There is a number I cannot stop thinking about since I started logging it: zero.
That is how many times, across two weeks of daily AI use, I was able to verify what actually happened inside the tools I relied on. Dozens of queries a day, real decisions resting on some of them, and not once could I check which model ran, whether it used the data it claimed, whether the reasoning shown was the reasoning that executed.
I had filed verification under "niche concern." Something auditors care about. A high-stakes edge case, irrelevant to ordinary use.
Logging it for two weeks reframed the whole thing.
Verification is not niche. It is a cost you pay silently every session it is missing — the price of acting on outputs you cannot confirm. Nobody notices, because the cost is spread across hundreds of tiny moments instead of arriving as one visible bill. Until the day an unverified output is wrong in a way that matters, and there is no trail to trace it back through.
I started calling it trust debt. Invisible, compounding, and growing the more decisions you route through AI you cannot inspect.
That is the problem @OpenGradient is building around. Verifiable inference as infrastructure, not a feature toggle — every call leaving a proof, TEE attestation or zkML, that the model you relied on is the one that actually ran.
I do not know yet whether the execution matches the ambition.
But I have a log now that makes the cost of the alternative very hard to unsee.