Most AI investing tools ask you to trust the output. This is the whole path a recommendation takes, including the parts that can say no.
Rumi reads research and surfaces ideas but never proposes a trade. When something is actionable, it goes to whichever specialist owns that asset class — Sage for stocks, Cleo for crypto. They write a specific proposal: a ticker, a side, and a dollar amount.
The proposal then hits a rule engine written in ordinary code. This is deliberate. Language models are unreliable at “never” — you can talk them out of a rule they were merely told to follow. A check that runs in code cannot be reasoned with.
These run on every proposal today:
Anything that clears the rule engine goes to Roan, the risk agent, who can approve it, amend the size downward, or veto it. An amendment can only ever shrink a trade, never grow one.
Below your thresholds, small trades fill automatically. Above them you get a card with the reasoning and a yes or no. Any trade of $5,000 or more requires your explicit approval, as does every stock trade and everything in real-money mode. That threshold is not a preference setting — it is the line the design is built around.
Every decision writes a record: what was proposed, who proposed it, what the risk review said, what happened, and the reasoning behind it. Each record is stamped with the exact model version and a hash of the exact prompt text used, so a decision made months ago can be reconstructed rather than guessed at.
Those records cannot be edited or deleted. That is enforced by the database itself, not by application code that could be changed later. If we altered a recommendation after the fact to make it look better, the write would be rejected.
“Our AI manages your portfolio” is not a checkable statement. What is checkable is whether limits are enforced where a model cannot override them, whether a human approves the consequential moves, and whether there is a record you can go back and read. Those are the parts we can show you.