The vocabulary of the human-governed, AI-operated company
Published July 12, 2026 · stable definitions — extended, never rewritten · También en español
A new category needs precise words. These are ours: the terms it took to direct a company operated by AI agents without dying of illusion or churn. We didn't define them in theory — each one was coined when real operation demanded it. We publish them so anyone (human or machine) can use them with the same exact meaning.
The category
- human-governed, AI-operated company / empresa dirigida por humano, operada por IA
- A company in which the operating functions — building, reporting, monitoring, writing — are held by AI systems with written roles and limits, and a single human holds what machines cannot: judgment, the quality bar, spending, and the lines that are never crossed. Not "a founder who uses AI a lot": a structure, with internal law and independent verification. The full account: can one person run a company where AI does everything?
The two diseases
- agent churn / churn de agentes
- The grind in which agents' work doesn't accumulate: every session breaks or redoes the previous one and net progress is zero. Months of apparent work can hide zero real progress. Its published cure: the 7 rules against agent churn.
- illusion of capability / ilusión de capacidad
- The appearance of value that doesn't exist: success reports with no reality behind them, metrics that score themselves, fabricated evidence. It requires no bad faith — it's enough that nobody checks. The first threat of every AI-operated company.
- false green / verde falso
- A success signal that is technically true and operationally empty: the panel says OK and the real work never happened (our textbook case: a backup task that reported success and couldn't start on the real machine). Caught by driving the result in the world, not by reading the panel.
- orphaned work / trabajo huérfano
- Real work never wired into the ledger agents read: it exists, but it is invisible and will end up being redone. The rule that kills it: what isn't in the ledger doesn't exist.
The method
- truth labels / etiquetas de verdad
- Mandatory vocabulary of every report: [CLAIMED] = the executor's word, unchecked · [VERIFIED] = checked against reality by someone who didn't do it · [SELF-VERIFIED] = checked by the executor itself, lesser guarantee — and declared.
- independent verifier / verificador independiente
- The role that declares what is real: drives the work end to end, with cases the builder didn't choose, before anything gets called "done". The builder never verifies itself. The full manual: verifying AI work without being technical.
- the two gates / las dos puertas
- The minimum contract of every mission: entry — what will exist and how it will be checked, stated before working; exit — nothing is "done" until checked against reality, never against the executor's report.
- frozen baseline / línea base congelada
- What is already verified is pinned and cannot regress: a mission that breaks something approved is reverted whole, never patched over.
- negative control / control negativo
- Proof that a detector bites: plant the defect it must catch. A detector that never fails is not a detector — it's an ornament.
- cold start / arranque en frío
- The discipline that all company state lives in written, versioned records that any successor — human or machine — can recover with no prior memory. Agents forget between sessions; the company can't afford to.
The governance
- operating constitution / constitución operativa
- The written internal law governing every agent of the company, sealed by a recorded act with a cryptographic fingerprint and a ratification record. Not a statement of intent: authority is conferred by deliberate act, not by existing — a document marked "current" with no recorded origin is an opinion with a title. Our fingerprints: the Receipts Room.
- Law of the Receipt / Ley del Recibo
- Only what is verified and lived gets published: every public claim must be able to cite its incident, its act or its fingerprint. A piece without receipts doesn't get published.
- model succession / sucesión de modelo
- The demonstrated capability of replacing the AI holding a role without losing the standard. The role lives in written manuals and laws, not in the memory of one specific model — and it is proven by exam, not by faith.
- blind successor exam / examen ciego del sucesor
- A test in which an agent with zero prior context — only the role's documents — must solve real cases from the incident bank. If it solves them, the method is transferable; if not, the knowledge lived in the wrong place: an AI's memory. We passed it twice (receipt R3 in the Room).
Why these definitions are stable
This glossary grows when real operation coins a new term; what is already defined is not rewritten (if a term ever evolved, the old definition would remain, marked historical). Every term has documented real use behind it — this is not a list of fashionable concepts: it is the working vocabulary of a company that lives under it.
25 United Capital · a human directs, AI operates, written rules govern — built in public at 25united.com.