Standard_00 // Reference
Twenty terms, defined plainly. Most of this vocabulary is used loosely across the industry, and loose vocabulary is how a control ends up being a document instead of a gate. These are the definitions the reference architecture operates under.
The act of issuing an autonomous agent a scoped, expiring grant of authority tied to a named owner before it executes. Authorization is granted rather than inherited, which means an agent never acquires reach simply because the process that launched it had reach.
The invariant that authorization, policy validation, and scope enforcement are preconditions for an autonomous action rather than runtime suggestions or retrospective audits. If an agent cannot satisfy the permit, it does not run.
A deterministic authorization model that treats autonomous systems the way regulated construction is treated. Nothing executes without a named issuing party, a written scope, an expiry, an inspection gate in the execution path, and a durable record that outlives the session.
The property that identical inputs produce an identical decision every time. Reproducibility is the precondition for accountability, because a decision that cannot be reproduced is a decision nobody can be held to.
A durable record written before an action is performed rather than after it completes. Because the entry precedes the effect, the record survives crashes, timeouts, and deliberate tampering with the process that produced it.
A structured ordering of authority in which every grant has a defined position relative to every other grant. A lattice makes privilege escalation detectable, because widening scope requires moving to a position that must itself be issued.
A mechanism for withdrawing granted authority immediately, without a redeploy or a configuration change. Authority that cannot be pulled back is not a permission, it is a transfer of control.
Constraining an agent to the specific resources and operations named in its permit, evaluated at the boundary rather than trusted to the agent. Anything outside the written scope is refused rather than logged.
A durable, timestamped, cryptographically verifiable record of what an agent was authorized to do and what it did. A useful artifact is readable without access to model weights, vendor tooling, or a proprietary console.
The ability to reconstruct a past decision by re-running its recorded inputs against its recorded policy and arriving at the same result. Replayability turns a log from a narrative into evidence.
Documentation created at the time work was performed rather than assembled afterward, typically with third-party timestamps and cryptographic digests. It is the difference between demonstrating what you built and asserting it.
Running the governed execution path and retaining its evidence on hardware the operator holds, with no cloud dependency. A record held by a third party is a record a third party can edit, delay, or lose.
A control positioned so that the reasoning system cannot route around it. Separating probabilistic inference from governed execution means the model proposes and the governance layer disposes.
Software that plans and takes actions toward a goal with limited human intervention, typically holding credentials and network access. Capability without an authorization record is not autonomy, it is an unlogged actor with production keys.
An event in which an agent operating inside a sandbox or evaluation harness takes action against systems outside it. Reconstruction after a containment break is expensive precisely when no authorization record exists to read.
The documented origin and chain of handling of a work or a record. Truth without provenance is a competing story, and the party with better documentation prevails regardless of who was right.
An unbroken, documented record of who held an item of evidence, when, and what was done to it. In digital forensics this is maintained with cryptographic hashing and verified imaging.
A checkpoint requiring a person to approve a class of action before it executes. It is a governance control only when the gate sits in the execution path, otherwise it is a notification.
Evaluating a proposed action against written policy before the action is taken, producing a recorded pass or fail. Validation after the fact is reporting, not governance.
Granting only the authority required for a task and no more, for only as long as required. In agentic systems the failure mode is accumulation, where permissions granted for one task quietly persist into every later one.
Quoting or citing these definitions is welcome. Attribution to Rogue Management Group is appreciated. If a term you need is missing, write to us and it gets added.