[I-D] New: draft-vidiniotis-crp-headers-00 — HTTP Header Fields for AI Context Safety and Governance

[Constantinos Vidiniotis <[e-mail]>]

Hi all,

I have submitted an Internet-Draft proposing a vocabulary of HTTP header
fields for AI context governance, safety signalling, and compliance evidence
on AI inference request/response cycles.

*Draft*: https://datatracker.ietf.org/doc/draft-vidiniotis-crp-headers/
*Repository*: https://github.com/AutoCyber-AI/crprotocol-specs
*Website*: https://crprotocol.io/spec/

*Motivation*

AI inference calls (requests to large language models) currently carry no
standardised metadata about the quality, safety, or governance status of
their responses. Every operator instruments this separately, producing
non-interoperable signals.

This draft applies the HTTP header pattern - the same pattern that gave us
Cache-Control, ETag, and Content-Security-Policy - to AI request/response
cycles. The draft defines six header namespaces:

  CRP-Context-*     envelope state, quality tier, ETag for conditional
dispatch
  CRP-Safety-*      hallucination risk, attribution, fidelity, oversight
mode
  CRP-Provenance-*  HMAC chain integrity, audit trail URI
  CRP-Compliance-*  EU AI Act class, NIST tier, GDPR PII flag
  CRP-Agent-*       agentic dispatch state, safety budget across agent
chains
  CRP-Memory-*      context cache layer signals

A separate draft (draft-vidiniotis-crp-safety-policy-00) defines a
declarative directive language for these headers, inspired by CSP.

*What this is NOT*

This is not a new inference protocol. The headers ride on existing HTTP
requests to existing LLM providers (OpenAI, Anthropic, etc.). The draft
explicitly mandates that CRP headers MUST be stripped before forwarding
to providers - the model remains ignorant of the protocol layer.

This is not a replacement for MCP or A2A. Both are based on JSON-RPC over
HTTP. The CRP headers govern the AI calls made by MCP tools and propagate
across A2A hops. CRP sits at a different layer.

*What I'm asking*

I'd appreciate feedback on:

  1. Header naming and namespace structure
  2. The interaction model between CRP-Safety-Policy and
Content-Security-Policy
     (which inspired it)
  3. Whether IANA provisional registration of the priority 10 headers
     (listed in §16.1 of the draft) is the right starting point
  4. Whether this work should pursue a new WG, individual submission,
     or seek adoption by an existing WG


Two independent implementations are in development to meet Proposed
Standard interoperability requirements.

Thank you for reading. Comments and critique welcome, both on-list and
to the issue tracker on the GitHub repository.

Best,
Constantinos Vidiniotis
AutoCyber AI Pty Ltd
[e-mail]