I design and build systems with AI for a living, and I write about them. The writing has accumulated a set of named concepts. Naming them was deliberate: a concept with a name can be reused and sharpened, while a concept described anonymously has to be re-explained in every piece. This piece collects the vocabulary in one place, with each term pinned to one definition, because a term that spreads without its definition drifts until it means nothing. Martin Fowler calls that failure semantic diffusion, and his bliki, a corpus of short entries, each named and attributed, is the prior art for publishing a list like this at all. Lexikon is Greek for dictionary, which is what this page is.
A few of these names are already in print: the enterprise brain has its own piece, and koine arrived with my writing on the knowledge system behind this work, which also named intent-readiness separation. Most are named in public here for the first time, which is deliberate. This list is the planting.
Two honesty rules govern every entry. The name is mine and the underlying ideas almost never are, so each entry says what it stands on. And one term carries one meaning: the definition given here is the definition, and everywhere else I use the term I mean exactly this.
The core concepts
The deterministic spine
Any claim an AI system acts on or presents as true (a balance, what a learner has mastered, whether two records reconcile) is computed by rules over observable events, never by a model’s guess. The model is kept out of the truth path entirely. This is what makes an agentic system auditable and safe in a regulated organisation, and it is why the system’s answers do not change when the model version does. It stands on event sourcing, pointed at a new problem.
The enterprise brain
The part of an organisation’s AI operating layer that remembers and reasons: the governed memory a workforce of agents and people read from and write back to. Every entry carries provenance and a human gates what gets in. A contradicting entry never overwrites the old one; it is surfaced for a ruling. It stands on two decades of business capability modelling for its decomposition, and on the personal knowledge-management tradition for the smaller-scale intuition.
The experience flywheel
The design rule that an agentic system must get better from being used: every interaction is captured, and what proves durable is promoted, through a gate, into the system’s memory or its deterministic spine. Name the capture-and-promote loop at design time or you have built a tool and called it an agent. The flywheel image is Jim Collins by way of Amazon; the gated promotion into the deterministic spine is the contribution.
The fractal loop
One loop shape (learn what is needed, give it form, build and prove it, feed what the building discovers back) repeating at every scale of work: a working session, a product, a portfolio, an organisation. The self-similarity is what lets learning compound across scales, and why one person with agents and an organisation of thousands run the same design. It stands on Mandelbrot for the image and on Stafford Beer’s viable system model for the organisational prior art.
The software factory
The durable system around the model: the skills, reviewers, gates, memory, and tracked work that turn a general model into repeatable, verifiable delivery. The model is the engine; the factory is everything that makes the engine’s output trustworthy, and the factory is the asset that compounds. The term itself is old (Japanese software factories in the 1970s, Microsoft’s Software Factories in the 2000s); what I claim is the AI-native, single-author instance and the argument that this is where individual leverage now lives.
The applied-AI-systems thesis
The position underneath all of the above: the value of AI in an enterprise sits in the operating layer around the model and the data the organisation holds about its customers and operations, not in the model. Concretely, a two-tier shape: a frontier model orchestrates and specialist subagents execute bounded tasks, while a deterministic spine holds the truth. The industry keeps rediscovering pieces of this under feature names; the thesis names the whole shape.
The operating discipline
Fog, and sharpening
Most work does not arrive buildable; it arrives as fog, a direction worth holding that nobody could act on yet. The discipline is to hold fog in a ledger of its own, out of the work queue, each entry carrying the condition that will make it worth looking at again. Nothing is dispatched until it has sharpened into a question concrete enough that an agent can be sent to answer it. Sharpening is the intake gate of the software factory: the tracker takes only sharp work and the build system runs only from the tracker, so the factory never burns effort on fog. The fog image is borrowed from the fog of war; the ledger with its surface-when conditions, and the dispatch gate on it, are what I claim.
Koine, and the Koine Loop
Koine was the shared Greek that let the Hellenistic world coordinate after Alexander; a team’s koine is the shared language its members speak when they work together, and Koine-OS is the pack that ships that language as a working method a team can install. The Koine Loop is its operating rhythm: prompt, create, review, adjust, and loop until clean, with the reviewer never being the generator and every iteration landing as a durable file.
Steward-as-orchestrator
What happens to the governance role when agents act: the data steward stops being the back-office fixer and becomes the person who writes the rules the agent runs against. Governance moves to the front office because the rules now execute.
Three-role separation
The build discipline that no agent judges its own work: a coordinator authors the judgment and never builds, a fresh builder authors the change, and a blind auditor verifies the artefacts cold, without the builder’s narration. It stands on the oldest idea in quality control, independent inspection, applied to agents because agents flatter their own output.
The blind audit
The verification half of the discipline above, worth its own name for one sentence: green gates verify your tests; the blind audit verifies your feature. An auditor that has read the builder’s story checks the story, and an auditor that has not checks the work.
Measure the judge
If a model gates anything (a review verdict, a promotion into memory), the gate’s own accuracy is a measured number against a ground-truth set, never an assumed virtue. Bounded autonomy is bounded by measurement, and an unmeasured judge is just a second opinion.
Intent-readiness separation
Every decision (intent) has a corresponding operational state (readiness), and conflating them in one artefact is how architecture and data catalogues lose their delivery teams’ trust. Keep a decision log and a readiness tracker as separate artefacts, and derive the gap report as a projection that is reconciled on a ritual cadence.
Forcing functions
A forcing function converts a should into a must: it makes the right action the path of least resistance or makes the wrong state visibly fail, so the right thing happens without anyone remembering to do it. The term is Don Norman’s from interaction design; what I claim is the taxonomy for AI-assisted work, five shapes. The gate blocks progress until a condition holds; the ritual recurs on a clock nobody can skip; the pairing ties two artefacts together so neither ships alone; the default projection makes the right state the automatic one; the tripwire makes a silent failure announce itself. A rule without a forcing function is decoration.
This list will grow. When a concept earns a name, its definition lands here, and future pieces will link back to it.
PS, for anyone interested in the machinery behind the writing: the working seats in my own system carry Greek names. Chiron builds, Kritos reviews, Mnemon remembers, Epistates coordinates, Hestia keeps the operational surface running. The names do real work, because a named seat keeps an agent inside its role. They are pet names though; the vocabulary above travels without them. So far only koine, and lately the tracker pinax, have graduated to public life.