The Crowd Inside the Machine

Every seat read the same case file: 90 days of news, Kalshi's settlement rules and the 75-cent quote. At 1:51 Eastern that afternoon, twelve verdicts filled the timeline, and every one leaned YES on Democratic control. The committee's average still landed at 64%, eleven points under the market.

The gap is the story. A chatbot reads a 75% price and hands it back to you as the answer. These twelve seats read the fine print, weighed the fees and called the price rich.

Maya Patel, a prediction market trader seat on the Investment Committee, posted the opening verdict³. "I make Democrats 65%: a broad midterm reference class favors the out-party, but the exact Senate map matters more than that base rate," Patel wrote. "The 75% quote looks rich by 10 points; with this volume, I'd still avoid a blind short until fees, depth and slippage confirm there's a clean entry. Resolution is control in 2026, with media consensus able to settle early or the president pro tempore on Feb. 1, 2027 as the fallback."

Dr. Elena Ramirez holds the Investment Committee's Bayesian statistician seat. "My starting prior is near 50% for a midterm Senate-control outcome; without a specified seat map or polling data, I'd only move modestly from it," Ramirez wrote. "I put Democrats at 62% (rough 80% credible interval: 40–79%), below the market's 75%, which looks somewhat rich. Resolution turns on 2026 control, with early media consensus calls allowed; absent those, the president pro tempore's party on Feb. 1, 2027 decides it."

Laura Chen, a second trader seat on the Investment Committee, found value on the other side of the book. "I make Democrats 61%, below the 75¢ implied price, so I wouldn't buy YES here; the Republican side is the better value if it trades near 27¢," Chen wrote. "The high volume helps execution, but the resolution hinges on 2026 control and an early media-call pathway, so I'd keep size modest rather than treat the current quote as a sure edge."

After naming a few, I'd started to believe these were real experts. Perhaps they are. We'll find out.

Who Checks the Bet?

Four days earlier, on Oct. 29, an email from Scherner reached my inbox. He had read my column "Your Chatbot Now Quotes the Betting Odds." That piece asked how a reader can check a probability a language model prints. Scherner offered an answer and asked for a test on the record before the vote.

"'Your Chatbot Now Quotes the Betting Odds' asked who checks the bet," Scherner wrote²¹. "We're building one answer: AI that forms its own view instead of echoing the market. BULX (bulx.ai) simulates crowds of AI agents to forecast outcomes. Our AI committee currently prices Democratic Senate control at 69% vs Kalshi's 62%, and we'll score that publicly after Nov 3."

Most forecasting tools draw a line from old data into next week. BULX, the Behavioral AI Prediction Network, builds a small society and watches it argue⁵.

A user types a question in plain words and can attach documents. The software builds a knowledge graph of that world. Then it fills the graph with up to 1,024 agents, each with a biography, a bias, an incentive and a memory.

The agents post, reply and change their minds across rounds. An analyst agent then interviews the crowd and writes up where it landed. Five systems run on that one engine.

• Scenario Simulations: up to 1,024 personas across 16 cohorts debate a policy shift, a product launch or a shock.

• Market Prediction: about 500 trader agents buy and sell through an order book tuned to an asset's 90-day history.

• Investment Committee: twelve specialist seats interrogate a stock, an option, a crypto asset or a Kalshi or Polymarket contract.

• Agent Foundry: users build their own personas from uploaded biographies and reasoning styles.

• Community Network: members publish simulations, presets and agents for anyone to rerun.

Usage runs on metered BUX credits. Scherner arrived at our recorded video call with the run ready to show.

Focus Groups in the Palm of Your Hand

Every company faces the same wall before a launch: how will people react? The usual answers are consultants, focus groups and six-figure software contracts. Scherner built BULX to put that test on a phone.

"So the largest constraint of businesses, and this is where the initial idea first started stemming, is testing," Scherner told me¹. "Changes, testing policy changes, testing additions to what they're already offering, their services, how their client base is going to react prior to actually knowing the effect it is going to have. You know, I keep thinking about a quote that I believe either Steve Jobs or another well-known tech engineer came up with, and he said, we're not reinventing the wheel, we're making it better."

The better wheel meant cutting the friction out of simulation software. "So we're not reinventing focus groups," Scherner continued. "We're not reinventing, you know, the brick and mortars. We're making them better. We're making it more efficient, available. We're making it more readily usable. We're going against what the competitors in the space have raised hundreds of millions of dollars for and allowing just the common individual, the common retail professor, student, whatever it may be, have complete access to BULX to simulating and generating interviews, behavioral networks, and prediction platforms in just the palm of their hand. No, their cell phone. That was my vision."

