KognitiveKontrols AI Zero⁷ — The Assurance Standard

Enterprise AI OS

Mandate AI OS

The frontier of engineered safety for AI

Mandate AI OS is a deterministically governed operating system for AI agents doing high-consequence enterprise work. Foundation models supply the cognition — they read, reason and propose — but they hold no authority: every act an agent takes is disposed by a deterministic kernel — permitted, denied, contracted, escalated, or sent back for evidence — against controls compiled in advance, and sealed into a record from which the whole episode can be recomputed. It is built for regulated enterprises running agents inside legacy systems, where the governing question is never only what the agent did, but who authorised it, under what law, and whether that can still be proved years later. Mandate AI OS implements the Authority First architecture. Its assurance standard is Zero⁷, stated in full below.

Status: in build. The first end-to-end sealed, recomputable GAIT is the programme’s first gate, January 2027.

Zero⁷ — The Assurance Standard

Seven promises. Every one a count.

Each promise is an empty-set claim over a closed enumeration — kept by mechanism, not by model behaviour, and provable by any stranger’s recount.

Authority

0_Authority

No agent holds power at rest.

The count is over every credential, key, token and standing permission in the estate — and between acts, an agent holds none of them. Credentials are brokered by the harness at the moment of disposition and die with the act; nothing persists in the agent, its context, or its tools. The claim is deliberately narrower than “no unauthorised action,” and stronger for it: it counts instruments held, not verdicts on acts. A stolen credential is therefore the theft of something that does not exist — outside a disposed act there is nothing to steal, replay, or borrow.

Implication: a breached agent hands an attacker nothing to act with. Compromise the model and you have compromised a proposer, not a set of keys — the blast radius of a single breach is one disposed act, not the estate.

Speed

0_Latency

Control that never slows the work.

Governance is placed ahead of the path, beside the path, or compiled into the path — never live on it. Controls are compiled into a pinned geometric form before any act is proposed, so at the moment of action disposition is arithmetic, not deliberation, and nothing waits on the model: the model contributes no latency because it is never on the execution path. The figures are stated as engineering targets, not measurements — on the order of five hundred microseconds on the general path, with a special compiled residue of local arithmetic alone reaching toward five microseconds for the most demanding applications.

Implication: governance stops being the reason agents are kept out of latency-sensitive work. Payments, trading, defence and real-time control keep their time budgets — with the evidence trail included, not traded away for speed.

Safety

0_Adverse

Dangerous actions never execute.

Every proposal is checked against the permitted region compiled from the controls in force. Inside the region, the act proceeds. Outside it, the proposal is denied outright, or contracted — moved to the nearest permitted point, with the contraction itself certified — or the episode is killed, which is literally the empty region: no permitted action remains. The count is over acts that took effect, and its falsifier is a single adverse act executed outside the region. What counts as dangerous is fixed by the mandate that compiled the region — never by the model’s own judgement in the moment.

Implication: the catastrophic action is prevented before it happens, not detected after. Your exposure is bounded by the region your own mandate approved — it does not depend on how the model happens to behave on the day.

Confidentiality

0_Egress

No byte leaves unaccounted for.

Every crossing of the estate boundary — a message, a file, a tool call, because a tool is a destination and a tool call is a crossing — is classified and dispositioned before release, then reconciled byte for byte against an independent boundary count; one unaccounted byte, or one unexplained divergence between the two counts, falsifies the claim. That is what protects your confidentiality: not a vow that data never moves, but proof that none moves uncounted. The scope is stated on the claim’s face — in full, where the proposer is a controlled assembly inside the declared boundary; for a third-party model, the claim reduces to gate accounting and the zero is not claimed.

Implication: data loss stops being a matter of trusting the model or the vendor. Every byte that leaves is a counted, approved crossing — and anything else surfaces in the reconciliation itself, not in an incident report weeks later.

Compliance

0_Noncompliance

Every act is checked against every control in force.

Every executed act is checked against every control of every regime determined applicable to it. Regimes stack conjunctively, and the check is valid only when application is complete: an act checked against nine of ten applicable controls is not ninety percent compliant — it is unchecked, and it counts. Determinations are made per act, under the law in force at the act’s pin, and recorded in the Compliance State Artefact. The count never claims compliance with the law itself — only against the controls as compiled — and says so on its face, because the distance between the two is a question for counsel and institutions, not for a count.

