Context Relay Protocol v3 · governing a real local LLM

Watch CRP govern the model on your machine

These demos drive a local LLM (LM Studio, Ollama or llama.cpp) through the Context Relay Protocol and surface every governance signal it produces — grounding, hallucination risk, prompt-injection shields, tamper-evident audit trails, HMAC provenance chains and the full CRP-* header set.

Live LLM detection

CRP introspects the inference layer before it sends a single token.

Scanning local runtimes…

All discovered models

Runtimes probed

APP 1

AI Safety & Governance Console

Enter a prompt + grounding context + a CRP safety policy. Watch CRP decide whether the answer is safe to release.

  • Prompt-injection shield (input-side)
  • Decision Provenance Engine: grounding, fabrication, risk
  • Policy enforcement → HTTP 451 halt
  • Tamper-evident audit trail (OCSF / SARIF export)
APP 2

Context Management & Provenance Explorer

Have a multi-turn conversation. Watch CRP remember facts and chain every window with tamper-evident HMACs.

  • Contextual Knowledge Fabric (persistent memory)
  • Per-window HMAC provenance chain
  • "Tamper" a window → chain turns BROKEN
  • Session tokens + ETag + CRP headers

What makes this significant

1
It detects the model. CRP reads each runtime's real context ceiling, quantization, tool-calling and reasoning capabilities.
2
It governs the output. Every answer is scored for grounding and hallucination risk, then enforced against a declarative safety policy.
3
It proves what happened. Hash-chained audit trails and HMAC window chains make tampering detectable after the fact.