--- source_hash: 283c407016ea9e27902d592a3e918cf76464caa19d02b64c5d6e9ded2c144430 post: /posts/tools-and-plugins/jev-the-decision-model-that-doesnt-chat --- # Context Pack: Jev The 19 September introduction now has an explicitly dated 17 October 2026 provider update. Route: /documents/tools-and-plugins/jev-the-decision-model-that-doesnt-chat. Jev is TypeSafe's official name; System One describes the decision-model category. The launch explanation attributes the naming to Kahneman's fast judgements/Jevons and describes parallel sampling/RLCD. Do not claim independent architecture validation. ## Verbatim excerpts > Jev takes a state and some questions. It returns values your code can use. > A typed answer makes integration easier. It does not make an incorrect option correct. > The model supplies a judgement. Your application owns the action. ## Core contract State plus typed questions enter; typed assessments leave. Choice returns an option, probabilities and confidence. Score returns a probability-weighted position along an ordered rubric (possibly between levels), legend, probabilities and confidence. Noul returns a probability from 0 to 1 and no separate confidence. Native confidence is a statistic of the returned distribution, not simply the probability that the selected answer is correct. Evaluate thresholds on representative cases/consequences; calibration is aggregate and high confidence grants no authority. State can be text/JSON/text array. Curate relevant facts, reference nested fields with backticked paths and write the full question in instructions; question IDs are not sent to the model. Ask independent atomic questions together; use another request only for a genuine answer dependency. Code combines weights/thresholds and owns side effects. Preserve act/gather/escalate routes and deterministic arithmetic/date/once-only action handling. ## Worked examples to retain - Support Export freeze: category Choice; severity Score; reproduction Noul; refund Noul; frustration Score. Illustrative bug/high-enough severity/usable repro/low refund/concerned result, not a live call. Ignore unused independent results. - RAG: relevant/evidence/contradiction/injection Nouls; route evidence/conflict/drop. EU help-centre versus US-only scope and hostile forum scrap are examples; filtering is not injection security. - Citation: exact string presence first, reject missing quote without model; surviving quote gets support/contradict/unrelated Choice plus review rule. - Guardrails: hazard Nouls/severity Score around a generative call; pass/review/block/support, including supportive routing for self-harm. - Dates: extract named parts, compute calendar in code. The parallel-questions cookbook reports 13 questions on a GDPR page with jev-1.12, 12.2× lower cost and 10.0× speed. Answers mostly unchanged, not identical: breach_72h mean 0.804 batched/0.814 single. Attribute the experiment, never generalise its multiplier. Launch claims of 100 ms/70–500 ms and high multipliers are vendor claims; no latency measured here or subsidy conclusion. The September 255-option limit is dated; check current Choice reference. ## Provider snapshot: 17 October 2026 Direct TypeSafe: 1.13.0 and 1.14.0; latest→1.13, preview→1.14. Both cost $0.042 per million input tokens, or $42 per billion; output is free. Version 1.14 was released on 14 October. Direct alias changes were planned, not yet completed. Text was the original essay's scope. Experimental images now work on both, Box only on 1.14; off by default, explicit TypeSafe-Danger-Zone-Images header, not recommended for production. No audio/video inference claimed. Defaults are 64k per request and 32k for the state plus any question. Experimental long-context header raises both to 262k with speed/accuracy caveats. Current rate limits are organisation-specific; September figures of 250,000 tokens/s and 1,200 requests/min are historical, not entitlements; 429 on reaching a limit. Pin versions/log results and re-evaluate thresholds when changing model. SDKs: typesafe-sdk, @typesafe-ai/sdk, TYPESAFE_API_KEY, typesafe-ai/skills. Vercel typesafe-ai/jev uses current experimental_decide and ai>=7.0.128; evaluate is historical. Cloudflare env.AI.run('typesafe/jev', ...) catalogue lists a 32,000-token context. OpenRouter ~typesafe/jev-latest tracks 1.14, unlike direct latest; access there has no waitlist or separate TypeSafe account and does not prove direct-service GA. LangChain langchain-typesafe/TypeSafeClassifier plus experimental routing/tool-risk middleware; refusal is not an approval workflow. Pydantic TypeSafeModel/typesafe:jev-latest maps output fields to questions. Provider names/shapes/limits need per-surface checks; yes/no may be called boolean. Maths, counting, dates, multi-step reasoning and generation belong elsewhere. Irrelevant/contradictory/adversarial state can mislead. English strengths do not transfer automatically; Noul/Choice about a similar fact need not agree. Typing is not truth or injection immunity. Provider privacy policy promises no training/fine-tuning on customer Input; enterprise ZDR is an arrangement, not default for everyone. Current enquiry contact sales@typesafe.ai. This is a published-commitment summary, not an independent audit. Original ImaginaryTools/CXN use is the September personal account, not fresh production verification. ## Primary references - https://typesafe.ai/blog/introducing-system-one-models-and-jev - https://typesafe.ai/blog/jev-1-14 - https://docs.typesafe.ai/models - https://docs.typesafe.ai/images - https://docs.typesafe.ai/confidence - https://docs.typesafe.ai/primitives/choice - https://docs.typesafe.ai/patterns/fan-out.md - https://docs.typesafe.ai/patterns/confidence-routing.md - https://docs.typesafe.ai/cookbooks/parallel_questions.md - https://docs.typesafe.ai/cookbooks/classifying_rag_passages.md - https://docs.typesafe.ai/cookbooks/citation_check.md - https://docs.typesafe.ai/cookbooks/llm_guardrails.md - https://docs.typesafe.ai/cookbooks/date_extraction_cookbook.md - https://docs.typesafe.ai/patterns - https://docs.typesafe.ai/sdk/python - https://docs.typesafe.ai/sdk/javascript - https://docs.typesafe.ai/agent-skill - https://ai-sdk.dev/docs/reference/ai-sdk-core/decide - https://vercel.com/i/jev-integrations - https://developers.cloudflare.com/ai/models/typesafe/jev/ - https://openrouter.ai/docs/guides/community/jev - https://docs.langchain.com/oss/python/integrations/providers/typesafe - https://pydantic.dev/docs/ai/models/typesafe/ - https://typesafe.ai/legal/privacy-policy - https://docs.typesafe.ai/legal Keep two original fenced diagrams, three image diagrams, related essays and the example/dated-claim boundaries. Old draft placeholders and unavailable private working-document references are not required reading.