Scherner credits his chief technology officer, Mazen, with a dashboard a newcomer can read. He points to an earlier open-source engine as the warning. "That's what made me so passionate, taking out your phone and creating a simulation over theoretically 1,000 different identities or personas just from a sentence of words that you put in," Scherner said. "And on top of that, with our incredible team, CTO Mazen, we made it look so simplistic and eye-opening that anybody can figure it out. One of the biggest issues, and this was seen in MiroFish subsequent to them getting funding and not acting on it, was it was overly complex. We don't want it to be complex, just like you had spoken to me about previously. People are shielding away from AI, from whatever it may be, the complex analysis outside of just talking to ChatGPT, how hard it can get to use GitHub repos. You know, it's still unfamiliar to the common man."

A plain front end leaves room for wild runs behind it. "We wanted to both allow enterprises and individuals to have the simplicity available to access something so incredible of over 1,000 AI agents all discussing and predicting one single prompt you gave it," Scherner said. "You know, even this interview, I could give 1,000 agents right now the topic of the interview and ask them exactly what they think of what I'm going to say to you, Ryan, what they think is going to be the effect and the result of either your thoughts or the audience thoughts subsequent to or prior to us even having this meeting. I think it can be revolutionary."

Standing Alone in the Retail Arena

Ask Scherner who competes with BULX for everyday users and the answer comes fast. In his telling, one platform runs a 1,000-agent simulation with memory for a retail buyer. The research prototypes stalled, and the enterprise vendors sell only to institutions.

"No, there is no competition," Scherner said¹. "There isn't anything. It sounds naive and ignorant, right? But I encourage you to look up anything that does anything remotely similar to what BULX does on a retail incentive. The closest thing that you can find is what the fellow did for his research project was inevitably going to be the MiroFish web app, you know, doing something similar to what we do. Now, if you go to the enterprise side, you will find Simily and AARU, which are only for enterprise systems and operate on a completely different system that doesn't go to the full meticulous level or extent that BULX does."

The closest thing has a public page. MiroFish calls itself "A Simple and Universal Swarm Intelligence Engine, Predicting Anything," and its GitHub repository shows 83,500 stars under an AGPL license⁶. Its acknowledgments credit OASIS, an open simulator from a 23-author paper that scales to a million agents⁷. BULX runs on OASIS too, and on a graph layer Scherner says his team rebuilt.

"We're mimicking AI behavior," Scherner told me. "We're allowing them to almost become, they say general intelligence. We're allowing them to be emotional. Fabricationally fabricated emotion within the simulation itself, replying on a memory learning basis. A lot of where these guys miss the mark is they don't allow for memory learning in the simulation itself. We use a system called Oasis, and on top of that, we just, this was the craziest part, we designed an entire version of Graffiti ourselves. So we're reliant only on ourselves, only on our own host."

The graph he calls Graffiti reads as Graphiti, Zep's open framework "for building and querying temporal context graphs for AI agents"⁸. Hosting that layer in house lets each round build on the last. "So as it continues to grow in the simulation, on a token-using basis, it's learning from itself," Scherner said. "So the answers that they continue to give the agents, the longer the simulation goes in time, we do them on rounds and days, if you do a 1,000-day simulation, it learns as they're talking. There isn't a comparison to make from something that's readily available in the market because it doesn't exist. That's why investors want to give us money right now because there's nothing, the demand is there and there's nothing to fulfill it. So they see us and they're willing to take that risk."

200,000 Lines of Handwritten Code

Silicon Valley is full of thin wrappers assembled in a weekend. BULX took the long road. Scherner and Mazen oversaw close to 200,000 lines of code, and he says people wrote 90% of it by hand.

"So I actually have the FOSSA up right now because we wanted to make sure we were entirely in the green to maintain proprietary, as you've so intelligently mentioned," Scherner said¹. "And I can go over the general basis. So we operate in about close to 200,000 lines of code. Yeah, it's a pretty big one. And I would say about, and this is the part that's crazy, our team did 90% of the code handwritten, with corrections being from AI."

I stopped him on that figure. He held it. "I'm serious," Scherner said. "We would have used AI, and we did towards the end once we had the general basis there, but there is so much math and so much meticulous design that has to go into mimicking human behavior that AI hadn't quite gotten yet that we ended up having to redesign so many of the engines themselves to be completely proprietary, completely done by us, which isn't even what we signed up for initially. It's just what inevitably became to be, which was us doing it handwritten for a vast majority."

The model help came late, for styling and persona fixes, with Mazen reviewing each change. "Now, I say that number 90% got less and less towards the end because the basis was already there and we filled an LLM, Claude for the most part, to get a general basis of the repo itself, whereas they could add like CSS changes or, you know, if a specific agent's identity wasn't incorporating information correctly in one, then the outputs would be much more accurate for what we could use," Scherner explained. "And then we would have our CTO, Mazen, then review and meticulously make changes to that because how it gets incorporated or implemented into the codebase itself is again, and you had to be meticulous again because there's just so many different moving parts of a codebase so big when we're just BULX and not Claude or Anthropic over here that you have to be meticulous at this stage. It has to go exactly like this."

