# Decision Models (MoC) ## Overview [[Decision Models (DMs)|Decision models]] return a decision (a choice, a yes/no probability, a score) instead of text. [[TypeSafe AI]] turned the idea into a product with [[Jev]] in September 2026 and calls it a [[System One Models|System One model]]. I went deep on it because it's a new building block between code and LLMs: bounded judgment that's cheap enough to call thousands of times, with probabilities you can actually put a threshold on. If you only read three notes, read [[Decision Models (DMs)]], [[Reinforcement Learning for Calibrated Decisions (RLCD)]] and [[Confidence-Gated Routing]]. ## The models - [[Decision Models (DMs)]]: the category - [[System One Models]]: TypeSafe's name for it - [[Jev]]: the first commercial one - [[TypeSafe AI]] and its co-founder [[Diogo Almeida]] - [[Machine Native Intelligence]]: the "prod, not God" thesis behind it - [[Jevons Paradox]]: where the name comes from, and the economic bet - [[System 1 and System 2 Thinking]]: where "System One" comes from ## How they're trained - [[AI Post-Training]]: the three branches (RLHF, RLVR, RLCD) - [[Reinforcement Learning From Human Feedback (RLHF)]] and [[InstructGPT]] - [[Reinforcement Learning with Verifiable Rewards (RLVR)]] - [[Reinforcement Learning for Calibrated Decisions (RLCD)]] - [[AI Model Calibration]] - [[Proper Scoring Rules]] - [[Mode Collapse]] and [[Typicality Bias]] - [[AI Sycophancy]], [[Reward Hacking]] and [[AI Hallucination]]: what preference training does wrong ## Building with them - [[System One Primitives]]: Choice, Score, Noul, state and confidence - [[Atomic Question Decomposition]]: code in control, one judgment per question - [[Relative vs Absolute AI Judgments]]: "which one" vs "does any fit" - [[Speculative Fan-Out]] - [[Confidence-Gated Routing]] - [[Composite Scoring]] - [[AI Model Cascades]] and [[Model routing]] - [[Candidate-Then-Select Extraction]] - [[Reranking]] - [[Hierarchical Classification]] - [[LLM-Generated Features for Classical ML]] - [[AI Guardrails]] ## Trusting them - [[Shadow Evaluation]]: test on your own traffic before switching - [[Conformal Prediction]]: coverage guarantees that don't depend on the vendor's calibration - [[Self-Consistency]] - [[Zero-Shot Classification]]: the "this isn't new" lineage - [[AI Context Rot]] and [[AI KV Cache]]: why Jev isolates questions over a shared state - [[Jagged Intelligence]] ## Open alternatives and tools - [[Laya]] - [[Kev]] - [[SemIf]] - [[Jevlike]] - [[fast-jev-compaction]] ## Notes <!-- QueryToSerialize: LIST FROM [[Decision Models (DMs)]] AND #type/permanent_note AND !#type/quote AND !#type/creation/quote WHERE public_note = true SORT file.name ASC --> <!-- SerializedQuery: LIST FROM [[Decision Models (DMs)]] AND #type/permanent_note AND !#type/quote AND !#type/creation/quote WHERE public_note = true SORT file.name ASC --> - [[AI Guardrails]] - [[AI Model Cascades]] - [[AI Post-Training]] - [[Composite Scoring]] - [[Confidence-Gated Routing]] - [[Conformal Prediction]] - [[Hierarchical Classification]] - [[Jev]] - [[Loop Engineering]] - [[Machine Native Intelligence]] - [[Mode Collapse]] - [[Model routing]] - [[Reinforcement Learning for Calibrated Decisions (RLCD)]] - [[Self-Consistency]] - [[Shadow Evaluation]] - [[System 1 and System 2 Thinking]] - [[System One Models]] - [[System One Primitives]] - [[Zero-Shot Classification]] <!-- SerializedQuery END --> ## Related - [[AI (MoC)]] - [[AI Design Patterns (MoC)]] - [[Large Language Models (LLMs)]]