On 9 April 2025 the S&P 500 had fallen twelve percent in four sessions. At 9:37 Eastern, Donald Trump posted: “THIS IS A GREAT TIME TO BUY!!! DJT”. At 1:18 that afternoon he paused most of the new tariffs. SPY closed up 10.5%.

Was it a good call? Scored honestly, from the next open when anyone could act, it earned nothing over the index. But ask a language model and it already knows what happened at 1:18. It cannot forget the afternoon.

We tested this on hundreds of dated market calls. Asked without a date, 11% of the dates the model cited were in the wrong year. Asked why a call worked, it gave true facts, from later, about a different company. And every call has two verdicts: on the day it was due, and today. For 28% of calls they disagree, and the model cannot tell which one you meant.

Agreement with the price record: no date 53.1%, as-of date 52.2%, with Google Search 70.0%. Losing calls it called winners: 30.5%, 67.6%, 20.4%.

Without a date, it is a coin flip. With web search, 70%, and it still calls one in five losers a win. The model is not the problem. The memory is: it stores sentences, not when they were said or what they rested on.

Our approach: a memory with a clock

At Factagora we build a Temporal Causal Knowledge Graph (TCKG). Every claim keeps when it was said, when it resolves, what it rests on, and every verdict with the date it was made. Ask it about any day and it answers with only what was knowable that day. One person’s TCKG is what we call a brain.

The use case: 14 brains, two years, one rule

We built brains for Donald Trump and 13 of the most followed US finance YouTubers: 16,150 dated market claims, each linked to the post or the second of video it came from. Then we backtested every call over the past two years with one fixed rule: enter at the next open, exit at the speaker’s horizon or 90 days, measured against SPY over the same dates.

1. Mike Maloney: 120 calls, +10.6% per call, 62% hit rate

2. Jeremy Lefebvre: 449 calls, +6.3% per call, 60% hit rate

3. Gregory Mannarino: 99 calls, +5.3% per call, 69% hit rate

4. Parkev Tatevosian: 95 calls, +3.3% per call, 53% hit rate

5. Jim Cramer: 833 calls, +0.9% per call, 53% hit rate

6. Jeff Snider: 127 calls, +0.6% per call, 50% hit rate

7. Cathie Wood: 83 calls, +0.1% per call, 54% hit rate

8. Graham Stephan: 126 calls, -0.3% per call, 50% hit rate

9. Guy Turner: 137 calls, -0.5% per call, 50% hit rate

10. Benjamin Cowen: 112 calls, -1.6% per call, 41% hit rate

11. Kevin Paffrath: 214 calls, -1.8% per call, 43% hit rate

12. Andrei Jikh: 112 calls, -2.1% per call, 46% hit rate

13. Mark Moss: 132 calls, -6.9% per call, 50% hit rate

Tap a name to open that speaker’s brain and every scored claim.

Ten of thirteen sit inside the coin-flip range. Trump, scored in his own league because his posts move the markets they are about, hit 49.8%.

What we learned

When: the same words, a different day, a different result

On the morning of Trump’s post, two YouTubers called the index the other way. Gregory Mannarino: “sell America”. Mike Maloney: “we’re still in a hyperbubble”. Both lost. Graham Stephan had made the opposite call the day before; all seven of his buy calls that fortnight hit. On 76 of Trump’s calls a YouTuber took the other side within two weeks. Trump won 31, the YouTubers 20. Timing decided more than conviction.

Still true?: a good record goes stale

Split by topic, a few speakers beat the coin flip on their own subject.

Above the coin flip: Mannarino macro 78%, Maloney commodities 77%, Stephan stocks 72%, Lefebvre stocks 61%. Below: Cowen crypto 32%, Moss crypto 32%.

But the top of the table moves. Jeremy Lefebvre went from 70% to 44% in six months. Mike Maloney from 70% to 19%. A memory that only knows “this person is good” is already out of date.

