Skip to content

The Black-Box Question

"If LLMs are black boxes, how can CRP make any guarantees about what they produce?"

This is the most important sceptical question, and it deserves a direct answer. Black-box governance is not only possible — it is the only honest form of governance for any system whose internals are opaque.

The Short Answer

CRP makes no claims about what happens inside the model. It governs what goes in and what is done with what comes out.

[What CRP controls]                    [The black box]

 ┌─────────────────┐                   ┌──────────────┐
 │ Context Envelope│ ── structured ──► │              │
 │ (what goes in)  │     input         │   LLM Model  │
 └─────────────────┘                   │  (CRP cannot │
                                       │   see inside)│
 ┌─────────────────┐                   │              │
 │ DPE Analysis    │ ◄── raw text ──── │              │
 │ (what comes out)│     output        └──────────────┘
 └─────────────────┘

The LLM is given a precisely constructed Context Envelope (the input layer). What it produces is then subjected to 13 stages of analysis — claim detection, attribution, fabrication, distortion, entailment, contradiction, repetition, completeness, flow, hallucination risk, quality tiering, policy evaluation, provenance binding. Based on those analyses, CRP decides whether to deliver, flag, halt, or re-dispatch the response.

Analogies That Dissolve the Objection

This is not fundamentally different from how every other form of governance works:

Governance system Inside the black box What is governed
Financial audit What was in the accountant's mind Whether the numbers are consistent and traceable
Breathalyser Liver chemistry What is present in the output
HTTP security headers What is in the database What gets delivered
TLS The content being encrypted Confidentiality, integrity, authenticity of transit
Drug test Body chemistry Whether specific substances are detectable
CRP Model weights and internal representations The observable outputs of every inference

In each case, the inability to see inside is acknowledged, and governance proceeds on the observable surface. The result is real, measurable, and verifiable — not theatre.

The Two Claims, Untangled

The sceptical "LLM atheist" position usually conflates two different statements:

  1. "You cannot fully understand what an LLM will do in all situations." — TRUE.
  2. "You cannot verify, govern, and record what it actually did in specific calls." — FALSE. CRP disproves this.

The first is an alignment research problem. The second is a protocol engineering problem, and it has a solution.

What CRP Therefore Provides

  • Verifiable — every assertion the protocol makes about a response is expressed in headers that are cryptographically chained to the audit trail.
  • Measurable — the DPE produces numerical scores (hallucination risk, flow score, completeness ratio), not vibes.
  • Enforceable — Safety Policy directives stop bad responses before they reach the application.
  • Regulatable — the evidence emitted satisfies EU AI Act Article 12 / 13 logging requirements, ISO 42001 control evidence, GDPR Article 22 traceability.

The black box is real. Its observable surface is also real. CRP is the governance layer for that surface.

Further Reading