AI Mini-Series Part III: From Votes to a View: How AI Has to Be Used Differently in a Systematic Fund

Part 3 of a series on building and testing an AI layer for our systematic framework. This one is the hinge: the difference between reading the Five Forces with rules and reading them with a language model, and why that difference is the entire point of the exercise.

The first piece of this series argued that the AI layer is a second, independent reader of the same Five Forces our deterministic strategies already read. The second piece was about the data that reader is allowed to take in. This piece is about the question those two raise and neither answers: once the AI has read the Forces, how does it actually reach a conclusion, and how is that conclusion different from the one our existing systems already produce? This is the hinge of the whole project. If we get it wrong, the effort collapses into an expensive way to restate what we already know.

How the deterministic book decides

Our deterministic strategies reach a call through a voting protocol, which we have written about before and which has always sat at the center of how the firm’s strategies work. Each Force produces signals, and each signal is, in effect, a vote, with multiple signals aligning to shape directions and magnitude. The votes aggregate, and a position falls out the other side: more constructive when the weight of the votes is constructive, more defensive when it is not. There is no separate singular score for the market sitting above all of this. The call is the sum of the votes. We do compute standing market scores on the deterministic side, including a pair of regime measures we have validated and watch daily, but they enter the book the way every input does, distilled into votes and guardrails. Nothing scored sits above the count.

That design works because every input has already been turned into a clean number with an unambiguous vote before it arrives. A moving average is above or below its level. A spread is widening or tightening. Counting those votes is mechanical, repeatable, and free of mood, which is the entire reason we trust it. Nothing about the AI layer changes how the deterministic book votes, and nothing is meant to.

Why a language model does not fit that mold

The trouble starts when you try to make the AI vote the same way, because the thing the AI is for does not reduce to a vote. Its job, from the first piece, is to read context: the language of a policy statement, the meaning of a surprise, the difference between a scare that fades by the close and one that reprices a quarter. When it reads that material, what comes back is reasoning, expressed in language, and reasoning expressed in language is not a decision. A paragraph, however sharp, cannot size a position or set exposure, and it cannot be combined with anything else the system produces. Our current deterministic system was spared this problem because its inputs arrived already scored. The AI hands you the opposite situation on the first day: a rich, contextual reading, and nothing structured at the end of it.

We need a different solution, then, and its shape follows directly from the problem. We do not want the AI to tally votes the way the deterministic book does, for two reasons. Its reasoning does not reduce to a vote in the first place, and even if we forced it to, re-tallying the same votes would tell us nothing we did not already know. What we want instead is for the model to commit to an actual read, and to express that read as a small set of defined numbers the rest of the system can use.

So that is what it produces, and it is deliberately not one number. The two scores sort the Forces by how fast they move; a Regime Score and a Composite Score. The regime score is the slow read, the macro backdrop set by the business cycle and by where corporate earnings and fundamentals stand. It turns over in months and quarters, because the forces behind it do. The composite score is the faster read, the near-term lean set by the Forces that move day to day: the action in price, the swings in sentiment and crowd psychology, and the positioning of options dealers. Between them, the two scores account for all five Forces, divided by how fast each one moves rather than thrown into a single pile. Around those, the model states a confidence factor, for how much conviction the read deserves, and a severity factor, for how extreme conditions are. There is no single magic figure that we can create that solves investing and that’s not the goal. Each of the four is a defined quantity with its own job, and together they tell the story in numbers rather than prose. That structure is the point: a single number would hide its own uncertainty, while a read that has to state its confidence and its severity next to its direction is forced to show how sure it really is.

The deterministic book never had to solve this. The AI creates the problem the moment you try to use it, and a great deal of the engineering in this project goes into solving it well.

The trap of the agreeable narrator

There is a tempting wrong way to build this, and it is worth describing because it looks like progress. You take the same signals the deterministic book already votes on, you hand them to the model, and you let it write a thoughtful narrative that arrives at a score. It reads beautifully. It also adds nothing. Same inputs, same underlying logic, better prose, same answer. You have spent a great deal of effort reproducing a conclusion you already had, with a confident essay wrapped around it.

This is the failure I worry about more than a hallucination, because it is invisible. The system looks like it is doing something new while quietly doing the old thing twice. You have created a second voice that agrees with the first by construction. Two readings guaranteed to agree are one bet counted twice, with the second copy dressed up to look independent. That is the very opposite of what the first piece promised, where the entire value of the second reader came from its independence.

The requirement that follows is strict, and it is the heart of this article. The AI’s view has to be built from what the votes never see. It has to draw on the contextual, language-based information the deterministic signals cannot represent, and it has to be free to reach a different conclusion than the vote count does. A second opinion that can only ever agree is not a second opinion. If the AI’s read is just a function of the same signals we already tally, it is redundant, and redundant is worthless here, however well it is written.

A read of its own, which then has to be trusted

This is why a read of this kind exists on the AI side and nowhere else. It is the model’s own structured view of the same Forces, reached by reasoning over context the votes cannot use, and deliberately built so that it can disagree with the vote count. When the two readings agree, the agreement means something, because they reached it by different roads. When they disagree, the disagreement is information, the kind the first piece called the raw material of real diversification.

Giving the model that job creates the hardest problem in the whole effort, and it is the subject of the next piece. The moment you ask a language model for a score instead of a paragraph, you hand it the power to be confidently, structurally wrong. A score carries an authority a paragraph does not, and people act on it. So the challenge becomes making that scored view trustworthy enough to stand beside a system we have trusted for years. Trustworthy, as the next piece will argue, is mostly a matter of what you refuse to let the model do.

Blackworks Capital LLC manages funds through Blackworks Capital Management LLC, an Exempt Reporting Adviser. Nothing here is an offer or solicitation or investment advice. The systems described are in research and paper-testing and do not manage client capital. Past performance does not guarantee future results.