Skip to main content

Data-class catalog

Every request and document is classified by content into exactly one data class. The class drives the policy matrix lookup. This is the authoritative catalog; for the concept, see data classes.

Classification resolves the highest-priority class when content matches more than one (for example, content tagged both Public and Unknown resolves to Public). Class names have short and long aliases (e.g. PII, PHI, and Identity all resolve to PII / Personal Data).

Data classRiskCoversDefault handling
Source Code / SecretsCriticalAPI keys, passwords, tokens, connection strings, private source codeBlock — every tier
PCI / Payment DataCriticalPayment-card numbers (PAN) and related payment dataEnterprise-managed · redact
Strategy / Board ConfidentialCriticalBoard materials, M&A, confidential strategy and forecastsEnterprise-managed · route
Legal / ContractualHighContracts, legal advice, settlementsEnterprise-managed · route
Financial DataHighRevenue, forecasts, internal financialsEnterprise-managed · route
HR DataHighCompensation, performance, employee recordsEnterprise-managed · route
PII / Personal DataHighSSNs, national IDs, passports, identity documents, PHIEnterprise-managed · route (sensitive spans redacted)
Customer DataMediumCustomer records and account detailsEnterprise-managed · route
Public / Non-sensitiveLowPublished or general contentPublic frontier · allow
UnknownHighThe classifier is not confidentEnterprise-managed · route
Unknown is protected, not permitted

A low-confidence classification is treated as elevated risk — routed to a trusted destination, never allowed through by default.

Detectors

Classes are assigned from a mix of deterministic detectors and a semantic classifier:

  • Deterministic (regex/validators) — secrets and API keys, payment-card numbers (Luhn), SSNs and national IDs, and prompt-injection patterns. These are high-precision and can hard-block on their own.
  • Semantic classifier — content-level classification for the business classes (Financial, Strategy, Legal, HR, Customer). Runs before routing, on a fixed safe classifier, and never on the chat destination.
  • Custom rules — organization-specific detectors you add (DLP & rules), mapped to any class.

Notes

  • The risk level is inherent to the class; it drives dashboards and prioritization.
  • The Default handling column is the shipped DEFAULT_TRUST_MATRIX — every row is tenant-tunable in the policy matrix.
  • See the policy-action catalog for what each action does.