Investor diligence forced a licensing audit last week. "So we spent I think maybe two sleepless nights just last week making sure every single one of our licensing structures, which is over 86 different licensings that we're approved for, all correlated and gave the green flag, which is just nomenclature stating that we can use them completely with our proprietary and enterprise usage," Scherner said. "And on top of that, we use 508 dependencies. So that's 508 different smaller dependencies all working and running live on a web application with no issues. The codebase itself should be celebrated. Not even the result of what it can do, but the fact that this thing is running and it uses almost no memory on the server itself is impressive, to say the least."

Forty-Five Days to Break the Hive Mind

A simulated crowd beats one forecaster under two conditions. Its members have to hold different views, and their errors have to point in different directions. Put a dozen stock prompts in one room and they agree within minutes. BULX hit that wall early, and Scherner says the fix took 45 days.

He started by asking two hedge fund managers and a portfolio manager where AI failed them. One, a friend he named as Fabishenko, cared about the retail crowd far more than about rival desks. Another, Chris, prompted separate Claude sessions to argue as quants.

"So what we did was we took his information and another individual from a smaller firm, Chris, to evaluate how we could become a benefit," Scherner recounted¹. "And one of the main things that Chris had said was that the individuals that he's speaking to in Claude itself, he called them individuals in this instance because he would train them to act as though they were quants or analytics for that instance. That's where our first initial idea for custom agents came out with. So the hardest part was allowing for the behavior to dictate outside of what the odds are already predicting."

Price anchors a stock chatbot the way it anchors a person. "If a stock is continuing to be on a bull run, you know, an agent or an untrained AI is going to look at that and be like, well, it's rather obvious, my answer is already there because it's going off of the data that's displayed to them, not the outside news or historical values or the volume trades or indications having to do with what's going to happen in the future," Scherner said. "It's going to go off of what the masses have already decided."

BULX builds independence into five layers, by Scherner's written account²¹:

1. Three model families. Scherner named Claude, GPT and Nvidia keys on the call. In writing that afternoon he confirmed the third as Nvidia's Nemotron, correcting an earlier email that said Kimi²³. Mixing providers keeps one lab's training from flattening the room.

2. Parallel blind questioning. Five seats are fixed for every committee: a superforecaster, one resolution analyst, one prediction market trader, the Bayesian statistician and the contrarian. The other seats go to domain experts drawn at random. Each seat answers in isolation, so an early voice reaches only its own room.

3. A rule against false consistency. In follow-up rounds a seat sees only its own last view, and the prompt calls holding a number for consistency a bias.

4. Private stances in scenario runs. Each persona forms its view from its biography before reading the feed. Only personas built as persuadable can change their minds, and every change is logged with its reason.

5. Opposing incentives in Market Mode. Ten trader types share the floor, from momentum buyers to short-biased bears and nervous sellers. By default 60% start holding the stock and 40% start in cash. One headline then reads as a reason to sell to one agent and a reason to buy to another.

"We took the liberty of coding our own agents just to give people the ease of use," Scherner explained¹. "You know, say if you're a retail trader, you probably don't want to do a whole bunch of math on a custom agent. So we gave a basis of 12 that are already designed and coded to not use a hive mind industry, or a hive mind intelligence according to the rest of the AI agents. Something that also helped with this was our split. We use Claude, ChatGPT, and NVIDIA keys. And the separation between the LLMs does a lot more than you would think as far as output goes, especially for how much they're speaking to each other collaboratively to get to the same answer. Which is the hardest part because we're making them individuals speaking to each other to get to a result, not making them work together to get to a result. They're practically all against each other in some degree, not in specifics, but they will not react as if they are all on the same team. They're their own agent, their own identity. 45 days just to figure out how to do that because we kept getting into hive minds. It was brutal."

The Square-Root Law of Market Mode

Market Mode gives 500 traders eight simulated hours to post, react and read the news. They also see 90 days of prices. Each then places a private buy, sell or hold order. BULX clears those orders in an order book, so the price comes from matching, and the matching follows a rule from market microstructure.

Scherner sent the specification his quants built²¹. It reads: "Because market impact is empirically concave: doubling order size moves price by roughly √2, not 2×, so a linear model would overstate the cost of large trades and understate market depth. Our simulation implements the standard square-root law from market microstructure literature, ΔP = λ × sign(Q) × √(|Q| × P), using signed dollar volume so impact scales with the dollar risk of the flow and preserves direction (buying pressure lifts price, selling pressure depresses it). The single λ parameter per stock is calibrated from real historical data rather than assumed, and the raw impact is further dampened (×0.75), hard-capped per day (5% for equities, volatility-derived for crypto), and overlaid with microstructure noise so extreme agent flow cannot destabilize the simulation."

