Skip to main content

​
Why TealTiger audits everything (safely)

TealTiger is built for security and governance workflows, where you need evidence of:
  • what was evaluated,
  • what decision was made,
  • why it was made,
  • and how it was enforced,
without leaking sensitive data (prompts, responses, secrets, PII). In v1.1.0, TealTiger’s audit posture is:
  • ✅ Redaction by default
  • ✅ Correlation-driven traceability
  • ✅ Deterministic decision evidence
  • ✅ Safe structured metadata for cost/security reporting
Audit logs are only useful if they are safe to retain. Redaction is not a nice-to-have — it’s the default security posture.

​
The audit pipeline (what happens to data)

​
Redaction strategy (v1.1.0 guarantees)

TealTiger applies redaction before persistence. Raw inputs and outputs are never written directly to audit storage.

​
Guaranteed redaction rules

In v1.1.0, the following are never persisted in audit logs:
  • raw prompts
  • raw model responses
  • secrets (API keys, tokens, credentials)
  • detected PII/PHI fields
  • full tool payloads containing sensitive values
Instead, TealTiger records safe derivatives.
If a value is redacted, it is redacted before the AuditEvent is created. There is no “raw audit” mode in v1.1.0.

​
What is logged (AuditEvent contract)

TealTiger emits a structured AuditEvent that captures decision evidence, not raw content.

​
Stable AuditEvent fields

The following fields are guaranteed and stable:
  • decision.action
    One of: ALLOW, DENY, REDACT, TRANSFORM, DEGRADE, REQUIRE_APPROVAL
  • decision.reason_codes[]
    Stable, machine-readable identifiers explaining why the decision occurred
  • decision.risk_scores (optional)
    Normalized risk signals (security, cost, reliability)
  • policy.id / policy.version
    Which policy produced the decision
  • enforcement.mode
    REPORT_ONLY, MONITOR, or ENFORCE
  • correlation.trace_id / correlation.span_id
    Used for end-to-end traceability
  • redaction.applied
    Boolean flag indicating redaction occurred
  • metadata.safe_context
    Sanitized, structured metadata (never raw text)
Audit logs are designed to be SIEM-safe by default. They can be retained long-term without leaking sensitive data.

​
SafeContent vs RawContent

TealTiger intentionally separates content from evidence.
  • Raw prompt: ❌ never stored
  • Raw model output: ❌ never stored
  • PII/PHI values: ❌ never stored
  • Decision outcome: ✅ stored
  • Reason codes: ✅ stored
  • Risk scores: ✅ stored when enabled
  • Tool name (normalized): ✅ stored
  • Redaction flag: ✅ stored
If you need full prompt/response retention, handle that outside TealTiger using your own storage, consent, and retention controls.

​
Redaction behavior by decision type

  • ALLOW: evidence only
  • DENY: evidence only
  • REDACT: sensitive fields removed before audit
  • TRANSFORM: only effective values are referenced (never originals)
  • DEGRADE: safe defaults recorded; raw inputs excluded
  • REQUIRE_APPROVAL: context summarized; not stored verbatim

​
Determinism & reproducibility

Audit events support deterministic replay without storing raw data. You can answer:
  • Why was this blocked?
  • Which policy fired?
  • Was this enforced or only reported?
  • Did redaction occur?
Without needing:
  • the original prompt
  • the original response
  • sensitive payloads
Unless explicitly overridden:
  • Redaction is ON
  • Audit emission is ON
  • Enforcement mode defaults to REPORT_ONLY
  • Correlation IDs are auto-generated if missing
These defaults are chosen to make TealTiger safe to enable early in production without increasing data retention risk.

​
What will NOT change

The following guarantees are locked:
  • Redaction happens before persistence
  • Raw prompts/responses are never logged
  • Reason codes are stable identifiers
  • Audit event schema remains backward compatible within v1.1.x
Existing meanings remain stable; future versions may add fields only in compatible ways.
  • /audit/audit-event-schema
  • /policy/reason-codes
  • /policy/risk-scores
  • /architecture/enforcement-flow
  • /versioning/stability-guarantees