A company called TrustScale has released Argus, a tool that claims to catch your AI’s lies before they reach a customer. The pitch is good. The claims deserve a second look.
On the 5th of August, TrustScale announced Argus, which it calls a ‘patent-pending AI assurance platform.’ The job it is built for is a real and expensive problem: large language models make things up, confidently, and someone downstream has to catch it. Argus sits over your AI output, flags claims it cannot support with evidence, gives each a ‘TrustScore,’ and suggests evidence-based corrections before the text gets published or sent.
The part TrustScale leans on hardest is the method. Most ‘guardrail’ products use one AI to check another AI, which has an obvious flaw: the referee hallucinates too. TrustScale says Argus instead uses deterministic verification against empirical evidence, meaning it checks claims against actual source data rather than asking a second model for its opinion. If that holds up in the wild, it is a genuinely better idea than the AI-marking-its-own-homework approach most vendors sell.
Read the numbers with your hand on your wallet
TrustScale cites research that even simple queries on frontier models can average up to 20% hallucinations, rising to as much as 88% for some industry-specific questions. Those figures are dramatic, and they are doing marketing work. An 88% number depends entirely on how you define a hallucination and which narrow domain you test, so treat it as a reason to care, not a measured fact about your own workflow.
A few things are missing from the launch, and their absence is the story. There is no public pricing. There is no independent benchmark of Argus against the existing fact-checking and grounding tools already on the market. And the launch arrived as a wire press release, which tells you it is aimed at enterprise buyers with procurement teams, not at a solo operator trying to keep a chatbot honest on a budget.
The angle: the underlying idea, checking AI claims against real evidence instead of a second model’s vibes, is the right direction, and worth watching. But ‘patent-pending assurance platform’ with vendor-supplied hallucination stats and no price is a sales motion, not a proven result. If you run AI in front of customers, ask for a trial on your own data and your own error cases before anyone quotes you an annual figure.
Did you know: the word ‘hallucination’ for confident AI nonsense annoys a lot of researchers, who argue ‘confabulation’ is more accurate, because the model is not seeing things, it is filling gaps with plausible invention.
Related on Top Tool Stack: The pass^k reliability test · Testing your AI automations (n8n)