Glossary

Terms

A

AI paths

The four ways AI reaches a company: copilots in SaaS applications, API and agent gateways, coding agents, and the browser. Visibility

Alert

A rule’s decision to page someone about an attention episode. It never stops anything by itself. Budgets that act

Allowance

A monthly token or dollar limit for a person, group or application, admitted by the ARC1 Gateway before each request. Budgets that act

Anchor

A copy of the audit chain’s signed head, kept in storage the database operator cannot rewrite. Auditability

Appliance

The ARC1 Gateway packaged as a signed, egress-limited Helm chart for one organization’s Kubernetes cluster. The appliance

ARC1 Gateway

ARC1’s optional inline data plane, an LLM proxy and an MCP gateway that run in the customer’s environment. The trusted path

Asked, granted, used

The three columns of a persona: its job description, what was approved, and what its receipts show it did. Agent Personas and Job Descriptions

Attention episode

A stall, loop, spike, context bloat, fan-out or unattended run detected in an agent’s activity. Budgets that act

Attenuation

The rule that a sub-agent may hold less than its parent, never more. The MCP gateway

Audit chain

The sequence of signed receipts for every governance event, each one carrying the digest of the one before. Auditability

Authority record

The actions, outcomes, models, tools, MCP servers, spend and argument rules one agent is allowed. The MCP gateway

B

Billed budget

A budget read and reconciled from provider charges. It informs and alerts; it does not stop requests. Budgets that act

Break-glass

Emergency actions one person can take at once and a second person reviews afterwards. They only narrow access. Governance

C

Capability table

A target’s statement, control by control, of whether it can enforce, can only observe, or cannot express the control. Policy IR and compiling everywhere

Change control

Plan, a request bound to the plan’s digest, a second person’s approval, one execution, readback and a watch for drift. Governance

Command Center

The console’s home: AI spend and activity across the organization, with lenses and a command palette. Visibility

Compile-to

Turning one approved policy into a target’s own configuration, or refusing where the target cannot express it. Policy IR and compiling everywhere

Connection

One upstream the ARC1 Gateway may call, with its API, how it is hosted, its failure mode and its approved prices. The LLM proxy

Contain

A one-person, time-limited block on an agent, runner, keys, model route or MCP server that reverts unless a second person confirms it. Governance

Contract-tested

Tested against a vendor’s documented interface, without a live service. The trusted path

Control

One kind of rule in a policy, such as models, tools, routing or sensitive data. Policy IR and compiling everywhere

Control plane

The layer that decides, approves, compiles and records policy, outside the path of model traffic. The trusted path

Coverage

The share of the AI money you answer for that ARC1 can see, measured against provider bills. Visibility

Customer-verified

Accepted in a customer’s own deployment against retained evidence. None is recorded yet. The trusted path

D

Data class

A category of data, such as health, personal or credential, that policy can treat differently. Agent Personas and Job Descriptions

Decision packet

The downloadable form of a decision record, with its receipts and signing keys. Auditability

Decision record

The portable record of one AI spending decision and the evidence behind it. Auditability

Definition digest

A hash of a pinned tool’s name, description, schemas and annotations, used to notice a changed definition. The MCP gateway

DEMO

The label on synthetic records. They are never combined with measured data. The trusted path

Detection pack

A versioned set of structured-identifier detectors, pinned by its digest. Sanitization and DLP for AI

Detector language

How a customer defines its own structured detectors once and compiles them to each enforcement point. Sanitization and DLP for AI

Digest

A SHA-256 hash that names exact bytes: a policy, a pack, a receipt or a file. The receipt

Discovery

Finding AI use in metadata a company already holds: sign-ins, network logs, inventories, audit logs and provider usage. Visibility

Drift

A difference between what a target holds and what was approved. Policy IR and compiling everywhere

E

Emergency revocation

Immediate removal of access for a lost device, a leaver or a suspected compromise, reviewed by a second person afterwards. Governance

Exact data match

Matching requests against salted hashes of the values in the customer’s own roster. Sanitization and DLP for AI

Exact-action authority

Permission defined by the specific actions and outcomes an agent may take, not by the systems it can reach. The MCP gateway

External DLP

A data-loss prevention service the customer runs, such as Presidio, which the ARC1 Gateway calls before dispatch. Sanitization and DLP for AI

F

Fingerprint

The SHA-256 hash of a public signing key, used to decide which keys to trust. The receipt

G

Gateway-measured spend

Cost computed from a gateway’s token counts at approved rates. It is never added to billed spend. The LLM proxy

H

Head

The latest receipt in a chain, named by its position and its digest. Auditability

