Concepts

Architecture overview of the Meibel policy plane: how requests flow through it, how policies are defined, and what the audit log captures.

Request lifecycle

Every call through the Meibel proxy follows this sequence:

1. Tenant resolution — The X-Meibel-Tenant header identifies which tenant context applies, which policy to load, and which rate budget to debit.
2. Entity detection — The full assembled prompt (including RAG-retrieved context) is scanned for entity types defined in the active policy.
3. Redaction — Detected entities are replaced with typed mask tokens ([SSN_001], [PERSON_NAME_001]). Replacement is deterministic within a call — the same entity value always maps to the same mask index.
4. Policy verdict — If no entities were found: ALLOWED. If entities were redacted: REDACTED. If a block rule matched: BLOCKED.
5. Forward to upstream — The redacted prompt is forwarded to the upstream model API with your credentials.
6. Audit write — An immutable audit record is written: entity types found (not values), verdict, model version, prompt hash, tenant ID, timestamp.
7. Response return — The upstream response is returned verbatim with Meibel headers attached.

Policy DSL

Policies are YAML documents with three top-level sections: entities, blocks, and audit.

version: "1"
policy_id: healthcare-v2

entities:
  - group: pii-standard
    action: redact
    mask_format: [TYPE_INDEX]
  - group: hipaa-safe-harbor   # 16 HIPAA identifiers
    action: redact
    mask_format: [PHI_TYPE_INDEX]
  - pattern: "NPI-\\d{10}"         # custom regex
    label: PROVIDER_NPI
    action: redact

blocks:
  - if_entity: SSN
    in_context: system_prompt   # SSN in system prompt = hard block
    action: block
    reason: SSN in system_prompt not permitted

audit:
  enabled: true
  store_raw_prompt: false
  store_hash: true
  store_entity_types: true

Entity type taxonomy

Isolation models

Tenant isolation is enforced at the database partition level. There are three isolation models available depending on your deployment:

Audit log schema

Each audit record is immutable once written. The schema is fixed across API versions.

{
  "request_id":      "req_01HX9K2B4Y",        // unique ID, immutable
  "tenant_id":       "finance-wealth",
  "policy_id":       "finance-default-v1",
  "policy_verdict":  "REDACTED",              // ALLOWED | REDACTED | BLOCKED
  "model_version":   "gpt-4o-2024-08-06",    // from upstream API response
  "pii_findings":    ["PERSON_NAME", "SSN"],  // entity labels only
  "redaction_count": 2,
  "prompt_hash":     "sha256:a3f9cc...",       // sha256 of pre-redaction prompt
  "latency_ms":      312,
  "created_at":      "2025-11-14T14:23:08Z"
}