k tokens/s · 1,200 requests/min · RLCD (Reinforcement Learning for Calibrated Decisions) · 70 to 180ms

Send judgment to Jev at 100ms. Call the frontier model for prose.

Definitions

What Jev is, in one minute.

What is Jev

Jev is TypeSafe's calibrated decision engine for semantic judgment: Choice picks a category from a fixed set, Score rates along ordered levels, and Noul answers yes/no with a confidence your code acts on. Latency is about 100ms and cost is $0.042 per million input tokens with free output.

Choice vs Score vs Noul

Use Choice when the answer is one of a fixed set of categories, like intent, component, or tool. Use Score when the answer is an ordered level, like urgency P0–P3 or fit 1–5. Use Noul when the answer gates action, like needs-human, has-PII, or injection-detected. Combine them per pattern.

When not to use Jev

Do not use Jev for prose generation, multi-step reasoning, or exact precision such as math, dates, or regex. Send those to an LLM or keep them in code. You send judgment to Jev, compose the verdict in code, and call the LLM only where prose is the product.

More in the FAQ.

Pattern catalog

You reuse these patterns.

8 patterns from the TypeSafe cookbooks. Match keywords, pull schemas from the catalog.

Match all 8 by keyword. Match your problem above