Concepts
Concepts for AI architecture, reliability, economics, and operating models.
Arizen concepts name recurring mechanisms in agent architecture, reliability, model economics, and decision rights.
Use the map, then follow each concept to its defining essay.
Choose the Decision
| If You Want To | Start Here |
|---|---|
| Bound agent behavior | Read Stochastic Gap, Probabilistic State Machine, and Agentic Contract. |
| Locate reliability bottlenecks | Read Validator Asymmetry, Golden Dataset, and Conservation of Chaos. |
| Understand model economics | Read Intelligence Arbitrage, Iron Triangle, and Utilization Paradox. |
| Design AI operating models | Read Judgment Compression, Human-in-the-Loop, and Jurisdictional Voids. |
Agent Architecture
How stochastic behavior becomes bounded, specified, and recoverable.
- Stochastic Gap - the boundary between deterministic software and probabilistic behavior.
- Probabilistic State Machine - a runtime architecture for routing model failure through explicit states.
- Agentic Surface Area - the portion of a workflow exposed to autonomous model action and control risk.
- Agentic Contract - typed behavior boundaries for one agent capability.
- Durable Execution - agent work designed to survive crashes, retries, and runtime interruption.
- Context Propagation Failure - why useful information degrades at agent and context boundaries.
Reliability and Evaluation
Where reliability effort pays, and where failure hides after components look correct.
- Validator Asymmetry Principle - why validation quality can buy reliability more cheaply than generation quality.
- Golden Dataset - evaluation data as the asset that survives model, prompt, and provider changes.
- Conservation of Chaos - constraints move uncertainty unless the system routes it deliberately.
- Jurisdictional Voids - handoff failures that no component, metric, or team owns.
Economics of Intelligence
How model cost, judgment, scarcity, and information compression change what is rational to build.
- Intelligence Arbitrage - routing work to the cheapest sufficient intelligence tier.
- Iron Triangle of Inference - the tradeoff between model quality, latency, and cost.
- Utilization Paradox - why a larger model can be cheaper when hardware utilization changes the denominator.
- Entropy Arbitrage - turning noisy, high-entropy input into compressed, actionable output.
- Problem Scarcity - why problem formulation becomes scarce when implementation becomes cheap.
Organizational Intelligence
How AI changes judgment, decision rights, ownership, and what organizations can see and do.
- Judgment Compression - expert decisions encoded into systems, including who owns and reviews them.
- Human-in-the-Loop - human review designed as decision rights, capacity, escalation, and appeal rather than a safety checkbox.
- Jurisdictional Voids - handoff failures that no person, team, component, or metric owns.
- Semantic Compression - meaning per word as a function of shared context and trust.
- Epistemic Decay - technical knowledge losing validity as models, tools, costs, and constraints change.
- Attentional Gravity - visible metrics pulling optimization effort away from harder truths.
- Expert's Shadow - strong priors making adjacent possibilities harder to see.
- Map of Ignorance - an explicit inventory of unresolved questions, assumptions, and experiments.
From Concept to Implementation
When a concept becomes an implementation choice, move to Patterns. For a longer synthesis, see Books.