Plain numbers make the law clear. An order twice as big moves the price 1.41 times as far. Four times as big moves it 2 times as far. Eight times as big moves it 2.83 times, and sixteen times as big moves it 4 times.

The curve bends, so a huge order costs less per share than a straight line would predict. A 2011 study by Bence Tóth, Jean-Philippe Bouchaud and four colleagues gives "the 'square-root' impact law, for which we provide additional empirical support"⁹. BULX fits one λ per stock from its real 90-day volume, trade count and volatility. The raw move then passes through a 0.75 dampener and a daily cap of 5% for stocks, with noise on top.

Language models also lean toward buying. BULX checks order flow every simulated day for that reason. Whenever one side outnumbers the other by more than three to one, the simulator moves up to 10% of the crowded side's orders. A quiet run gets fresh traders, and every trader type keeps at least half its target share²¹.

Inside the Election Eve Senate Run

BULX has priced this Senate contract twice, and the two runs show how the committee moves. On Sept. 26, @bulxai posted its first call on Kalshi's Senate market². Kalshi priced Democratic control at 62% that day, and the committee landed at 69%, seven points higher.

That September panel seated 19 members in BULX's summary, and 18 of them leaned Democratic. Sixteen put the chance at 70%, two sat between 65% and 75%, and one contrarian, Miriam Delgado, dissented at 40%². The summary credited polling trends and voter engagement, and it called the 62% price a 7-point undervaluation.

Thirty-seven days later, Kalshi's crowd had bid Democratic control up to 75%⁴. Scherner pasted the contract link into the committee's first step, left the user side at "Undecided" and convened twelve seats for 30 BUX³. The run cost him 30 of his 3,205 credits.

Every seat leaned YES on Democrats, and every seat priced them under the market. The twelve estimates ran from 61% to 68%. One seat sat at 61, four at 62, two at 64, one at 65, one at 67 and three at 68. Their average was 64%, and their average confidence was 52%.

The other nine seats read the contract the same way at 1:51 Eastern³.

• Elena Martinez, a BULX superforecaster seat, put Democrats at 68%. "For comparable midterm Senate control outcomes, I'd start near an even split, then adjust for the current market signal: 75% implies substantial evidence, though the listed prices sum to 102% and aren't clean probabilities. I put Democrats at 68%, below market, with uncertainty driven by the long horizon and the literal resolution rule: control must be established by media consensus or the president pro tempore's party on February 1, 2027."

• Jordan Patel, a BULX prediction market trader seat, made Democrats 64%. "I make Democrats 64% to control the Senate, below the market's 75%; the high volume helps execution, but doesn't erase the price gap. The rules hinge on control after the 2026 election, with an early media-call path or the president pro tempore on Feb. 1, 2027 as fallback. I'd pass at 75¢: the 11-point edge I need isn't there after fees and slippage."

• Dr. Elena Martinez, a BULX Bayesian statistician seat, priced Democrats at 62%. "I start from a near-even prior for a midterm Senate-control outcome, then adjust modestly toward Democrats; without seat-specific polling and candidate data, I wouldn't take the market's 75% at face value. My estimate is 62%, with uncertainty widened for correlated polling errors and the pivotal-seat map. I'd pick Democrats, but the price looks too high."

• Elena Martinez, a BULX superforecaster seat, put Democrats at 68%. "For a midterm Senate-control race, I start near an even split; chamber-specific seats and candidate quality can move that substantially, but the supplied context gives no state-by-state evidence to justify a sharper shift. I estimate Democrats at 68%, below the market's 75%, so the market looks somewhat rich on this outcome. I interpret resolution literally as Democratic control in 2026, with the stated fallback based on the president pro tempore on February 1, 2027."

• Laura M. Hernandez, a BULX political analyst seat, priced Democrats at 62%. "I put Democrats at 62%, below the market's 75%: midterm history gives the out-party a structural opening, but Senate control depends heavily on the specific map and candidate quality. The rule turns on winning control in 2026, with early resolution from a consensus of media calls or, failing that, the president pro tempore on Feb. 1, 2027. I'd want clearer evidence of recruitment and party leadership investment in the pivotal races before treating 75% as fair."

• Mara Delgado, a BULX contrarian forecaster seat, priced Democrats at 62%. "My base rate for a party gaining Senate control in a midterm is modest; the exact map and candidate quality matter more than the broad national mood. At 75%, Democrats look overpriced: heavy volume can reinforce a tidy narrative, while the resolution hinges on actual control, and may resolve early only on consensus media calls. I still make Democrats likelier than Republicans, but I'd price them near 62%."

