SEATTLE, Sept. 6 — A Seattle-based startup is betting $25 million that it can teach artificial intelligence to tell the truth — and to prove it when challenged. Resect AI, an artificial intelligence research company, announced Thursday that it has raised $25 million in early funding to build what it calls an “accountability layer” for enterprise AI. The goal: capture and reduce hallucinations at runtime, stopping AI models from confidently presenting fabricated answers as fact.
The problem is a familiar one to anyone who has watched a chatbot stumble. Hallucinations, or confabulations, happen when a model replies with a false response and delivers it with total confidence.
Models don’t remember; they predict patterns based on training. When data runs thin, they guess. Chief Executive Kevin Owens put it bluntly.
“AI has prematurely been put in a position of trust,” he said. “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.” Resect’s answer is an open-source offering that opens the black box. The company is building tools that look inside large language models to observe, detect, interpret, audit and modify behavior — a level of transparency the industry has long dismissed as impossible. “Many argue that understanding the black-box internals of LLMs is out of reach, but we fundamentally disagree,” said Chief AI Officer Tim Walton.
“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.” The company’s path began with its own models, built for high factuality with a training process it calls behavior and reinforcement training. That approach, Resect found, could be applied to existing open models — DeepSeek, Qwen and Llama among them. The task then shifted to building tools that work inside the model architecture itself, intercepting misbehavior before a hallucination ever surfaces.
Those internal visibility tools will anchor an enterprise-level audit product called the NeuroWave Product Suite — described by the company as a polygraph for neural networks. “Through this process, we’ve developed technology that observes exactly when and how models fail, and surgically fixes them,” Walton said.
The name Resect is a medical verb — to cut out or remove part of an organ, tissue or bone. The surgical metaphor runs through the company’s approach. Two models are already live on HuggingFace.
Both are based on the Qwen3 architecture and operate in non-thinking mode, meaning they lack an underlying chain-of-thought layer. The smaller, a 600 million-parameter fact checker called Veritas, scored an average of 72.3% on the LLM-AggreFact benchmark — a 7.4% improvement over Qwen3.
An 8 billion-parameter model is also available. A GitHub repository for open-source tools is linked but not yet populated. Resect did not announce the names of its backers, saying only that the funding came from private equity investors.
The money will go toward research and development, go-to-market efforts and hiring locally in the greater Seattle and Portland regions.