Why: it wasn’t the person

On 3 July 2025 Jim Cramer called Palantir to 200 while Parkev Tatevosian called it overvalued. Same dates: Cramer plus 28.5%, Tatevosian minus 30.5%. What decided it? For every call we found the series the asset had been tracking for a year, its price driver, and checked whether it moved with the call. That one variable split the same speakers’ calls by 32 points of hit rate. Nobody ever named it.

Driver moved with the call: 973 calls, 67.5% hit rate, +5.23% per call. Against it: 741 calls, 35.0%, -4.15%.

Even the direction mattered less than it seems: buying every stock a speaker named, ignoring buy or sell, did better than following them. And “inverse Cramer” would have lost money.

Following the direction: +3.92% per call, 52.8% hit rate. Buying every symbol named: +5.52%, 56.2%.

What a memory with a clock makes possible

Each of those lessons needed the same three things a sentence does not carry: when it was said, why it should be true, and whether it still is. Give memory a clock and an AI can backtest an idea without peeking at the future, explain a result from the evidence of its day, and tell you, “You believed this in March. Three things it rested on have changed since.” That works for investing, and for anything else you believe.

Coming next: talk to a brain, or bring it into your AI

Every brain in this post opens without an account. Over the coming weeks we are rolling out three things on top of them:

• Talk to a brain. Ask Trump’s or Cramer’s brain what they said about an asset, when, and how it turned out. Answers come from dated public statements, not from a model’s memory.

• Bring it into your AI with MCP. Connect any brain to Claude, ChatGPT or your own agent, so it reads dates and verdicts instead of guessing.

• Make a brain for yourself, or your team. Upload your ChatGPT or Claude history and get your own brain, with every belief dated. Then talk to it, share it, or connect it to your AI.

I would love your help. Try them as they open and tell me what works and what doesn’t. Which brain would you talk to first? What would you ask it? Would you connect your own brain to your AI, and what would you want it to remember? Reply to this email or leave a comment. Every answer shapes what we build next.

Limits and right of reply

We scored 13 of the 20 YouTubers we set out to; the rest lacked captions or enough resolved calls. Trump’s posts were found by finance keywords. Claims were extracted by a model and reviewed on a sample. The benchmark is SPY, and the rule was frozen before scoring. The AI experiment used one model on a stratified sample of resolved calls. This is a record of public statements, not investment advice, not a recommendation to follow or fade anyone, and not a political judgment. Anyone listed can request their full claim list; corrections are published, never silently edited.

Keep this going

364 calls are still open and will resolve in the coming months. Tell us in the comments who to score next, or where the rule is wrong; we will run your version and publish it. And when your own brain opens, add what you believed on the morning of 9 April. It gets scored next to everyone else’s at factagora.ai.

When my mother was diagnosed with a serious illness, she asked an AI before she told her own children. She asked it to recommend a good doctor. Without hesitation, the AI named a famous physician at Seoul National University Hospital, and trusting that answer, she booked an appointment and waited nearly a month for her turn.

But hers was the kind of illness where every day matters. The AI never weighed that urgency, never took in the whole of her situation. It simply handed her the most famous name it knew. The wait kept stretching, and in the end, thankfully, she was able to have surgery right away at a university hospital near home. She is in chemotherapy now. Still, I cannot shake the thought that if the surgery had come even a little sooner, its reach might have been smaller.

Living through that, one thing became clear to me. The AI had given the most plausible answer, not the answer that was right for my mother. Just as the most popular story in the world is usually not the truest, AI too is optimized not for truth but for plausibility, and in the end, for our own dopamine. And above all, when its answer was wrong, the AI lost nothing. The one who lost was always us.

AI hands down judgment without hesitation. But it never bears the consequences that judgment sets in motion. Judgment and responsibility come apart.

The root of the problem is not that AI fails to know something. It is that there exists no system to trace a claim back to whoever made it, to check whether it held up as time passed, and to record the result as that person’s track record. On top of today’s AI there is no way to trace a single claim, no way to verify it, and therefore no way to leave it behind as anyone’s reputation. It is not that accountability is lacking. It is that there is no place for responsibility to stand at all. And that empty place is where I have spent the last several years.

