AI Mini-Series Part VII: Hindsight at Machine Speed: Why the Council Does Not Learn From Its Own Trades
Part 7 of a series on building and testing an AI layer for our systematic framework. This one is about the most tempting idea we have examined since the build began: letting the AI learn from its own history.
I ended the last piece by saying the work from here is to keep finding the next thing to doubt. This piece is about one that volunteered itself, and it arrived looking like a gift.
Here is the gift. Every trading day, the Council leaves behind a full day of its own thinking: the reads it formed, the deliberation behind them, the dissent, the reports, and then, later, the grades, what actually happened at one horizon after another scored against what the system said would happen. Stack that up for months and you are holding something most firms would love to have about their own decision process, a complete and honest record of what a manager believed, when, and why, with the outcomes attached.
The idea that walks in behind a record like that writes itself. Feed it back. Let the system study its own history, keep what worked, discard what did not, and get a little wiser every day. Give the Council a second brain. The pitch is compounding, and compounding is the one force everyone in this business is trained to respect. I wanted it to be true the way you want any free lunch to be true, and the fact that I wanted it is exactly why it went into the same process everything else goes into here. Much of our research is AI-assisted now, and we pointed that research at the question in the most direct form I could phrase: who has made this work, and what happened when they tried?
Two shapes, and neither is a loop
Start with who has made it work, because an idea this obvious has been available to every large fund and every well-funded lab for a while now. If a self-improving trading memory worked, someone credible would be running one, and given how loudly this industry markets its AI work, someone would be saying so.
We could not find one. No fund we could identify, including the largest and most public about their AI ambitions, has shown a working version of the loop: a model that learns from its own trade history, updates its own beliefs, and gets better because of it. What we found instead, everywhere the work was real, were two other shapes.
The first shape is a curated memory with receipts. Morgan Stanley built an assistant for its financial advisors on exactly this pattern. The model retrieves only from a corpus of documents that human experts vetted, every answer cites the source it came from, and the system is graded against a fixed set of golden questions, re-asked with every change, so that a change that makes it worse is caught before an advisor ever sees it. The executive who led the build summed it up in a sentence I have not been able to improve on: “All the boring stuff that nobody cares about is what makes this successful.”
The second shape is idea generation behind a hard gate. Man Group built a system called AlphaGPT that proposes trading hypotheses, and the discipline is all in what happens next: each hypothesis is frozen before it is backtested and is never edited after the results come in. The machine is allowed to suggest. It is not allowed to look at the answer and then revise the question, which is the exact move that turns research into curve-fitting.
Notice what the two shapes have in common. In both, the machine is held one step away from its own beliefs. It can retrieve what humans curated. It can propose what tests will judge. What it cannot do, in either shape, at any firm we could find doing serious work, is hold the pen on its own permanent record.
What happens when a model keeps its own diary
Absence of evidence is suggestive, and it is not proof. Maybe everyone is simply early. So the second question was what happens, measured in controlled conditions, when you hand a model the pen anyway.
That experiment has been run, and the results are not close. Agents given an unfiltered memory of their own past work performed worse than the same agents given no memory at all. The mechanism matters more than the headline, because it is not a bug that will get patched away. A language model imitates what it retrieves. Researchers studying these systems call the behavior experience-following: show the model its own precedent and it follows the precedent, with very little regard for whether the precedent was any good. A stored success becomes a template, and so does a stored error. An unfiltered memory does not just preserve mistakes; it propagates them into new decisions, which is close to the opposite of learning.
It gets worse when the model is allowed to maintain the memory itself. In one documented case, an agent permitted to rewrite its accumulated playbook compressed 18,282 tokens of strategy, tokens being the units a model reads and writes in, down to 122 tokens in a single rewrite. It kept the cover page and shredded the binder, confidently, as an act of tidying up. A memory is also an attack surface: security researchers have shown that an agent’s long-term memory can be poisoned through ordinary-looking inputs, planted records that quietly steer decisions long after the planting.
The self-learning loop does compound, and that much of the pitch is honest. It compounds whatever falls into it, without first checking what that is.
The strongest case on the other side
An honest survey has to include the best evidence against its own conclusion, and here the best evidence comes from AQR. Cliff Asness spent years as one of the loudest voices against overfitting in quantitative work, and his firm now lets machine learning set factor weights that people used to set. He has described the machine’s work as “annoyingly better than me,” and he has been candid that the change is uncomfortable and sits in tension with the firm’s long discipline. When a person with that history reports that the machine earned the pen, the report deserves weight rather than a dismissal.
