https://dryhurst.io/articles/ Field notes from Dryhurst on applied AI, typed model judgments, and shipping AI into production software. en Mon, 28 Sep 2026 00:00:00 GMT https://dryhurst.io/articles/what-is-jev-system-one-model/ https://dryhurst.io/articles/what-is-jev-system-one-model/ Mon, 28 Sep 2026 00:00:00 GMT Jev answers typed questions about your data with calibrated probabilities, and never writes a word of prose. What it is, how it differs from an LLM, and where it fits in real software. TypeSafe Jev System One decision model structured outputs API Jev model AI decision engine https://dryhurst.io/articles/choice-score-noul-question-types/ https://dryhurst.io/articles/choice-score-noul-question-types/ Mon, 28 Sep 2026 00:00:00 GMT Every question you ask Jev is a Choice, a Score or a Noul. Here is what each one returns, how to pick between them, and how to compose them into decisions your code owns. AI structured outputs primitives Jev question types Noul primitive AI decision scoring TypeSafe primitives https://dryhurst.io/articles/decision-models-vs-llms/ https://dryhurst.io/articles/decision-models-vs-llms/ Mon, 28 Sep 2026 00:00:00 GMT Most AI features need a decision, not a paragraph. How decision models like Jev differ from LLMs, and a practical test for when to stop generating text. decision models vs LLMs System One vs System Two AI Jev vs LLM structured outputs alternative AI classification model https://dryhurst.io/articles/confidence-vs-probability-ai-escalation/ https://dryhurst.io/articles/confidence-vs-probability-ai-escalation/ Mon, 28 Sep 2026 00:00:00 GMT An AI answer is only half a decision. The other half is knowing when not to trust it. How Jev separates probability from confidence, and how we set the thresholds. confidence gated routing AI confidence score Jev probability calibration automated escalation AI decision model thresholds https://dryhurst.io/articles/speculative-fan-out-batch-queries/ https://dryhurst.io/articles/speculative-fan-out-batch-queries/ Mon, 28 Sep 2026 00:00:00 GMT Ask every question you might need, in one request, and let code ignore the rest. Why that is cheaper and faster than branching, and when a second call is genuinely required. speculative fan out Jev batch queries LLM parallel processing TypeSafe batch performance AI decision latency https://dryhurst.io/articles/jev-openrouter-decisions-api/ https://dryhurst.io/articles/jev-openrouter-decisions-api/ Mon, 28 Sep 2026 00:00:00 GMT OpenRouter serves Jev through a separate decisions endpoint, not chat completions. Here is the request shape, what it costs, and how fast it answered for us. OpenRouter decisions API typesafe/jev-1.13 Jev OpenRouter Jev API pricing OpenRouter decision model https://dryhurst.io/articles/claude-code-model-router-jev/ https://dryhurst.io/articles/claude-code-model-router-jev/ Mon, 28 Sep 2026 00:00:00 GMT Not every message to a coding agent needs the biggest model. How we used Jev to size each request and hand small jobs to Haiku and Sonnet, including the limits we hit. LLM model router Claude Code router Jev model router AI cost optimization Claude Code subagents https://dryhurst.io/articles/ai-agent-skill-picker-jev/ https://dryhurst.io/articles/ai-agent-skill-picker-jev/ Mon, 28 Sep 2026 00:00:00 GMT Agents pick skills from a cramped one-line index and often guess wrong. Here is how we put Jev in front of that choice, and what it costs. AI agent skill selection Jev skill picker Claude Code skills agent skill routing TypeSafe agent skill https://dryhurst.io/articles/ai-guardrails-citation-checking-jev/ https://dryhurst.io/articles/ai-guardrails-citation-checking-jev/ Mon, 28 Sep 2026 00:00:00 GMT A guardrail should hand you evidence, not make your policy for you. How we use Jev to screen LLM traffic and catch citations that do not hold up. LLM guardrails AI citation checking RAG verification prompt injection detection Jev guardrails https://dryhurst.io/articles/ticket-email-triage-fraction-of-a-cent/ https://dryhurst.io/articles/ticket-email-triage-fraction-of-a-cent/ Mon, 28 Sep 2026 00:00:00 GMT Sorting a pile of text is a decision problem, not a writing problem. Our Jev triage setup, what one email cost, and the rules we use when Jev is unsure. LLM triage AI email classification AI lead triage ticket classification AI low cost AI classification