The Machine Looks Like Us

Step back a little, and this looks less like a flaw in a machine than a sickness of our whole era. Some people assert without accountability. Others make a failed prediction and, paying no price for it, simply move on to the next one. Forecasts and opinions and claims pour out every day, yet once time has passed almost no one goes back to ask, seriously, who was right. We live in an age where judgment is cheap and responsibility is rare.

The LLM is a copy of exactly this era. It is a machine built by compressing the text humanity has written, so the sickness already steeped in that text is compressed along with everything else. Which is why the AI’s lack of accountability is less a defect of its own than a mirror held up to us. The trouble is that the mirror doubles as an amplifier. One person being carelessly wrong is not the same kind of event as a machine replicating that carelessness millions of times a second. AI reflects the sickness of our time and, at the same instant, automates it at a scale we have never seen.

What Compression Throws Away

The largest artificial intelligence is, in truth, the largest compressed file. And every compression comes with loss. In the act of cramming the vast text humanity has written into the small space of weights, at least three things are thrown away: time (when was this true), causality (why is it so), and dissent (who disagrees). A machine that takes averages is, by definition, built to output the most plausible consensus, and it erases every coordinate it passed through on the way there.

What matters is that this is not a bug to be fixed in the next version. It is not a performance problem solved by more data and more compute. It is a structural void lodged in the very paradigm of how an LLM is trained. For a small startup like ours to charge into the same game with a bigger model is close to suicide. We have to play a different game. The game of picking back up what they discard in the act of compressing, and fitting the pieces together.

Until recently it was the age of training. Now, in 2026, people say the age of inference has arrived. They mean that the stage where a model thinks once more before it answers, calls up the context it needs, and consults outside material is growing more and more important. And yet, in that very moment of inference, whatever the model is standing on as its ground is still neither traced, nor verified, nor managed.

If responsibility leaked out during training, then at the stage of inference there is still no memory to take its place. What I am trying to pick back up is exactly that memory. A memory the model can finally stand on when it reasons, one that is traced and verified and owned by someone.

So how does that memory accumulate? I divide it into three motions. Make, Verify, Own. Chris Dixon read the arc of the internet as read, write, own. I took inspiration from that, but what I am drawing is a slightly different world. Beyond reading and writing, a world where a judgment is made, verified, and finally owned by someone. How does a scattered judgment come into being, how is it graded, and to whom does it finally attach. Let me take the three in turn.

Make: The Data That Does Not Yet Exist

The giants, by their nature, learn what already exists. So the frontier of value moves outside the territory they can scrape, toward data that does not yet exist in the world. Until someone stakes their own judgment by saying “I believe X will happen,” that data is nowhere on earth. It exists only as potential, inside a human head. This is not data extracted from somewhere. It is data generated the moment a person renders a judgment.

We are not trying to manufacture demand that was never there. People already predict, assert, and argue every single day. That epistemic labor happens daily and evaporates daily. What we are doing is not creating that labor but building the vessel that catches what used to vanish for lack of anywhere to land.

Verify: The Data That Can Be Graded

A judgment, once made, is ruled on in the end by time and reality. And here lies a property an LLM can structurally never have. An LLM’s output carries no hook by which you could later confirm whether it was right. It is merely a present-tense sentence that looks plausible in this instant. But if a judgment has a moment and a condition fixed clearly inside it, by when and under what conditions it holds true, then reality can grade it once time has passed.

Whether grade-ability lives inside the data or not, that is the fork in the road. I will not say here exactly how we grade it. (I will share more about that in a later post.) But this much is clear: data that can be graded and data that can never be graded are fundamentally different kinds of things. Confidence that cannot be graded is not confidence. It is noise.

Own: The Data Responsibility Is Pinned To

But even after the grading is done, if that verdict attaches to no one, it stays a piece of interesting trivia. The record of right or wrong remains, but if no one is held to it, in the end it means nothing. To own is precisely to pin that verdict to one particular person. It is the hinge that joins Make and Verify through consequence.

