AI Substrate Glossary: Definitions for the Terms KoldOps Uses

This page is the definitional reference for the terms KoldOps uses in its writing about AI substrate, storage, and context infrastructure. Each entry is short on purpose. Each entry links to the longer piece that defines the term in depth, where one exists.

Living document. Refreshed quarterly. Last reviewed May 29, 2026.

AI Substrate

The operating discipline that keeps a business's decision-state airlocked to its codebase's state. Same version control, same review gates, same diffs, same audit. The substrate is not the model, the vector store, or the compute layer. It is the underlying record that the AI reads from and writes to, managed with software-grade rigor. Full definition: Decision-State, Airlocked to Code-State.

Storage for AI

The persistent layer an AI agent reads from and writes to. Plays the role for AI workloads that a database plays for an app or that S3 plays for objects. Five required properties: versioned, reviewable, retrieval-native, replayable, LLM-native. A vector database is an index, not a storage layer, and satisfies at most 1 of the 5. Full argument: Vector DBs Aren't Storage. They're Indexes.

Context Engineering

The engineering discipline of gathering, ranking, and delivering context to an AI agent at request time. Distinct from prompt engineering (which is single-turn copywriting) and from retrieval-augmented generation (which is one technique within the discipline). Most production AI systems require a context engine, whether the team building it uses the term or not. Pillar piece forthcoming.

Decision-State

The current and historical record of a business's binding decisions. Vendor selections. Routing standards. QC thresholds. Hiring rubrics. Pricing rules. The questions code-state answers (what version is in production, who changed it last, why) applied to business decisions. In most businesses, decision-state is unmanaged: it lives in PDFs, decks, Slack threads, and the heads of senior employees.

Code-State

The current and historical record of a software system, managed by git, pull requests, CI, tags, and a deployment process. Every change has a named author, a timestamp, a review, and a diff. Code-state is the solved problem that decision-state is supposed to imitate when you build an AI substrate.

Decision-Code Airlock

The fusion of decision-state and code-state into one reviewed, versioned, diffable record. An airlock is a mechanical seal between two pressure systems; a change on one side propagates and is acknowledged on the other before either continues to operate. The airlock is what an AI substrate enforces. Without it, the AI reads stale folklore.

Intent Layer

The record of what the business intended, kept in machine-readable form, version-controlled, and queried by AI agents at request time. The intent layer sits inside the substrate. It is what the AI consults when asked "what is our policy on X." Distinct from the operations layer, which is what the business actually does. Drift between the two is the failure mode the substrate is designed to detect.

Drift Detection

The process, manual or automated, that flags when business operations have diverged from the recorded decisions in the substrate. When the shop floor stops following the routing standard, when the AR team waives the credit hold, when the QC team passes a part that the threshold rule says fails. Drift detection is the substrate's immune system. Without it, the substrate and reality silently desync until an audit, a hire, or a lawsuit reveals the gap.

Substrate Audit

The 5-question diagnostic for whether a business has an AI substrate. A new hire asking about a policy finds a single authoritative document (yes/no). That document has a version history with named authors (yes/no). Changes pass through review before taking effect (yes/no). An LLM can read every relevant decision document without a vendor portal (yes/no). A process detects drift between recorded decisions and operations (yes/no). Score 0 or 1 on each. Most businesses score 0 to 1 of 5. Full audit and scoring: Decision-State, Airlocked to Code-State.

Where these terms are used

KoldOps writes about AI substrate, storage, and context infrastructure for physical-industry businesses: manufacturing, construction, energy. Knoxville, Tennessee. The terms above are used consistently across blog posts, product pages, and customer engagements. When a term needs a longer treatment than this glossary provides, a dedicated pillar piece is the home for the definition.

See also: all posts, Business System Review, on-premise AI.