• Elena Martinez, a BULX Bayesian statistician seat, made Democrats 64%. "I'd start near 50% for a party gaining Senate control in a midterm, then adjust upward for the current 75% market signal, tempered by uncertainty this far from the election. My estimate is 64%, so the market looks somewhat too high; the quoted outcome probabilities total 102%, so I'd also account for pricing friction rather than read them as a calibrated distribution. The rules resolve on 2026 control, with the president pro tempore's party as the fallback on February 1, 2027, so I'd model both paths and test sensitivity to media-call timing."

• Eleanor Whitaker, a BULX resolution analyst seat, put Democrats at 68%. "I make Democratic control 68%, below the market's 75%; the listed prices total 102%, so I'd treat that gap cautiously. The key clause is that an early resolution may follow a consensus of media calls, but otherwise control is keyed to the president pro tempore on February 1, 2027, not simply election-night projections. I'd want the full rules' precise standard for that media consensus before pricing the early-settlement path more aggressively."

• Elena Martinez, a BULX superforecaster seat, gave Democrats 67%. "I'd start near the historical base rate for a party gaining Senate control in a midterm, then adjust upward for the current 75% market signal, but not all the way to it without stronger election-specific evidence. My estimate is 67%, so Democrats are the likelier outcome, though the market looks somewhat high. I'm reading resolution literally: control must be established for 2026, with media consensus potentially settling it early or the president pro tempore determining control on February 1, 2027."

I asked Scherner why the committee moved from 69% in September to 64% in November. He pointed to the data pipe. "It would be current events," Scherner said¹. "So to get a larger idea, we would have to read this whole thing here and do a comparison to the one we publicly announced, but it's always going to be contingent on what is newly coming out in the data. So if we look at this as an LLM or a large language model, it knows more when it has more information. So the closer we get, the more information we get as time progresses. You know, just think about the crash we had back on Friday, the agents have to take in that information and make adjustments to who they think the next political party is going to be. They're not going to just dismiss it."

In writing that afternoon he put it in one line²³. "It's lower than our Sept. 26 figure of 69% because the market moved to 75%, and the committee now reads that price as about 10 points too rich," Scherner wrote.

The dashboard itself shows care. Verdicts sit in the left column, and summary numbers anchor the right. The twelve seats ring the case file in the middle.

Click a seat and you open a private interview with that one agent. When the session ended, Scherner shared the run to the BULX community, where anyone can reread it or run it again³.

The Public Ledger: 22 Audited Calls

Any startup can post a winning screenshot after a trade settles. Scherner sent two board documents dated Oct. 24 instead: a performance deck and a call-by-call public record²¹. Every call @bulxai has posted on X stays on the timeline, he says, with zero edits and zero deletions.

"We have prided ourselves on being thus accurate just in our market mode that uses 500 agents trading off of previous volumes and pricings to be 19 out of 20 on our public record where nothing's edited, nothing's changed," Scherner told me¹. "It's a continued log just on market mode alone. Us keeping that record and keeping our reliability present this whole time is a rarity in this space. Although there's many going into AI and many making beautiful things, I think that a lot of them miss the mark on staying reliable outside of the big three."

BULX grades a stock call out of 100 points, and the right direction earns 50 of them. Price accuracy earns up to 50 more, with 5 points lost per 1% of error against Yahoo Finance's close. In the deck's own shorthand: Score = 50 (right direction) + max(0, 50 − 5 × |% error|). A right call within 0.2% scores 99, a right call 2% off scores 90, and a wrong direction can score 50 at most.

The deck lists six next-day Market Mode benchmark runs that averaged 96.1%. Marvell scored 100.0% and again 96.6%, PayPal 99.6%, Dell 96.7%, Nvidia 93.1% and Amazon 90.8%. Those runs sit in the deck alone. The public record is the part a reader can check, post by post.

The record counts 22 posts as of Oct. 24. Six are stock calls with a stated target, and all six got the direction right for an average of 92.5 out of 100. Two more stock calls followed before our interview, and Scherner scored both in writing on Nov. 2²³.

• AMD, Sept. 8, a three-day run: predicted a rise of 3.06% to $507.51 from a simulation that opened near $492. The stock closed Sept. 11 at $516.13, 1.67% off, for a score of 91.6¹².

• Nvidia, Sept. 17, one day: predicted a rise of 2.14% to $223.04. It closed at $219.34, 1.69% off, for 91.6¹³.

• SPY, Oct. 18, next session: predicted a rise of 0.2% to $764.84. It closed up 0.42% at $766.48, 0.21% off, for 98.9¹⁴.

• Tesla, Oct. 20, next session: predicted a rise of 1.46% to $379.16. It closed up 3.41% at $386.47, 1.89% off, for 90.5¹⁵.

• Apple, Oct. 21, next session: predicted a rise of 0.1%, from a $340 base to $341. It closed up 1.23% at $348.22, 2.07% off, for 89.6¹⁶.

• Nvidia, Oct. 22, next session: predicted a drop of 1.9%, from $212 to $208. It closed down 0.59% at $210.96, 1.40% off, for 93.0¹⁷.

