Hallucination is the fundamental trust problem in AI. Language models confidently generate plausible-sounding but factually incorrect information. Prajna-V2 is our approach to solving this — a lightweight Cognitive Resonance Network (CRN) adapter that achieves 100% exact recall on verifiable factual queries.
The CRN Architecture
Instead of modifying the base LLM's weights, Prajna-V2 attaches a 6.7M parameter CRN adapter to a frozen Gemma-4-E2B backbone. The adapter implements an EpisodicMemory module — a 17,810-entry retrieval table that maps query patterns to verified factual responses.
How It Works
When a factual query arrives, the CRN first checks its EpisodicMemory table. If a verified answer exists, it overrides the base model's generation via logit adjustment. If no match is found, the base model generates normally. This gives you the best of both worlds: perfect recall on known facts, and generative flexibility on novel queries.
Results
On our CEHRI benchmark (Contextual Episodic Hallucination Resistance Index), Prajna-V2 achieves 100% exact recall on verifiable facts, 100% accuracy on arithmetic, and strong performance on implicit goal reasoning tasks. The adapter runs efficiently on Apple Silicon via MPS.
Provisional Patent
The CRN architecture is covered by a provisional patent filing. The model weights are available on HuggingFace under the Gemma license.