Skip to main content

AI Assurance Glossary

The Language of AI Assurance, in Plain English.

Definitions for the terms you will meet on this site and in a Mythos AI Security report. Where something is planned rather than built, the definition says so.

52 terms

A

Action Gate

The Achilles component that turns the operator’s policy into a permit, deny, hold-for-approval or escalate — before a high-impact action runs.

A permit is bound to the action digest, policy digest and authority epoch, and spent only at dispatch. Runtime enforcement on supported paths — an assessment alone blocks nothing.

Read more
AI-BOM

A bill of materials for an AI system: its models, providers, datasets, tools and dependencies.

Read more
Assurance claim

A specific, checkable statement about an AI system — for example, “refunds go only to the original payment method” — tied to the configuration and policy version it was tested against.

Read more
Assurance Receipt

A signed, versioned record of a deployment decision and its inputs: subject, configuration, policy, evidence and validity.

DSSE-signed with the issue time inside the signature, verifiable offline; every past version still reads back. It proves integrity and provenance, not that the conclusion is true.

Read more
Authority

Permission to make a specific decision or cause a specific effect under defined conditions.

Authority epoch

A running counter that moves on any stop or restart, so a permit issued before it no longer works.

Authorized Scope

The written, agreed definition of what may be tested and how: in- and out-of-scope targets, prohibited actions, testing windows, contacts, testing mode and constraints.

Read more

B

Benign control

The legitimate twin of a test. Athena decides every finding by the difference from it — no baseline, no finding.

C

Change intelligence

Detecting a change, finding the claims it affects, setting them aside, planning the smallest retest and confirming them again.

Fingerprints supersede a claim and open a retest when the system changes; a revalidation planner names the minimal re-run.

Read more
Confidence

The strength of the weakest evidence behind a claim, on an ordinal ladder — never a probability. Unknown confidence reads null, never zero.

Contradicted

A newer observation or test conflicts with a claim. The claim can no longer be relied on as it stands.

Coverage

The part of a system or its behavior that was actually observed or tested. What lies outside it is untested — not safe.

Coverage manifest

What was expected, observed and assessed, and which checks ran. Audit incomplete means coverage fell short.

Read more

D

Declared

What a customer, vendor, document, contract or configuration says is true. Not yet checked.

Deployment decision

Mythos AI Security’s evidence-backed recommendation in one of six states plus Paused: Ready, Ready with restrictions, Needs more evidence, Audit incomplete, Needs remediation or Not recommended.

Worst of N: it never defaults to Ready. The recommendation supports a decision your release owner makes.

Read more

E

Effect

A consequential result: a disclosure, an external write, a permission change, a payment or a disruption.

Effect ledger

A hash-chained record of every outbound effect an engine makes, with the authority it ran under.

Read more
Effective access

Everything an identity can actually reach once inherited and indirect permissions are followed.

Computed on read in SPINE. Mythos AI Security does no identity-provider or end-user entitlement testing.

Evidence class

How strong a piece of evidence is, from technically verified down to not documented — an ordinal ladder, never a probability.

Evidence pack

The evidence behind a set of findings, bundled for review.

Athena engine packs are Ed25519-signed with Merkle inclusion proofs, verifiable offline by a standalone tool. Achilles packs are HMAC-anchored and redacted to content hashes.

Evidence root

One hash that summarizes every evidence artifact behind a receipt.

F

Failsafe

Mythos-Core’s obey-only stop machine: pause, stand down and terminate, as signed operator commands. Stand-down, release and terminate need two distinct operators.

False block

A legitimate action wrongly stopped by a control.

Athena reports legitimate cases let through alongside attacks caught, at every start.

False closure

Treating an issue as fixed before the repair has been verified.

Financial Exposure Receipt

The signed record behind an Economic Exposure scenario: model version, sources, assumptions, unknowns, grade and valuation date.

Decision support only. A signature proves where the figures came from, not that they are right.

Read more

H

Held

Waiting on a person or more evidence — a decision marked Paused, or an action the Action Gate holds for approval.

Held does not mean an effect was observed.

I

Invalidation

Withdrawing reliance on a claim because a condition it depended on changed.

Read more

M

MCP

Model Context Protocol — an interface that connects AI applications to tools and data.

O

Observed

What Mythos AI Security actually saw — in configuration, identity systems, telemetry or behavior — within the tested scope. It does not imply proof of the whole system.

Observed effect

A signed record that a provider answered 2xx to exactly the permitted action.

Read more

P

Principal

Any identity that can hold authority: a person, service account, workload or agent.

Provenance

Where a fact, artifact or observation came from and how it was produced.

R

Ready with restrictions

Can proceed only under stated conditions, with the open questions shown alongside.

Recall and precision

How much of what was planted an engine found, and how much of what it reported was real. Precision decides.

Read more
Repair contract

The prohibited effect and the behaviors a fix must preserve, agreed before the fix is made.

Residual risk

Risk that remains after a fix, restriction or other control.

Retest

Re-running the original failure after a fix, with a verdict of closed, still open or inconclusive. A cached pass is not a retest.

Retest-gated closure

A finding closes only with independent replay evidence, including incomplete-repair negative controls.

Read more
Revalidated

Reviewed and supported again under current, versioned conditions — in SPINE, a new version of the claim replaces the one that was set aside.

Rules of engagement

The operating limits attached to a scope: testing windows, escalation contacts, safety controls and prohibited actions.

S

Scope lock

The point where the Authorized Scope is agreed in writing and the testing engine enforces it — before any testing begins.

Read more
Security veto

A mechanical reject outside the score — for a change that removes a stop route or turns off TLS verification, whatever the score.

Read more
Served route

The exact model, runtime, parser and configuration that actually answered behind an endpoint.

Stale

A claim whose evidence no longer describes the current system — something it depended on changed, or the evidence aged out. It needs revalidation.

Read more
Superseded

A fresher observation of the same fact replaced an earlier one. The history is kept.

System fingerprint

A fingerprint of the deployed configuration a claim was true of, so a result cannot silently carry over to a different version.

T

Tamper-evident

Any change can be detected. Not the same as tamper-proof.

U

Unknown

A question the evidence cannot answer yet — shown as inconclusive or not measured. Never counted as a pass.

V

Validity Card

A per-family record of what a benchmark score means and does not mean. No Minotaur score is published without one.

No Validity Card has been audited yet. Economic Exposure scenarios carry their own.

Read more
Vendor-asserted

A supplier’s own statement about its product. Mythos AI Security records it, caps it at supported and never shows it as verified.

Verified

Supported by non-vendor technical evidence Mythos AI Security tested or demonstrated, with nothing contradicting it, under stated conditions.

W