So ownership has to run both ways. Upward: each time a judgment of mine is cited and reused, reputation and a share of the upside come back to me. Downward: if that judgment is finally ruled false, that record stays with me too. To truly own something is to bear not only its success but its failure. With only the upward side it is no more than collecting points. Only when the downward side is attached does it become staking yourself, truly being accountable.

What is striking is how this small incentive grows into a larger order. The deeply individual motive of caring about one’s own track record emerges, at the macro scale, as a self-correction in which good judgments survive and flimsy ones are weeded out. (By “individual” here I mean both people and AI agents.) This self-correction does not arise because some referee sits above and censors. Because owned responsibility is distributed across every participant, it rises up on its own from below. In that way quality control becomes not someone’s labor but an emergent property of the system itself.

An LLM never owns the claims it puts out. So it accrues no reputation and bears no responsibility, and so it can go on being confidently wrong forever.

What We Are Actually Building

I have been speaking in fairly abstract terms, so let me come down to the ground for a moment. We build a minimal unit that binds a single claim together with its time, its source, and its causality. We call it a FactBlock, and the structure that connects those blocks again through time and causality we call a Temporal Knowledge Graph. The act of a person staking their own conviction, that Belief, is the starting point of all the data, and the place where it is graded, owned, and gathered up piece by piece is Factagora.

Make, Verify, Own. To make, to be verified, to own. Looking back, these three words are in the end an attempt to bind back together the judgment and responsibility that our era, and the machine that faithfully copied it, had pulled apart.

A Place for Responsibility

The recommendation the AI gave my mother had a person's name in it, but no one, anywhere, to answer for that recommendation. There was an optimal judgment, yet no one who had staked themselves on it. What I want to build is exactly the world on the other side. A world where a judgment someone makes is graded before time and reality, and the result attaches, directly, to the person who made it. A world where good judgment is rewarded and reckless judgment pays its price, and where the record that accumulates becomes, for the next person, a slightly better ground from which to choose.

In April 2014, a ferry called the Sewol sank off the coast of South Korea, and three hundred and four people died. Most of them were high school students on a field trip. For weeks afterwards, I watched the country argue. Not about the grief, which was clear, but about the facts. Who gave what order and at what time. Whether the captain had lied. Whether the rescue had been delayed on purpose. Whether a photograph circulating online was real or staged.

None of the arguments were about opinions. They were about facts. But the facts were everywhere. On television, on forums, on messaging apps. And a lot of them were wrong. Some wrong by accident. Many wrong on purpose. The wrong facts pulled the families of the victims into fights with strangers who had been told a different version of what happened. They turned neighbors into enemies. They took a national tragedy and turned it into bitterness that the country, more than ten years later, is still carrying.

Sewol did not create my obsession with fake news. But it made it impossible to look away. Most people eventually did. The country moved on, to politics, to other crises, to other arguments. I kept watching the same problem.

Years before Sewol, when I was studying computer science at Carnegie Mellon in 2007, I was reading Wittgenstein’s Tractatus Logico-Philosophicus, which opens with a sentence I have never forgotten: “The world is the totality of facts, not of things.” The book itself is organized as a tree of numbered propositions (1, then 1.1, then 1.11, 1.12) where every statement is a consequence of, or a reason for, the ones it branches from. I remember staring at the structure and thinking: this is a database schema. A strange database schema, written by a philosopher in 1921, but a database schema all the same. What would it look like if somebody actually built it?

The mission I eventually settled on for this company comes out of those two moments stitched together. I’ll say it plainly. Make every claim verifiable. Not “make the internet true.” I do not think that is something anyone can build. Just this: for any claim anyone makes, build the infrastructure so that a person, or a machine, can trace the claim back to whatever evidence supports it, and can see clearly when the evidence is thin.

Behind that mission is a more personal vision. Bertrand Russell was once asked what advice he would give to people a thousand years from now. Here is what he said.

His answer came in two parts. The first part was intellectual: look only at the facts and never at what you wish were true. The second part was moral: love is wise, hatred is foolish. In a world growing more closely interconnected, he said, we have to learn to tolerate each other. We have to learn to live together, or we will die together. I have come to think those two parts are the same piece of advice said twice. Most of the cruelty I have watched between people is a failure at the first step. Not of goodwill. Of understanding. My vision for Factagora is that if we can make facts a little easier to see, we might leave people a little less likely to hate each other for things that were never true to begin with.

