Skip to content

The Safety Policy Directive Language

CRP-Safety-Policy is an HTTP header that lets the caller (or a gateway, or a WAF) declare what the protocol should do when the DPE flags a response.

It is to AI safety what Content-Security-Policy is to web XSS: a wire-level, declarative, transport-layer enforcement mechanism. Full normative grammar is in SPEC-006.

Examples

CRP-Safety-Policy: halt-on CRITICAL; redact-on HIGH PII; warn-on MEDIUM

CRP-Safety-Policy: classify-eu-ai-act MANDATORY;
                   audit-mode CHAIN;
                   max-hallucination-risk 0.30

CRP-Safety-Policy: default-src self;
                   allow-claim-source envelope, ckf;
                   block-claim-source model-knowledge

Why Declarative Beats Code

Every line of application code that forgets to call a safety routine becomes a hole. A policy header is enforced by the protocol layer for every call, regardless of whether the application code remembered. The pattern is exactly the one CSP solved for browsers a decade ago.

Directive Categories

Category Directives Purpose
Action halt-on, redact-on, warn-on, dispatch-on What to do when a severity threshold trips
Source allow-claim-source, block-claim-source Which knowledge sources are admissible
Classification classify-eu-ai-act, classify-nist-rmf Mandatory regulatory tagging
Audit audit-mode, audit-export-to How the verdict is recorded and where
Limits max-hallucination-risk, min-quality-tier, max-windows Numeric guard-rails

Enforcement Verdicts

When the DPE evaluation completes, the gateway returns one of:

  • 200 OK — policy passed, response delivered.
  • 200 OK with CRP-Safety-Verdict: WARN — delivered with a warning header.
  • 200 OK with redacted body and CRP-Safety-Verdict: REDACTED — sensitive content stripped.
  • 451 Unavailable For Legal Reasons with CRP-Safety-Verdict: HALT — response withheld; reason in CRP-Safety-Reason.
  • 409 Conflict with CRP-Safety-Verdict: REDISPATCH — the caller is asked to re-dispatch with adjusted envelope.

451 was chosen deliberately to make policy halts unambiguous and machine-readable across CDNs, WAFs, and observability tools.

See also

  • SPEC-006 — full normative grammar (ABNF).
  • SPEC-005 — how verdicts are produced.
  • SPEC-014 — interoperability tests for policy enforcement.