FROM THE TEAM

What happens when the AI is wrong

It will be wrong sometimes. What matters is what the system does about it: a risk reviewer that can only shrink trades, receipts rendered by code rather than a model, and a record you can read afterwards.

By the Vestya team · Published July 29, 2026

It will be wrong. Any honest page about an AI investing product has to start there. Language models produce confident text whether or not they are correct, markets humiliate confident forecasts of every kind, and no amount of engineering changes either fact. What engineering canchange is what a wrong answer is allowed to cost you. Vestya's design assumes error and works backwards.

Layer one: a reviewer that can only say no

Every proposal that clears the deterministic rule engine goes to Roan, a separate risk agent whose entire job is to push back. Roan can approve a trade, veto it, or amend it — and an amendment can only ever shrink a trade, never grow one. That asymmetry is enforced in code: even if the risk model itself misfires and suggests a larger size, the system clamps to the smaller of the original and the amendment. A malfunction in the safety layer fails toward caution, not toward exposure.

Layer two: the boring parts are not written by a model

A subtle way AI products go wrong is hallucinated bookkeeping — a model narrating a trade that didn't happen, or misquoting a price that did. So in Vestya, receipts are not the model's job. When a trade fills, fails, or is declined, the confirmation you see is rendered by ordinary code from the actual database rows. The conversational agent can phrase a greeting; it does not get to improvise your account activity.

Layer three: consequence requires consent

Anything at $5,000 or above, every stock trade, and everything in real-money mode waits for your explicit approval. A wrong idea that needs your signature is an email; a wrong idea with autonomous execution is a loss. The approval gate converts the worst class of error into the survivable class.

Layer four: the record survives the mistake

When something does go wrong, the question that matters is what exactly happened? Every decision — including the bad ones, including the vetoes that should have been approvals — sits in a permanent record stamped with the model version and instructions that produced it. We cannot quietly bury a bad call, which has a useful side effect: we are forced to learn from specifics rather than impressions.

What we deliberately do not promise

No loss protection, no accuracy percentage, no “our AI beats the market.” Anyone quoting you those numbers for an investing product is selling something they cannot substantiate. What we can promise is structural: errors are contained by limits, reviewed by an adversary, gated by your consent, and recorded beyond our ability to revise. That is what being wrong safely looks like.

Not investment advice. Investing involves risk, including loss of principal. Paper-trading results are hypothetical and for educational purposes only.

Can you trust AI with your money? · Meet Roan, the risk agent