Hallucination detector works across any generator model without retraining

The key insight is portability through architectural decoupling. Prior probe-based hallucination detection required read access to the generator's internal states, making it incompatible with closed-weight APIs and forcing retraining per model update. This work eliminates that dependency by training a separate observer model that only needs the context and the final response, not the generator's hidden layers.

This connects directly to the source-aware verification work Hugging Face published on MCP agents (late September). Both papers recognize that hallucination detection must move upstream and away from tight coupling to specific model architectures. The observer-model approach here is a complementary strategy to the agent-layer verification that HF explored. It also echoes the architectural introspection pattern from the MoE routing work (also late September), which showed that detection can happen by reading internal signals rather than post-hoc response filtering. Unlike HARPO's reinforcement learning path to grounding, this is purely a detection layer, making it applicable to any generator regardless of training methodology.

If this observer approach maintains detection accuracy when the generator is swapped from Llama 3 to GPT-4 to an open-weight model without retraining the observer, that validates the portability claim. If adoption appears in production RAG systems within six months (watch for mentions in Hugging Face or Anthropic deployment case studies), the practical constraint has been genuinely removed.