The smartest tool in the system has exactly one job: doing less.

An LLM is a next-token predictor trained on human text. It has no memory of your position, no access to your margin, and no consistent definition of "risk." Ask it the same question twice with a slightly different preamble and you may get opposite answers. That is not a flaw to patch — it is the fundamental character of the architecture. The question is not whether to use it, but what role it is structurally permitted to play.

The trap is flattering and common. A developer watches a language model correctly summarize a central bank statement, correctly identify a project as a known exit scam, correctly flag a liquidity anomaly. The model sounds like it knows what it is doing. So the developer gives it a signal weight. Maybe a small one at first. The model now influences when the system enters a trade. The problem is that you cannot backtest a language model the way you backtest a rule. Its outputs are non-deterministic — the same input does not reliably produce the same output — and its reasoning is not inspectable at the weights level by the people deploying it. When it is wrong, you cannot run the tape back and find out why. You have introduced an un-auditable variable into the entry side of your book, and you will not know it is killing you until the drawdown is already there.

What the v3 build exposed

WiseBot's v3 systematic two-sleeve book was designed around a backtested carry-trend blend weighted roughly 0.6 carry / 0.4 trend. The backtest Sharpe came in at approximately 0.93 against a buy-and-hold Sharpe of 0.62 over the same period, with a max drawdown near -21% versus buy-and-hold's -59%. Those numbers gave the team confidence in the rule-based core. During design review, one proposal put on the table was to let the LLM layer, which was already being used for news parsing, contribute a small positive signal weight — essentially a soft vote to open or size up a position when the model flagged a macro tailwind. The proposal was rejected, and the rejection is worth explaining plainly: a signal that cannot be backtested cannot be sized. If you cannot run it through the same historical data that produced that 0.93 Sharpe and see what it does to the distribution, you do not know whether you are adding edge or adding noise dressed as edge. The math does not care that the model's summary sounded smart.

What the LLM is permitted to do in v3 is narrow and explicit. It monitors an incoming news feed for a defined class of events — protocol exploits, regulatory bans, exchange halts, severe macro dislocations — and if it fires on one of those categories above a logged confidence threshold, its only available action is to cut exposure or force flat. It cannot open. It cannot add size. Every trigger is written to the log with the raw input, the model output, and the action taken, so the record is auditable after the fact even if the model's internals are not. The architecture treats it as a circuit breaker, not a signal generator.

The asymmetry matters. If the LLM hallucinates a crisis that does not exist, the cost is a missed opportunity — you were flat when you could have been long. That is recoverable. If the LLM hallucinates an opportunity that does not exist and you have given it entry authority, the cost is a position opened on fictional reasoning with no backtest defense. That category of error is not recoverable in the same way, and in a leveraged book it can be catastrophic. The brake-only constraint converts the model's worst failure mode from "opened a bad trade" to "closed a good one too early." One of those is a tuning problem. The other is a strategy problem.

The portable rule: if a component of your system cannot be backtested, it is only safe to wire it to the exit, never to the entry.

The v3 architecture decisions, including the LLM constraint, are documented in the whitepaper and cross-referenced against the on-chain record at the-wisebot.com — there so the logic can be checked, not taken on faith.