# Jev Bureau (jevai.fyi)
> The Independent Technical Reference for TypeSafe Jev

Jev Bureau is an independent technical reference and evidence authority for Jev, the System One decision model developed by TypeSafe AI. It provides verified model specifications, API quickstarts, primitive schemas, and empirical pricing.

## Model Identity & Technical Facts
- Model Identifier: jev-1.13.0 (resolved by default alias 'jev-latest')
- Developer: TypeSafe AI (Founders: Diogo Almeida [CEO], Sasha Sheng [COO], Erik Gafni [CTO])
- Architecture: Non-autoregressive System One decision model trained via Reinforcement Learning for Calibrated Decisions (RLCD)
- Context Window Architecture:
  * 64,000-token total budget per request across shared state + all questions combined (TypeSafe direct)
  * 32,000-token limit for shared state + the single longest question (TypeSafe direct)
  * 32,000-token context window listed in OpenRouter model catalog
- Output Mechanism: Strict mathematical objects (Choice label, continuous Score float, or binary Noul probability float). Schema adherence is structurally constrained; calibration is trained across groups of predictions and does not guarantee individual answer correctness. No free-form prose text or chat generation.
- Latency Profile: TypeSafe reports response times between 70ms and 500ms on early access infrastructure.
- Token Pricing:
  * Input: $0.042 per 1,000,000 tokens ($42 / billion tokens)
  * Output: $0.00 / Free (non-generative)
- Official Endpoints:
  * TypeSafe Direct: POST https://api.typesafe.ai/v1/systemone
  * OpenRouter System One: POST https://openrouter.ai/api/v1/systemone (model: typesafe/jev-1.13 or ~typesafe/jev-latest)
  * OpenRouter Decisions Alpha: POST https://openrouter.ai/api/alpha/decisions

## Core Decision Primitives
1. Choice: Discrete categorical selection from an unordered closed set of up to 255 choices. Returns selected label, confidence float, and full probability distribution.
2. Score: Ordered rubric evaluation across 2 to 10 defined levels. Returns a probability-weighted continuous float score (which may land between levels) and confidence score.
3. Noul: Binary proposition evaluation. Returns probability float [0.0, 1.0] representing P(True). Does not include a separate confidence field because the probability expresses the binary degree of belief.

## Canonical Documentation Pages
- https://jevai.fyi/ : Canonical Jev entity overview, primary query owner, and search routing hub.
- https://jevai.fyi/primitives/ : Complete selection matrix, decision rules, and JSON input/output schemas for Choice, Score, and Noul.
- https://jevai.fyi/access/ : Production API quickstarts, client SDKs (Python: typesafe_sdk, TypeScript: @typesafe-ai/sdk), and endpoint specifications.
- https://jevai.fyi/jev-vs-llm/ : Architectural and operational comparison of System One decision engines versus generative language models.
- https://jevai.fyi/pricing/ : Published token rates, provider matrix, 64k/32k token limits, and client-side token cost calculator.
- https://jevai.fyi/typesafe-jev/ : Creator-to-model relationship, founding team biographies, RLCD training claims, and documented specifications ledger.
- https://jevai.fyi/changes/ : Permanent, dated ledger of model releases, platform expansions, and pricing updates.

## Editorial & Affiliation Notice
Jev Bureau is an independent publication. It is not affiliated with, endorsed by, or an official property of TypeSafe AI.