What each agent is allowed to do — authorized vs. observed.

Honesty marker

Authorization is the default-action setting (defaultDomainAction) plus the guardrail adapters. We compare what's allowed against what each agent is observed doing; gaps are unsanctioned actions.

Default stance · defaultDomainAction

The baseline decision applied to any action that no guardrail explicitly handles.

Default-allow

Every agent action is permitted unless explicitly denied.

Authorization matrix

Per agent, per action class: what's used, blocked, or unsanctioned — derived from observed actions.

Agent

Model provider

MCP server

Tool API

Internal service

Unapproved / unknown

marketing-gptungoverned

●

◼

support-copilot

●

data-pipeline

●

claude-code

●

cursor

●

● used· none✕ unsanctioned◼ blocked

Guardrail adapters · enforcement stack

The @qpoint/* packages that intercept actions — plus the one that's missing.

@qpoint/scanSensitive-data scan (PII/PCI/PHI)Detects qscan sensitiveDataType in payloads.ondatasupport

@qpoint/errorsError & anomaly detectionFlags failed/again actions.onengdatasupportmarketing

@qpoint/access_logsAction audit loggingRecords every action for the activity log.onengdatasupport

(none)Require owner/identity before actionNo adapter enforces agent ownership before an action — why an unowned agent can act.OFF · gapno coverage

Unsanctioned actions

Allowed-but-shouldn't-be, or blocked — the gap between authorized and observed.

AgentTool / modelDecisionWhy

marketing-gpt104.18.32.7ALLOWunapproved model endpoint · unowned agent

marketing-gptapi.unknown-llm.exampleBLOCKblocked · unowned agent

support-copilotgenerativelanguage.googleapis.comALLOWunapproved model endpoint

Fields: defaultDomainAction · action (ALLOW|DENY|BLOCK) · endpointId · class · guardrail packages (@qpoint/*) · riskLabels