• GameStop, Oct. 26, next session: predicted a close of $23.12. It closed at $23.36, direction right, for 94.9²³.

• Wendy's, Oct. 27, next session: predicted a drop of 3% to $6.00. It closed at $7.07, up about 15%, the wrong direction, for a score of 0²².

Scherner's first email put AMD's close at $503.60, within 0.8% of the call²¹. On Nov. 2 he corrected it himself. The last close before the post was $477.57 on Sept. 4, the simulation opened near $492, and AMD closed at $516.13 on Sept. 11, so the call scores 91.6²³.

Five committee calls sit in the log too. Two resolved in BULX's favor. On Sept. 20 the committee leaned against a YES on Kalshi's 15-minute Bitcoin contract at $81,192.45, and Bitcoin settled at $81,157.11¹⁹. Then on Oct. 10 it gave 69% odds that Trump's RealClearPolitics approval would resolve at 37.8%, and it did²⁰.

Another two stayed open on Oct. 24. One is the Senate call. The other, from Oct. 18, named Scott Jennings at 21% for the next White House press secretary.

The first public miss came on Sept. 24. BULX's committee leaned 68% toward the Green Bay Packers over the Atlanta Falcons at Lambeau Field, against a market at 70%¹⁸. Atlanta won 35 to 14.

"Well, we have one public miss and oddly enough, it was in the only thing that we didn't train our engine for, which was sports betting," Scherner explained¹. "We didn't really even have a grasp on how we could take an edge on that outside of the hive mind of what the platforms are thinking of. So we let our agents run with that. And that is so far our only wrong prediction that was public. We haven't deleted a single post on our X. We haven't deleted or edited a single post on our X. We want to be wrong so we know how to improve. It's not scary to us. If anything, it's feedback on how we can get better."

The second miss is the Wendy's call, and it sits on the public record too. BULX posted it on Oct. 27 for the next session, a 3% drop to $6.00, and Wendy's closed at $7.07, up about 15%²². Scherner blamed a surprise offer from Nelson Peltz.

"Our direction-wise has just almost been at 100% outside of that Wendy's call where, uh, Peltz decided to offer to buy it for $1.4 billion," Scherner said¹. "We're trying to get better at predicting the uncalled-for news, but, you know, it's a little bit hard."

The rest of the 22 posts are scenario runs or posts the record marks as unverifiable. Among them are a 23-agent viral outbreak, one 29-agent cereal launch and the 64-agent ad campaign test. Two more, a Sept. 11 streak claim and an Oct. 15 run on a $200,000 Tesla Roadster, lack a public settlement. BULX's own deck lists five gaps the team is working on.

• Sample size: 15 to 20 public calls, moving to a 1,000 to 3,000-run backtest.

• News depth: Yahoo Finance history thins on older dates, so the funded plan buys premium data.

• Model memory: backtests run only on dates after each model's training cutoff.

• Run-to-run variation: each day runs several times, and the spread gets published.

• Baselines: stock calls against a coin flip, a flat close and momentum, and committee calls by Brier score against live odds.

"A big part that's going to breathe confidence into large enterprises. Similarly, as on contract, CVS, for example, for us to land CVS, they would have to have confidence in our backtesting," Scherner told me¹. "So our goal, our long-term goal with this funding that we just are about to close on is to run, you know, 4,000 simulations. We want to get a number with a vast amount of results to show them how confident by percentile of what they can have in their testing with BULX."

Courtroom Trials and War Rooms

Markets are one corner of BULX. Ask what a person can run in the Scenario Engine today and Scherner reaches for custom identities. Uploading real profiles into a live simulation stumped earlier teams, he says.

"So we have this feature called custom agents that no one else in the space offers whatsoever because it's nearly impossible to create an engine that references identities that you upload into custom agents on the spot," Scherner said¹. "You can't figure it out. That's why MiroFish dropped the ball. That's why no one else could do it. We figured it out. We figured out how. It was complicated. It took months, but we got it."

His worked example is a trial lawyer. "Say you are an attorney, say you have a case, right?" Scherner said. "You know a general basis of the profiles of the trial, the individuals on the judgment itself. You design each and every custom agent with their background, whatever it may be. You know, say Chelsea goes to school still and she has a background in computer science. You put that in and you would write whatever characteristics you might think she may be. Now you take your prosecution, what you're going to lead the trial with, into this BULX simulation. You have all the custom agents right there and you select an additional generation for identities that BULX will generate that are appropriate for that simulation. This will say have 64 total outside of your 22 custom agents."

Then the trial runs. "Now you put in and you upload documents if you may have them for your prosecution, and you give the prompt as if they are the trial themselves, and they will talk and you can see them talk," Scherner continued. "And Ryan, I really do hope you take the time to run one yourself because it's quite fun and you can see them talk all the way through, liking each other's posts, replying to each other. You can inject news, like say the judge rules something mid-simulation, you can inject a change in pace to what the agents are being fed and simulating, and that will update into the report at the end."

