Research
We experiment based on research data from Conscience Intelligence (US R&D organization, upcoming), and share findings that contribute to the operational stability and performance optimization of AI systems
research|2026年4月
Failure-Aware Prompting: Injecting Past LLM Failures into Prompts Boosts Accuracy from 81.2% to 90.0%
Collect actual LLM failures, abstract into patterns, inject into prompts. Random injection achieves the same effect as retrieval
research|2026年4月
Nora Hallucination Detector: Frontier-level Hallucination Detection with a 9B Model
980 samples, 22 minutes of training, 510MB adapter. A 9B parameter model achieved 89.6% agreement with Claude 4.6