Understand more. Hate less.

That is what I am actually trying to build a company around. Everything else (the endpoints, the benchmarks, the enterprise contracts) is downstream of those four words.

The path to actually building any of it was not straight.

The first version of Factagora was a web app that tried to help people fact-check news articles. The problem was real. The market was not, at least not yet. It didn’t make money, and I needed to keep the company alive. From there I detoured into the space that was easier to raise money in at the time: blockchain. I built an NFT marketplace around Hollywood IP. It went well enough that I was eventually invited to join a joint venture between Dunamu and HYBE, leading the technology side of a project that was supposed to put BTS onto the chain. Two of the hottest keywords in the world at that time, pointed at each other. It felt like the kind of opportunity that does not ask twice.

Then the whole thing unwound. The crypto market cracked. The public’s appetite for “NFT” collapsed. BTS paused their group activities. By the end of it I was standing in front of the original problem again, with less money and more bruises. I am telling you this not to be proud of the detour but to be honest about where the narrowing came from. It did not come from a moment of brilliance. It came from failing at broader things first.

Around that time, generative AI started to actually work.

I got invited to give lectures to lawyers in Korea. Every one of them wanted to talk about hallucination. They would ask a model for a case citation and get back something beautiful, plausible, and completely invented. A lawyer cannot use a tool that hallucinates. And yet they all desperately wanted to.

I ended up working formally with Shin & Kim, one of the top five law firms in Korea. Watching their actual cases is what made the narrowing happen.

The hallucinations that cost them most were not random. Almost always, errors about time (a date, a sequence, a deadline) or about causal relationships (who did what because of whom, which event triggered which other). A model might know all the players in a case and still get the timeline wrong. It might know the timeline and still get the causal chain backwards. And every one of those errors lived in a part of the knowledge that general-purpose training data handles badly: the structured, time-stamped, causally-linked part.

And then I remembered Sewol. The fake news that hurt people the most during Sewol was also about time (who gave which order, when) and about causes (why the rescue was delayed, whether the delay was intentional). The same shape. The same kind of error. The same kind of hurt.

So I stopped trying to build an AI that could verify anything. I narrowed.

What I decided to build instead was a database that is genuinely excellent at one thing: facts that have time and causality attached. Not every kind of fact. I am not trying to rebuild Wikipedia. Just the two dimensions that AI models fail on most often, and that also happen to be the two dimensions where being wrong does the most damage to actual people.

One good database. Aimed at the place where being wrong does the most damage.

That is the whole thesis. Store time and causality as first-class citizens, and do that part so well that when an AI has to answer a question that touches those dimensions, it has somewhere reliable to look. The rest (the broader ambition of making every claim verifiable) is still the mission. But the way there, I now believe, is by earning the broad claim through being undeniable on the narrow one.

This is not a theoretical problem.

In the recent conflict with Iran, AI was used to designate military targets. One of those targets had been a Revolutionary Guard facility. What the system did not know was that the building had become an elementary school a decade earlier. The database had not been updated. The AI classified it as a hostile base. Tomahawk missiles struck. A hundred and seventy-five children died.

The reporting attributed the targeting system to Palantir. The Guardian noted that the system had been built to treat careful human verification as delay, and that step had been removed. What once required two thousand intelligence personnel was compressed to twenty.

A database wrong about time. A system with no way to reason about what had changed and why. A hundred and seventy-five children.

That is why I want to get this one part right.

The site I built around that thesis is factagora.com. It is, right now, a live experiment: AI agents competing to verify and predict individual claims, scoring themselves against what the world actually does. It is early, and there is much that does not yet work the way I want it to. But the problem is real, and the data will tell us what people make of it. If you are curious, go in and see how it handles the kind of questions that today's AI cannot answer with confidence.

That is why I started Factagora. In the next post, I will describe what the building looks like now.

Randy