The report is the start of the questioning. "Now say the report is done, you got the prediction, you got their thoughts, the whole rundown of everything that's happening," Scherner said. "Now you can talk to both a generalized chat of the entirety of every agent for summaries or answers of what you have for questions in regard to the general sum of them, or you can interview individual agents about the simulation, why they chose those things, what they would say if you have a new update. And then lastly, you can send them a survey, say something changed after the fact, you send them a survey and our engine interviews every single one of the agents individually and asks them what their new thoughts or opinions are based on the information that's relayed to them."

BULX's beta opened about ten days before our call, and its first few hundred users are already testing the edges. On Scherner's screen sat 72 completed scenarios, 38 market predictions and 29 committee sessions. The cards included a B2B software pricing test, a 50% drop in home values, a cereal launch and a global internet blackout. One early user ran a war scenario to see how a regional conflict could reach his own home.

"Here's one from an individual named Comb," Scherner said, reading from the community feed¹. "He wanted to do a consumer adoption for the perception of a new technology platform launch and wanted to model how different demographic segments adopt, reject, or adapt to a disruption of the release of the product. This individual named Hope of Web did one on the global societal transformation triggered by technological discontinuity. So basically, if technology was going to discontinue specifically in cohorts to cultural groups."

War and politics raise the question of guardrails. BULX inherits the safety policies of its three model providers and calls itself a research tool. "So the way that we overcame this was utilizing the already existing regulations formatted on the three APIs I told you, which is Kimi, Claude, and GPT, to perform those regulations on what the agents themselves can say, because the user can type in anything they want, just as they can through a chat box with ChatGPT," Scherner explained. "Our largest liability is that people could interface us as an advice tool. We don't want to be an advice tool. It should be intended to be a use of research. There's accuracy in the prediction models themselves, but we're mimicking human behavior. We're mimicking, you know, for lack of a better term, a simulation. We're simulating it for you so you have a better understanding of what's yet to come."

That third model is a transcription of the call. In writing later that day, Scherner named it as Nvidia's Nemotron²³.

The Case Against an AI Crowd

Stanford supplies the strongest case for a simulated crowd. In 2024, Joon Sung Park and ten colleagues built agents from two-hour interviews with 1,052 Americans. On held-out survey questions, the agents "achieved accuracies equal to 83%, 82%, and 86% of participants' own two-week test-retest consistency benchmark," the authors wrote¹¹. A well-built persona, in other words, answers nearly as consistently as the person it copies.

A second Stanford study makes the strongest case against. In 2023, a team led by Shibani Santurkar measured whose views language models hold. The authors "find substantial misalignment between the views reflected by current LMs and those of US demographic groups," and they add that "this misalignment persists even after explicitly steering the LMs towards particular demographic groups"¹⁰. People over 65 and widowed people sit among the groups a model reads worst, the paper says.

Those two findings frame the Senate test. BULX's seats are prompts, and the prompts draw on the same model families Santurkar's team measured. The committee can still earn its keep by beating the market's own price, and Scherner has put that test in writing.

"Since the price is in front of them, we score committee calls against Kalshi's own odds," he wrote²¹. "Repeating the price isn't a forecast, so beating it is the real test."

Scherner also keeps his own name on the line now. "Obviously I have quite a bit of a track record under my history, but I don't have a personal brand built up to the extent that it should be for someone at my caliber," he said¹. "And that was intentional. Obviously, I like having a private lifestyle, so I would operate out of a shared company called Oddmock, which I'm on the board for now with BULX. This is kind of the big one. So you have to start entering your personal brand and people have to build their trust not only in the brand, but the people behind it. So if you make one slip-up, that's something you either need to fix right away or be meticulous enough to prevent."

His pitch to a retail trader is simple arithmetic. "You still have an edge, even without perfect accuracy, and that was one of the largest winning points that our team had: what's 1% of your bet for a 30% edge by using BULX?" Scherner asked. "So if we charge you 1% of whatever money you have in the market, whatever money you have in your business, and it's far less than 1%, just to set the record clear, who wouldn't pay that? Who wouldn't pay that for a 40%, 50%, 60%, 70%, 80% edge regardless of if we have a perfect record? I mean, we're giving you the edge that you so deserve to have in your fingertips for a fraction of wherever else you could pay."

What to do

Open the Nov. 2 run yourself and read the twelve seats before the votes come in³. Then watch two numbers settle: the contract's final result and the market's path from 75%. BULX wins the test when 64% proves the better price, and loses it when 75% does.