Look closely at what was handed over, though, because the difference carries the whole argument. Machine-learned factor weights are fit on enormous samples, long histories across wide cross-sections of securities, inside a validation regime built by people who spent their careers hunting overfitting. The machine is learning from data deep enough to push back. That is a different activity, in kind and not only in degree, from a language model reading its own last dozen graded calls and writing down what it believes it learned. The first is estimation at scale under adversarial testing. The second has an older name, and the name comes from the poker table.
Lessons that write themselves
Poker players call it resulting: judging the quality of a decision by how the hand happened to end. Annie Duke built a book around the error, because it is the default way human beings process a small number of high-variance outcomes. Fold a strong hand, watch the bluff get shown, and the lesson your brain writes down is that the fold was a mistake. It was not. On a dozen outcomes in a game with real variance, grading decisions by results manufactures lessons that are worse than no lessons, and every one of them arrives wearing the feel of hard-won experience.
Hindsight does the bookkeeping for that error. Once you know how a decision ended, your memory of why you made it quietly rewrites itself, which is why the classic prescription in decision science is a journal: write the reasoning down at decision time, seal it, and grade the sealed version later, because the version in your head after the outcome can no longer be trusted. The prescription exists because paraphrase is where hindsight lives.
Now hand that job to a language model. Ask it, each night, to read its last dozen graded decisions and extract the lessons. It will produce lessons. It will produce them whether or not any exist, because producing a fluent answer is the one thing it is built to do, and it will produce them from exactly the kind of small, noisy sample that behavioral finance has spent decades warning humans about. The loop everyone imagines is a machine that learns from experience. The loop you actually get is resulting, automated, run nightly, in confident prose.
We did not have to take the base rate on faith, because we had already measured it on ourselves in a different costume. When we ran our own candidate regime signals through the validation battery the fifth piece described, block-bootstrap confidence intervals, walk-forward testing on held-out history, a correction for how many ideas we tried, the result was zero of fifteen. Fifteen candidates, each one plausible, each one with a story good enough to repeat in a meeting, and none of them survived. That is what a pile of attractive lessons usually is: noise wearing the costume of experience. If our own carefully built hypotheses clear the bar at that rate, the nightly conclusions of a model grading its own trades will do worse. A brain that stores noise does not make the Council smarter. It makes it confidently, cumulatively dumber.
The memory the Council already has
While the research was stacking up reasons to refuse the loop, it kept bumping into something I had half stopped seeing, the way you stop seeing furniture. The Council already keeps a memory. We built it piece by piece, under the same discipline as everything else in this series, long before anyone here used the phrase second brain.
Its reasoning is frozen at decision time. Every read the system produces is written down in full before the outcome is known, which means the Council has kept, from the start and by construction, exactly the sealed journal that decision science prescribes for people. Every decision is graded after the fact at horizons from one trading day out to ninety, so a call is judged at the tempo it was made for. Each night the Chair, the agent that writes the final read, opens with its own record in front of it: its last ten graded decisions, its hit rate Force by Force, its accuracy in market regimes like the current one. The full deliberation is preserved append-only, which means the record can be added to and never edited. And the agents draw their reference knowledge from a library that is tiered and curated rather than scraped and hoped for.
The Council, in other words, remembers a great deal. What it cannot do, and has never been able to do, is write its own beliefs unsupervised. It holds a complete record and almost no license to draw conclusions from that record on its own authority. Set that against the survey and the shape of the answer appears: the two patterns the survey found working elsewhere, a curated library with receipts and self-produced ideas held behind a gate, are the two halves of what we already run. The research question was never memory versus no memory. It was who holds the pen.
Who holds the pen
The design we chose keeps the pen human, and it fits in five rules.
Anything the AI drafts about its own performance passes a human signature before it touches the permanent record. The model can propose a lesson. A person signs it in, or it does not go in.
Post-mortems may quote the frozen decision-time reasoning, and they may never paraphrase it after the outcome is known. Quotation preserves what the system actually believed at the moment of the call. Paraphrase is the door hindsight walks through, so the door stays shut.
Every candidate lesson carries its sample size and its evidence status on its face, and nothing under-evidenced is allowed near a live decision. A lesson graduates the same way a signal graduates, by surviving validation, and zero of fifteen stands as a permanent reminder of how often that happens.
Beliefs are never deleted. They are invalidated, with a date and a reason, so the record keeps what we used to believe next to when we stopped believing it. An audit trail of being wrong is worth more to me than a tidy story of always having been right.
And new memory reaches the Council shadow-first, watched before it is trusted, like every other capability this series has described.
That is the second brain we are building, and it is boring on purpose, in exactly the sense the Morgan Stanley executive meant. The self-improving loop is the better story, and I understand why it is the one people want to tell. We will keep watching for evidence that someone has made it work safely, and if that evidence arrives we will test it with the same suspicion we point at everything else, and hardest at our own record. Until then, memory earns its way into the loop the way everything else here does, by passing the test, and not a day sooner.
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.