Glossary

The words Quard and its dashboard use, in alphabetical order.

Term	Meaning
Action guard	A guard on tools that change something, such as payments, emails or deploys. It checks the arguments and their labels against rules. See Guards.
Agent	An AI program that calls a model and tools to do work. Each agent has a name, such as `billing`, and versions.
Agent key	The secret the SDK uses to reach Quard, and the only Quard secret an agent needs. Created and revoked in Settings. It is shown once, when you create it.
Agent version	The model, instructions and tools an agent ran with. Recorded once per version, so an incident ties to the version that caused it.
Allow	A guard decision that lets a call run.
Always approve	An answer that lets later calls from the same agent to the same tool, with exactly the same arguments, run without asking until someone revokes it. Also called a grant.
Approval	A person’s answer to a call that waits: approve once, always approve or deny. See Approvals.
Approval guard	A guard that always asks a person before its tool runs.
Approve once	An answer that runs one call, with exactly the arguments shown, one time.
Ask	A guard decision that pauses a call until a person answers.
Block	A guard decision that stops a call. The tool never runs, and the agent gets a refusal.
Block rate	The share of guard decisions that blocked a call.
Carrier	The run id, parent step id and label reference that travel with a message between processes. See Multi-agent.
Chunk	A piece of public content, up to 4,000 characters, that the detector labels in one request.
Context label	The least trusted and most sensitive label among everything the model read before it asked for a call. See Labels and trust.
Damage	The harmful call at the end of an incident’s path, or the tool call whose content a guard flagged when no call was blocked.
Delegation	One agent handing work to another. A delegated agent can only use the tools its parent granted.
Deny	An answer that stops a waiting call. The agent gets a refusal.
Detector	An AI model that labels public content, such as `invoice` or `phishing`. It only makes things stricter. Quard’s detector is off until set up, then acts by default.
Egress guard	A guard on tools that send data out. It keeps internal data away from destinations that are not allowed.
Entry point	Where untrusted content first came into an incident’s path.
Fan-out	How many helper agents one agent starts in a run. One of the run limits.
Flag	A mark on content, such as `instructions` or `detector:payment_fraud`. Values from flagged content never count as coming from their origin.
Fleet	All your agents together.
Fleet check	A check that blocks a new recipient, IBAN or domain everywhere once a fifth separate run uses it within 24 hours. See Quarantine.
Grant	A standing always-approve answer. Grants are listed, and revoked, on the Approvals page.
Guard	A check that wraps a tool and decides whether each call may run. See Guards.
Guard decision	The result of a guard, with its rule and reason: allow, ask or block for a call, and pass, flag, strip or block for content a source tool returned.
Handoff	An agent passing a task or a message to another agent.
Hash key	The key that IBANs and emails are hashed with. Each project has its own, made from the server’s `QUARD_HASH_KEY`. The SDK gets it with its agent key.
Incident	A record that Quard opens for a run when a guard blocks a call or flags content, also in observe mode, with its path, verdict and replay. See Incidents.
Influence path	Every step from the entry point to a call, shown on an approval request.
Influenced	A call whose context label is untrusted, because the model had read untrusted content before asking for it.
Label	Where a piece of content came from (its origin), with its trust and sensitivity. See Labels and trust.
Label reference	A key to the labels a sender stored for a message. The receiving side looks the labels up with it.
Limit guard	A guard that caps calls, amounts and cost per run, agent or day.
Limited replay	A replay that can’t run: its request was not recorded, the suspect content is not a tool result, or no harmful call came. See When replay is limited.
Loop	Handoffs back and forth between the same two agents. One of the run limits.
Mask	The short form of a sensitive value shown in the dashboard, such as `DE89…3000`.
Model-generated	A value that appeared nowhere earlier in the run. See Value tracing.
Observe mode	A rule mode that records what the rule would have done but lets the call run.
Origin	The tool, domain, sender, server or agent that let content in, such as `web:supplier-portal.example`.
Origin override	A change, made in code, to the default trust or sensitivity of one origin. Settings lists the overrides your runs report.
Paid, not delivered	An x402 payment that settled while the paid response was an error. See Payments (x402).
Quarantine	The values the fleet check blocked everywhere. A value leaves it when someone marks it as known.
Redaction	Removing secrets, and replacing sensitive values with keyed hashes and masks, before anything is stored or sent to the detector.
Refusal	The message an agent reads instead of a tool’s result when a call is blocked or denied.
Replay	Rerunning an incident’s turning-point model call with and without the suspect content, to test the cause. It runs when someone starts it on the incident page. See Replay.
Root-cause finder	The part of Quard that traces an incident’s path, writes its verdict and runs replay.
Rules hash	A short fingerprint of an app’s active rules, recorded with every guard decision.
Run	One task from start to finish, across every agent that worked on it. See Runs.
Run graph	Who delegated to whom in one run.
Run limits	Caps per run on delegation depth, fan-out, loops, model calls and cost. See Run limits.
Sensitivity	Whether content is internal or public.
Shared memory	A store that agents read and write, wrapped so its items keep their labels. See Shared memory.
Source guard	A guard on tools that bring content in. It labels and scans their output.
Stand-in	A made-up value in the same format as a masked one, such as an IBAN with a valid checksum, that replay sends in its place.
Step	One thing that happened in a run: a model call, a tool call or a guard decision.
Strip	Removing part of the content before the agent reads it: suspect lines for the source guard’s rules, or a whole chunk for the detector.
Trust	Whether content is trusted or untrusted.
Turning point	The model call that decided to act on untrusted content.
Untrusted	Content from an origin Quard does not trust by default, such as web pages and outside email.
Value tracing	Finding every place an argument’s value appeared earlier in the run. See Value tracing.
Verdict	An incident’s finding: its category and the guard that was missing.
Watching	New values the fleet check is counting toward its limit.
Would block, would ask	What a rule in observe mode would have done.
x402 guard	A guard on an x402 payment client. It checks each payment before it is signed. See Payments (x402).