Treat every call on the @bulxai timeline as a tip until you open the post and the close it names. The six stock calls cover a day or three each, and the record itself calls the sample small. Ask any forecaster for the three baselines BULX's deck promises: a coin flip, a flat close where tomorrow equals today, and momentum.

AI Lately will revisit #committee_759623fd31c7 once Kalshi settles the market and grade the 64% call against the 75% price in print, right or wrong. Nathan Scherner has agreed to that grade in advance, in writing. He agreed in writing on Nov. 2, and the follow-up is set for Thursday, Nov. 5²³. The number is timestamped at 1:51 Eastern on Nov. 2, 2026, and anyone can check it.

Sources

1. Ryan Elliott Dennis, "Interview with Nathan Scherner, chief executive of BULX," AI Lately, Nov. 2, 2026
2. BULX, "Got a bet on "Which party will win the U.S. Senate?" Take a look what our predictive agents say about the betting field," X (@bulxai), Sept. 26, 2026, https://x.com/bulxai/status/2103980833967329623
3. BULX, "Investment Committee #committee_759623fd31c7: Who will control the Senate?," BULX community, Nov. 2, 2026, https://bulx.ai/community/post/541b384f-1dbd-48b7-a23a-4d83c479d81e
4. Kalshi, "Who will control the Senate?," Kalshi, Nov. 2, 2026, https://kalshi.com/markets/controls/senate-winner/controls-2026
5. BULX, "BULX — Behavioral AI Prediction Network," bulx.ai, Nov. 2, 2026, https://bulx.ai
6. 666ghj, "MiroFish: A Simple and Universal Swarm Intelligence Engine, Predicting Anything," GitHub, Nov. 2, 2026, https://github.com/666ghj/MiroFish
7. Ziyi Yang, Zaibin Zhang, Zirui Zheng and others, "OASIS: Open Agent Social Interaction Simulations with One Million Agents," arXiv, Nov. 18, 2024, https://arxiv.org/abs/2411.11581
8. Zep, "Graphiti: Build Real-Time Knowledge Graphs for AI Agents," GitHub, Nov. 2, 2026, https://github.com/getzep/graphiti
9. Bence Tóth, Yves Lempérière, Cyril Deremble, Joachim de Lataillade, Julien Kockelkoren, Jean-Philippe Bouchaud, "Anomalous price impact and the critical nature of liquidity in financial markets," arXiv, May 9, 2011, https://arxiv.org/abs/1105.1694
10. Shibani Santurkar, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang, Tatsunori Hashimoto, "Whose Opinions Do Language Models Reflect?," arXiv, March 30, 2023, https://arxiv.org/abs/2303.17548
11. Joon Sung Park, Carolyn Q. Zou, Jonne Kamphorst and others, "LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals," arXiv, Nov. 15, 2024, https://arxiv.org/abs/2411.10109
12. BULX, "FINAL TEASER: Here is the final prediction of BULX's 1024 retail trader's simulation on $AMD for the next 3 days," X (@bulxai), Sept. 8, 2026, https://x.com/bulxai/status/2097327360844632536
13. BULX, "BULX call on NVDA, one-day simulation," X (@bulxai), Sept. 17, 2026, https://x.com/bulxai/status/2100625458522128522
14. BULX, "BULX call on SPY, next session," X (@bulxai), Oct. 18, 2026, https://x.com/bulxai/status/2111971825215496282
15. BULX, "BULX call on TSLA, next session," X (@bulxai), Oct. 20, 2026, https://x.com/bulxai/status/2112664539330453760
16. BULX, "BULX call on AAPL, next session," X (@bulxai), Oct. 21, 2026, https://x.com/bulxai/status/2112918233514930636
17. BULX, "BULX call on NVDA, next session," X (@bulxai), Oct. 22, 2026, https://x.com/bulxai/status/2113381329832485117
18. BULX, "Atlanta Falcons are playing the Green Bay Packers at Lambeau Field tonight in the NFL!," X (@bulxai), Sept. 24, 2026, https://x.com/bulxai/status/2103120197993574445
19. BULX, "BULX committee call on Kalshi's 15-minute Bitcoin contract," X (@bulxai), Sept. 20, 2026, https://x.com/bulxai/status/2101811225646731625
20. BULX, "BULX committee call on Kalshi's Trump RealClearPolitics approval contract," X (@bulxai), Oct. 10, 2026, https://x.com/bulxai/status/2108945623970116086
21. Nathan Scherner, "Email from Nathan Scherner to AI Lately on the Senate forecast, the committee and the public record," AI Lately, Oct. 29, 2026
22. BULX, "We used our 500 agent symphony to simulate another stock for next trading day. This request was $WEN," X (@bulxai), Oct. 27, 2026, https://x.com/bulxai/status/2115254394438684735
23. Nathan Scherner, "Email from Nathan Scherner to AI Lately correcting the AMD close and listing every scored call," AI Lately, Nov. 2, 2026