Glossary
The core concepts from the Vita Indarra field manuals, in one place — each defined the way we'd say it across a table, not the way a textbook would.
Fluent is not the same as true. AI produces confident, expert-sounding language whether or not a word of it is true. Polish is the factory setting, not evidence. (From Don't Trust the Robot.)
Hallucination. When an AI states invented information as fact. Not a malfunction: a language model predicts the most plausible next words, and a plausible falsehood satisfies that goal exactly as well as the truth does.
Leverage on a leash. Use AI aggressively for everything it's genuinely good at — and keep a human hand on anything that touches a customer, money, or a promise. AI drafts; you approve. (From The Leverage and the Leash.)
Capability is not control. The spine of the whole catalog: you don't make a powerful system safe by making it more powerful, but by building a boundary outside it that it cannot reason, persuade, or be injected past.
The deterministic envelope. That boundary in practice — budgets, allowlists, rate limits, deny-by-default rules enforced by plain code outside the model, so even a fully compromised agent cannot exceed it. (From Building Reliable AI Agents.)
The sovereign leash. The strongest version: move the boundary into the money itself, with cryptographic controls that make overspending impossible rather than forbidden. (From Sovereign Machines.)
The verification bottleneck. As generation gets cheap, checking becomes the limit. How much autonomy you can grant an agent is set by how fast a human can verify what it's about to do — not by how smart it is. (From The Verification Bottleneck.)
The autonomy ceiling. That limit as a number: actions per day × honest seconds per review, against the reviewer attention you actually have.
The rubber stamp. What human review becomes at scale without information design — approving faster than honest reading allows. The more reliable the system looks, the worse the oversight gets.
Human-in-the-loop vs. human-on-the-loop. Approving every action vs. supervising by exception. The whole game is how the exceptions get chosen: by the trustworthiness of the verdict, not by how busy the queue is.
Detection is not calibration. An AI reviewer can notice a danger and still rate it "warn" instead of "stop." Noticing and grading are different abilities — so computable dangers belong to rules, never to model judgment.
The severity floor. Deterministic rules that compute an action's objective dangers from structure and state, returning fixed severities. Because the floor never reads prose, nothing written in the prose can talk it out of its verdict.
Gate vs. escalate. The doctrine adversarial testing forces on you: rule-backed or independently corroborated criticals can stop an action with no human; a critical resting on one AI reviewer's judgment must escalate — because anything that reads can be persuaded.
The receipt. A tamper-evident record of each oversight decision — action, flags with their backing, what the human saw, who decided, hash-chained to the record before it. Oversight an auditor can verify without trusting anyone's word.
The Five Sovereignties. What running AI on your own hardware buys you: privacy, control, permanence, introspection, and boundability. (From Private Intelligence.)
Sovereignty over secrecy. The real prize of local AI isn't that your data stays home — it's that you control the model, its lifecycle, and its limits. Privacy is the side effect; control is the point.
The glass box. A model you own, whose internal concepts you can read while it thinks — and whose spoken answers you can compare against its internal beliefs, to catch it lying from the inside. (From The Glass Box.)
Steering. Adjusting a model's internal activations to bias its behavior — a power only ownership grants. Revealing, useful, and advisory by nature: steering shapes behavior; it is never the safety layer.
World model. An AI that simulates an interactive world frame by frame in response to your actions — a game engine with no engine. Distinct from video generation, which produces a fixed clip you can only watch. (From The World in the Machine.)
Every definition above traces to a system that was actually built and tested — the method is on our standards page.
