Resect launches with $25M to reduce hallucinations in AI models

Hallucinations, also called confabulations, occur when an AI model replies with a false answer presented as confident. Since models predict patterns from training data rather than "remembering," gaps in information can lead them to guess. During refinement and post-training, many models are designed to appear helpful and may confidently give incorrect answers instead of saying "I don't know."

"AI has prematurely been put in a position of trust," said Chief Executive Kevin Owens. "Adding labels such as 'use at your own risk' flies in the face of proper governance or compliance. AI must be anchored in truth to be widely adopted across the enterprise." Owens said Resect's goal is to bring accountability and transparency to AI systems, which have so far been black boxes. The company is building an open-source offering that can look inside large language models to observe, detect, interpret, audit and modify model behavior.

Chief AI Officer Tim Walton said, "Many argue that understanding the black box internals of LLMs is out of reach, but we fundamentally disagree. We've spent an extensive amount of time and resources researching how models think, and what causes them to choose the answers that they do."

Resect started by building its own high-factuality models, the company said. It then found that its model-building and training approach could be applied to existing open models, including DeepSeek, Qwen and Llama. Rather than focusing on building a better model, the start-up turned to developing a suite of tools that work within the model architecture to intercept misbehavior and prevent hallucinations before they occur. The company said those internal visibility tools will form the basis of NeuroWave Product Suite, an enterprise audit tool it describes as a polygraph for neural networks.

The start-up currently offers two models on HuggingFace: a 0.6-billion-parameter fact checker called Veritas and an 8-billion-parameter model. Both are based on Qwen3 architecture and operate without a chain-of-thought layer. According to the model cards, Veritas scored an average of 72.3 percent on the LLM-AggreFact benchmark, an improvement of 7.4 percent over Qwen3. Resect also links to a GitHub repository for open-source tools, but the repository was not populated at the time of publication.

Resect did not identify its backers, but said the funding came from private-equity investors. The proceeds will go toward research and development, go-to-market initiatives and hiring in the greater Seattle and Portland regions.