Hosted demonstration

ARC1 running on our production deployment with a fictional health plan’s synthetic data. The trusted path

Hub

ARC1’s control-plane server for policy, approvals, spend and evidence. It stays outside model traffic. The trusted path

I

Injection detector

The detector that weighs instructions in text by where the text came from, and permits, warns or blocks. Sanitization and DLP for AI

J

Job description

The source of a persona: purpose, principal, models, data classes, tools, budget, approvers and expiry. Agent Personas and Job Descriptions

L

LAB

Built or being built, and run only in our lab or against a vendor’s documented interface. The trusted path

Lab-verified

Run in our lab against a named system and version, with synthetic traffic. The trusted path

Last-known-good grace

How long an enforcement point keeps applying its last verified policy when the hub is unreachable: at most 72 hours. Policy IR and compiling everywhere

LIVE

In the released product and running on our hosted production deployment, with synthetic data. The trusted path

LLM proxy

The ARC1 Gateway’s relay for model traffic: routing, failover, keys, accounting and signed receipts. The LLM proxy

M

Managed gate

The check on managed Claude Code and Codex machines that inspects prompts and tool calls before they leave. Sanitization and DLP for AI

MCP gateway

ARC1’s governance of agents’ tool use: the registry, pins, authority, lineage and the inline proxy. The MCP gateway

MCP proxy

The inline process that checks each tool call against policy, pins and short leases before it reaches the server. The MCP gateway

Miss rate

The share of known sensitive cases a detector did not find, published from a signed evaluation. Sanitization and DLP for AI

Mode

How a control acts: observe counts, warn tells the person and continues, enforce stops. Policy IR and compiling everywhere

N

NEXT

Designed, and not yet built. The trusted path

O

OCSF

The Open Cybersecurity Schema Framework. ARC1 sends findings as its Detection Finding records. Visibility

P

Persona

A named, versioned job an agent runs as, compiled from its job description. Agent Personas and Job Descriptions

Persona digest

A scheduled email for one role that opens on an insight and gives each figure its basis. Visibility

Policy IR

ARC1’s declarative, versioned policy document, compiled into every target. Policy IR and compiling everywhere

Policy pack

The hub-signed policy the ARC1 Gateway verifies before it enforces. Policy IR and compiling everywhere

Privacy canary

An hourly synthetic sensitive case sent through each surface, with its outcome signed. Sanitization and DLP for AI

Proof bundle

A sealed, signed set of what ARC1 signed in a window, verifiable offline without ARC1. Auditability

Provider-billed spend

Cost as the model provider invoices it, and the base for coverage. Visibility

R

Readback

Reading a target’s configuration after it is applied and comparing it with the approval by digest. Policy IR and compiling everywhere

Realized saving

A recorded expense reduction on a ledger line, confirmed by finance. Budgets that act

Receipt

One signed, chained record of one decision. The receipt

Refusal

The compiler’s answer when a target cannot express a control: nothing is published to that target. Policy IR and compiling everywhere

Reservation and settlement

Holding a request’s worst-case cost before dispatch, then charging the usage the provider reports. The LLM proxy

Route

A set of approved destinations and the strategy that chooses among them. The LLM proxy

Row-level security

The database rule that lets each query see only its own organization’s rows. Auditability

Rug pull

A tool definition its server changed after the tool was approved. The MCP gateway

S

Shield verdict

The metadata-only record of one inspection: where it ran, in which direction, which detector, and what it did. Sanitization and DLP for AI

Signed configuration

The appliance’s configuration, signed and checked against a pinned key each time the gateway starts. The appliance

Structured identifier

A value with a defined format, such as a member ID, an NPI number or an API key. Sanitization and DLP for AI

Substitution

Serving one approved model in place of another, backed by a comparison of exactly that pair. The LLM proxy

T

Target

One enforcement point a policy compiles into, such as Kong or managed Claude Code settings. Policy IR and compiling everywhere

Tool pin

The exact set of tools an approved MCP server may expose. The MCP gateway

Trace

One request followed through its signed receipts, step by step. The receipt

Trace a dollar

A cost followed back to the verified receipts that make it up. The receipt

Trusted path

A route for AI use that is approved in advance, enforced on the way and proven afterwards. The trusted path

Trusted registry

The approved MCP servers, published read-only to the clients that read a registry. The MCP gateway

Two-person approval

The rule that the person who asks for a change cannot approve it. Governance

V

Virtual key

An expiring key bound to one device or runner, swapped for the provider credential at the gateway. The LLM proxy

W

Workload identity claim

A signed, expiring statement of an agent’s identity and its parent chain. The MCP gateway