The Anthropic distillation scandal changed how I think about verification
Quick context if you missed it: in February 2026, evidence surfaced that Anthropic models had been trained on distilled outputs from other frontier models. The boundaries between Claude, GPT, and Gemini turned out to be blurrier than anyone had assumed.
For me, this hit personal. I'd just left my Fortune 500 job to build CrossCheck AI — a tool that runs your prompt through 3+ models in parallel and flags disagreements. The whole premise rested on one assumption: that different models trained by different companies have different blind spots. If three models agree, the answer is probably right. If they disagree, one of them is hallucinating.
Distillation breaks that assumption.
If Claude was partially trained on Gemini's outputs, where they agree isn't independent verification anymore. It's the same source thinking out loud through two different mouths.
So what I thought I was building — a consensus-detection tool — needed to flip. The new product doesn't celebrate consensus. It hunts for adversarial disagreement. Three models criticize each other's answers, and the moments where they fail to agree are exactly where the signal lives. When they all converge on the same fact, sometimes it means truth. Sometimes it means shared training data. The disagreements are where verification actually happens.
I shipped public beta this week. It's still rough — known issues with Gemini's stream parsing, occasional pipeline stalls. BYOK only (you bring your own API keys, no markup, no token proxying through us).
If you've ever wasted time fact-checking AI output by manually pasting between Claude and ChatGPT, this is exactly that workflow automated. ~10 min to set up keys, then verification in 60-180 seconds.
Would value early eyes on it from this community especially. We're looking for sharp critique while it's still pre-traction.
