What happened
QUASA reports that Resect emerged from stealth in Washington state on September 3, 2026, with $25 million in financing and a proposal for software that would observe, detect, interpret, audit, and modify large-language-model behavior while answers are being generated.
QUASA says the financing will support research and development, commercialization, and hiring around Seattle and Portland. The report says the investors, valuation, conventional financing stage, and technical milestones tied to the funding have not been disclosed. Citing GeekWire’s launch coverage, QUASA places Resect’s headcount at 30, but says the staffing information does not show whether customers are running controlled pilots or when a commercial system might become generally available.
The proposed technology is described as operating during model generation rather than simply evaluating completed text. According to QUASA, it is intended to identify a developing failure, change the model’s behavior before the answer reaches a user, and preserve an audit trail. The report says public materials do not explain the trigger signals, intervention method, required model access, supported architectures or inference engines, or whether the system works through proprietary model APIs.
QUASA says a RuntimeWire review found Apache 2.0-licensed Veritas fact-checking models with 0.6-billion and 8-billion parameters. The report cites a company-published average balanced accuracy of 72.30% for the smaller model on LLM-AggreFact, compared with 64.93% for a specified Qwen3 baseline in non-thinking mode. QUASA calculates the difference as 7.37 percentage points, while emphasizing that this is a self-reported result for a particular fact-checking setup, not an independent evaluation of Resect’s proposed in-stream control.
Why it matters
If demonstrated, an intervention that acts during generation could give organizations a control point before an answer reaches a user, while creating records for later review. That could be useful in high-stakes deployments where post-generation checking is too late. The practical value is not established, however: the evidence described by QUASA concerns fact-checking models, not independent testing of a separate system that detects and corrects errors in real time.
The distinction between checking finished output and changing a model’s behavior during inference is consequential. A real-time control would need to recognize an error early enough to intervene, avoid suppressing unusual but correct answers, and improve the final result rather than merely blocking it. It would also need to work across the models and deployment environments that enterprise buyers actually use.
The source does not independently confirm that the $25 million financing closed, identify the investors, or establish that Resect has a functioning enterprise product. It also does not provide evidence that the cited Veritas results transfer to production applications, closed model APIs, self-hosted systems, or fine-tuned models.
What to watch next
The key test is whether Resect can publish reproducible evidence that its proposed intervention prevents hallucinations without imposing unacceptable latency, compute costs, or false positives. Access, pricing, supported models, and general availability are currently unknown.
Meaningful proof points would include versioned software and evaluation artifacts, reproducible in-stream benchmarks, exact model and inference-environment details, a clear hallucination definition, and separate measurements for detection, successful correction, false positives, and false negatives.
Operational results will matter as much as accuracy. Resect should disclose latency, throughput, and compute consumption with the control enabled and disabled, plus a compatibility matrix covering open-weight models, fine-tuned variants, self-hosted inference stacks, and closed APIs.
QUASA says the enterprise suite is not broadly available as a finished product. The source provides no access process, customer list, pilot results, release date, or pricing, so those remain unknown.