Governance Pipeline
Request lifecycle
Every request that passes through SAIG is evaluated by a deterministic 9-step governance pipeline. The same input with the same policy always produces the same decision.
Pipeline steps
Parse
Extract messages, metadata, model reference, and attachments from the incoming OpenAI-compatible request.
PII Detect
Scan content using hybrid NER + regex across 6 EU languages (EN, DE, FR, IT, CS, SK). Identify entities: PERSON, EMAIL, PHONE, IBAN, ADDRESS, and more.
Classify
Determine intent (24+ business categories), ownership (first-party, third-party, unknown), and abuse type (22 threat categories including prompt injection).
Policy Decision
Evaluate against the deterministic 6-rule decision matrix. Output: ALLOW, ANONYMIZE, DENY, or SYNTHETIC_ONLY. Generate decision trace.
Anonymize
If policy requires anonymization, replace detected PII with stable reversible placeholders ([PERSON_1], [EMAIL_1], etc.).
Route
Select provider based on sovereignty mode, model alias, availability, and circuit breaker state. Enforce residency constraints.
Forward
Send the governed request to the selected LLM provider. Perform outbound PII verification before egress.
De-anonymize
Restore original values from placeholders in the provider response, returning natural text to the caller.
Audit
Write tamper-evident audit record with SHA-256 hash chain reference and Ed25519 signature. Attach governance receipt to the response.
Deterministic guarantee
The pipeline is deterministic: same input + same policy = same governance decision. This is critical for compliance evidence — decisions are reproducible and auditable.