AI should bring evidence, not verdicts

Traders route real money through automation only when they can see into it, override it and trust how it behaves under stress. That was the brief at MarketAxess, where I led the trust layer for Pragma AI, an execution engine handling $2T+ in order flow.

The lesson carries into everything since. A machine can be fast and right most of the time. People still need to see why, and to step in when it is wrong.

In Halo the rule is strict. Ask Halo can rank what drives a value, fetch the evidence and run a stress test. It cannot set a value, change a convention or sign off. Every input shows its source, its age and its confidence, or shows as absent.

That last part matters most. An AI that fills a gap with a plausible guess is more dangerous than one that says it doesn’t know.

Three tests I use before any AI feature goes near high-stakes work. Can the user see what it used? Can they override it in one step? Does it say when it is unsure? If any answer is no, it isn’t ready.