if an AI can spend crypto, its refusal log matters more than its win rate

Most trading bots are judged by the trades they take.

I think the more important record is the list of trades they refused, and the exact information they used at that moment.

A bot can follow every configured rule and still make a bad decision if one input is stale, one threshold changed without being versioned, or a transaction moved from pending to expired while the interface still treated it as live.

That creates a strange accountability problem. A trader can see the final PnL, but not necessarily answer the basic questions behind it.

• what data did the agent see?
• which checks passed and failed?
• which configuration version made the decision?
• was the transaction confirmed, expired, dropped, or simply unknown?
• did a human change the strategy before or after the trade?

For me, a useful AI trading system needs three separate records.

1. A decision receipt for every accepted and rejected setup
2. A transaction receipt that distinguishes submitted, confirmed, expired, and unknown states.
3. A version history that shows exactly when risk limits changed.

The refusal log may actually be the most valuable of the three. It exposes whether the system is being disciplined or just inactive, and it makes false confidence harder to hide behind a clean equity curve.

I am building TrenchersAI, so that is the context behind this question. I am not presenting it as an independent review, and I am not linking the product here. The broader issue is bigger than one tool.

What would you require in an audit trail before allowing an AI agent to control a crypto wallet?