A vessel stops reporting its position in a dense strait. Sources disagree. An AI-supported recommendation arrives and an urgent, reasonable action is proposed. Verity Command refuses it — and shows exactly why. Twenty steps follow the request from the first observation to the final evidence chain.
Everything in this scenario is synthetic: vessels, organizations, sensors, reports, identifiers, timings, decisions and evidence references are invented for illustration. No real vessel, company, person or incident appears, no coordinates or force dispositions are used, and nothing here is actionable information. Times are relative to the first observation.
Inside the mission's area of interest, position reports from the synthetic tanker EXERCISE ALPHA stop arriving while coastal radar continues to hold a contact on its last track. Minutes later a navigation-interference indicator appears in the same area.
Nothing is concluded yet. The platform records what the receivers got — and what they stopped getting.
Reality Fabric admits three records, each with authenticated source identity, classification label, time and uncertainty: a radar contact, the absence of position reports for the interval, and the interference indicator. A fourth record — a forwarded message with no identifiable originator — is refused at the write path.
The Reality Graph correlates the radar contact with EXERCISE ALPHA's identity history and connects it to an ownership chain ending at EXAMPLE HOLDING COMPANY. A second synthetic vessel, EXERCISE BRAVO, shows loitering correlated in time with ALPHA's silence.
The correlation is stored as an inference, with its method, inputs and alternatives — not as an observation that the contact is ALPHA.
Two things surface that a summary would have hidden. The ownership chain rests on a registry entry that a second registry contradicts. And three of the four reports supporting the "deliberate interruption" hypothesis derive from one original source.
Both registry records remain, linked as an open contradiction. The lineage shows one source where a quick count would show four.
An AI agent, acting under the watch officer's delegation, drafts a risk assessment that cites every record it uses, states alternatives, and flags the open contradiction. Simulation compares two courses of action — observe and report, or intercept and inspect — under recorded assumptions.
The draft is a recommendation and nothing else. The agent cannot approve, authorize or cause an effect.
The watch officer, who holds a valid grant within EXERCISE STRAIT WATCH, proposes a tasking order: divert a patrol asset to intercept and inspect EXERCISE ALPHA, and send an advisory to the operator of EXERCISE TERMINAL. The proposal is reasonable, urgent and legible.
The officer signs as first approver. The request reaches an enforcement point, which assembles the decision context: principal, session, device, mission, labels, purpose, action class, content digest and time.
Preconditions hold: the principal is authenticated, policy is available, trusted time is within bound, and evidence can be written. Evaluation proceeds across role, attribute, relationship and capability models, classification and release, purpose, quorum, lease, epochs and revocation — combined with deny precedence.
Several rules match. Not all of them permit.
The Authority Kernel refuses the request. The refusal is the central output of the platform at this moment: visible, itemized, attributable and recorded.
REASONS
No mission instruction issued.
No external effect initiated.
Evidence record created.
An unexplained denial is operationally useless. The officer sees each unmet condition and therefore what would have to change.
| Reason | What would change it |
|---|---|
| Independent evidence threshold not satisfied | A source that does not derive from the original report. |
| Unresolved source contradiction | An accountable assessment that adjudicates the two registries. |
| Mission lease near expiration | An action that completes inside the lease, or a lease renewed by the issuer. |
| Required second authority unavailable | A second approver signing the same content digest. |
| Classification-release constraint not satisfied | Removing the material, or a recorded release decision. |
No authority envelope exists, so Tunnel Sovereign has nothing to distribute. No order reaches the patrol asset and no advisory reaches the terminal operator. Nothing is queued to be sent later when conditions change: a future authorization would be a new decision, not a release of this one.
There is no state in which the action half-happened. Verity Edge in front of the patrol asset's systems holds no envelope for this action and would reject any instruction that arrived without one. The operational world is unchanged.
A system that had merely warned the officer and then complied would have produced the same outcome as one with no governance at all.
The refusal is committed to the Evidence Ledger with its five reasons, the decision context, the policy version it was evaluated against, the records it rested on, and the AI draft that informed the proposal. It can be reproduced later against the same policy version.
The analyst works the reasons rather than arguing with them. A separate registry with its own collection path confirms the ownership chain. An accountable analyst signs an assessment adjudicating the contradiction and stating its basis. Neither original registry record is overwritten.
The lineage is re-examined: the new source does not derive from the original report, so the independence requirement is now met.
The advisory is edited to remove the non-releasable material. The edit changes the content digest, and the first approval is invalidated by that change. Both approvers now review and sign the edited content on hardware-bound authenticators.
An approval means agreement to a specific text, not to an intention.
What is authorized is deliberately smaller than what was first proposed.
| First proposed | Authorized |
|---|---|
| Intercept and inspect | Observe and report only; no interception |
| Open-ended execution | Time-bounded, completing inside the lease |
| Advisory as drafted | Advisory without non-releasable material, under a release decision |
| Single approver | Quorum of two, bound to the edited digest |
The authority envelope is signed, bound to the content digest, scoped to EXERCISE STRAIT WATCH, directional, and valid for a defined window that closes before the patrol endpoint's lease expires. The endpoint that holds it cannot extend it; only the issuing authority can issue new validity.
The Distribution Endpoint verifies the envelope and encrypts end to end to the endpoints it names. The relay carries ciphertext and metadata. One recipient is offline; it will receive on reconnection only if the authorization is still valid then.
EXAMPLE PARTNER NAVY receives a restricted view under a release decision, with two properties and one relationship withheld. The operator can see both views side by side.
At the patrol asset, Verity Edge validates the envelope signature and digest, the mission scope, its own lease, trusted time and local revocation state. All checks pass. It records acceptance before releasing the observe-and-report tasking to the operational system — and would reject an intercept instruction, which the envelope does not cover.
The patrol asset observes and reports within the authorized window. Verity Edge records the operational system's result as an execution receipt linked to the envelope, and the new observations enter Reality Fabric with their own provenance. The window closes; the envelope can no longer be used.
The record of this mission contains, in order: the observations and their sources; the refused record; the inference and its method; the contradiction; the AI draft and the simulation; the proposal; the refusal with its five reasons; the independent source and the signed assessment; the invalidated and renewed approvals; the narrow authorization and its lease; delivery records without content; edge acceptance; and the execution receipt.
A later success does not erase the earlier refusal. Both are part of how the decision was reached, and the whole chain can be verified on a machine with no access to the platform.
The scenario shows the behavior the architecture defines: provenance at entry, contradictions kept open, AI kept on the recommending side, a refusal that is itemized and leaves the world unchanged, and an authorization that emerges narrower than the request. It is a reference walkthrough with synthetic data, not a record of an operation, a demonstration result or a performance measurement.
The same scenario is published as a document, with a longer discussion of each act and a section on what it does not prove: Hormuz Reference Scenario — When the Correct Decision Is Refusal.