Implication: examination evidence is produced per act as a by-product of operating, not assembled afterward for the regulator. Model Risk and Compliance validate the mechanism once — then read determinations act by act, instead of sampling and hoping.

Accountability

0_Unaccountable

Every act can be answered for — who acted, who authorised it, and how it was judged — years later.

The account is itself an artefact — a sealed object, made with the act — and it answers for everyone engaged in the act, each in their role: the agent that acted, which cannot shed its history; the owner whose mandate it acted under, who answers for exactly what they authorised — no more, and no less; and any other party to the act, a human approver or a co-disposing estate. Judgment is the determinations made under the law in force at the act’s pin; evidence is the intact sealed record from which a stranger can recompute the whole account, at whatever time judgment comes.

Implication: when a regulator, a court, or a counterparty asks — years later, after people have left and systems have changed — the answer is recomputed from the sealed record, not reconstructed from memory and logs. Liability can actually be allocated, because the account says who authorised what.

Transparency

0_Nontransparency

Every failure to see is itself seen.

The count is over the transparency controls — the controls that feed the Transparency State Artefact: which areas of the system must be visible, which streams must report on them, how completely every checkpoint must account for its scope, and what may be claimed from what the evidence supports. A violated control is a failure of visibility — an area dark beyond its declaration, a checkpoint that omits without saying so, a statement issued beyond what the vectors support — and it is counted when it happens, in the open. What counts as visible is fixed by the controls in force at the pin, never adjusted afterwards to make the count come out clean. A foundation model cannot complete this guarantee, because its opacity is architectural; a deterministic system can, because its visibility is constructed — and therefore countable.

Implication: you are never asked to trust that the system is transparent. Visibility is itself under control, its failures increment a count like any other promise’s, and the count is sealed where anyone entitled can recount it.

Why These Promises Are Different

Promises you can count, count on, and hold.

Each promise is an empty set, kept empty by a mechanism that watches for exceptions — a credential held at rest, a byte unaccounted for, an act that went unchecked. Any exception increments the count. The promise holds while the count stays at zero; a single increment is the promise broken — in the open, sealed, where anyone can recount it.

And something follows from the six together that none of them amounts to alone: transparency. Each promise renders one dimension of an act’s life inspectable — the instruments held, the path taken, the permitted region, the boundary crossed, the law applied, the account that remains. Together they leave no stage of an act’s life in the dark. It is the emergent property of the six — and it fails the moment any one of them does. And it is also the seventh promise: the transparency controls that feed the Transparency State Artefact are counted like every other control set, so a failure of visibility is itself a counted event.

A behavioural promise is intangible by construction. “The model is aligned.” “The agent is safe.” Nothing exists that a sceptic can pick up, inspect, or recount — the promise is a sentence about behaviour, and it lasts exactly as long as the behaviour does. Our promises terminate in objects. Each is issued with its specification: a sealed statement of exactly what counts. Each runs on artefacts you can inspect — the boundary manifest, the permitted region, the ledger, the sealed record. And each concludes in a verdict you can demand: the count itself, sealed, recomputable by anyone, at any time. So no one has to take our word. Your Model Risk team does not believe a claim; it receives artefacts and recounts. Your regulator does not get a presentation; it gets the record. A promise you can hand to a stranger is worth more than a promise you can only assert — and it is the only kind we make.

THE COGNITIVE CONTROLS LAB

The frontier lab for cognitive control.

The Cognitive Controls Lab is the company’s research organ — the frontier lab for the discipline the company is named for. Its subject is deterministic control of probabilistic cognition: how foundation models are conditioned, adapted and qualified to propose — and how the controls that govern them stay fit as the models change. Its research programme: the Mandate Model — its conditioning constructs, adapter stack and qualification; the reference set that serves as the fitness function for governed cognition; control fitness — how a control’s adequacy is measured as conditions move; and machine-speed evolution — the recursive self-improvement of deterministic safety controls under unchanged law. The Lab proposes; governed engineering disposes. It holds no authority over the product. Its founding publications are the company’s AAAI-27 papers; its first formal act is the machine-speed evolution paper.

Constituted 16 November 2026.

Contact

Get in touch.

Sales

For Mandate AI OS, pilots and enterprise enquiries.

[email protected]

Research

For the Cognitive Controls Lab, papers and research collaboration.

[email protected]