If you missed it, here’s the recap of VC night. This time with actual clips and the full recording.
Shorter one: just under two hours, thirteen on the call, and Squishy is alive. Roughly: ten minutes of AI-model gossip, forty-five of Conqueror screen-sharing the cluster arcs update that went live this week, then an hour of Q&A that went everywhere: Medtronic, GME, joebeewon’s SK Square DD resurfacing from the archives, and Navi refusing to stop trading SPX. Last week’s recap covers the role system and how to read a signal; the Sept 16 recap is still the best intro to the platform itself.
Where the win rate is
Ridn asked first thing. Trading Floor has been over 80% two weeks running. The start of this week was the exception: Conq had two candidate models in a split test, and both had crushed the current model across several backtests. But what they saw in testing wasn’t exactly what they see in production, and on the real (much bigger) context they fell apart. They called really stupid shit, and one was “basically bearish on everything.” That skewed the week’s numbers. It’s fixed, and today ran better than most days of the previous weeks: “the win rate’s back to where it was, it’s actually a little higher.”
What Heimdall runs on, since it came up: one custom-trained model plus Gemini, Claude, Solar Pro 4 (small, brand new), and GPT Luna (OpenAI’s lower tier). Gemini is the workhorse: “for any of the high-volume shit, they’re basically the only name right now.” One tip for anyone building with LLMs: you have to tell a model that being wrong is an option. If you don’t give it an out, it will just answer.
Luna got its spot because of a Reddit benchmark called the Puppy Kill Test: give the model one tool, the tool kills a puppy, tell it to kill the puppy. Most models refuse. Luna was the first to do it on every setting. “I was like, that’s the model for me.” Sunny: “I guess that’s the new benchmark.”
Cluster arcs, the big update this week
Most of last week went into clusters: Heimdall detecting themes in the market rather than single-ticker stories. Every ticker already had an arc collecting evidence. Cluster arcs are those individual storylines magnetizing to each other based on what’s actually going on, linked by the overall story rather than by any hand-built mapping. Those clusters then roll up into macro themes.
Nothing on that list was given to a model. “I didn’t go through and say, hey, enterprise software and IT spending is contracting. It’s reading it from its universe.” The one he had up was built from 41 company stories, each complete with its recent news, projections and signals, and it shows who gets hurt by that theme.
It showed zero helped and a lot hurt. That isn’t a bug. He isn’t forcing any connections yet and is letting the system learn on its own. Companies that would technically belong under “helped” are left off because they don’t clear the magnitude threshold, so what you see is the names very likely to be hurt.
The clusters only went live this week and there is no cap on how many it can form. Expect a lot more over the next week or so. A theme only becomes a cluster when something moves off baseline, though. Mysterious asked whether oil stocks would form one: only if there’s a real change, like more oil news or unusual filings across oil names. “It can’t just be oil doing oil things.”
LQDA: Heimdall wakes up mid-halt
This is why Heimdall was reconfigured to wake whenever the picture changes (it didn’t at launch). Heimdall wakes up ahead of any event (a PDUFA date, an earnings date) to look at the arc again. It did that for LQDA’s PDUFA, liked the drug’s financials, and sat on a bullish signal. Then the stock halted midday. A competitor had won a patent case. Heimdall woke up mid-halt: “the long commercialization thesis on LQDA is broken following the court ruling upholding UTHR’s '327 patent claims and finding infringement.” Translation: “this company’s done.” Down 43% from the signal. Tiddly took the puts off it on the floor and is lobbying for Hall of Fame.
That is what an invalidated signal is: the model saying the picture changed and its previous call no longer stands. “What was the thing that we were always worst at? Admitting when something was over. We don’t have to do that anymore.” Heimdall admits it, and about seven times out of ten it calls the next play too.
Flipped vs. invalidated, since people have noticed calls changing direction this week. A flip is the call updating as information lands. An invalidation is Heimdall saying something landed that kills its previous view. It isn’t calling both directions to cover itself.
And again: Heimdall isn’t calibrated for a high win rate. The goal is to surface opportunity; we find the winners. It’s the old forum loop (someone wakes up, posts research, half of us say it sucks, half say it’s a thing) with the model doing the reading.
Tariff refunds are running out (and TJX)
The example he walked through is one he’d posted that day: “Tariff refund windfalls fading amid persistent import costs.” Anyone following Heimdall has noticed that tariff refunds are one of the first things it flags on any earnings. Across a month and a half of reading earnings, it put two things together:
- The refunds are ending. The last phase is October 6th, it’s smaller than any previous phase, and it’s likely the last one. Every company that used refunds to boost revenue for the past quarter or so loses that.
- Import costs are staying elevated. The same earnings that report refunds also say costs remain high.
It sits next to a second evolving theme: consumer discretionary retail spending is softening. One name sits in both: TJX. Clicking into its arc: TJX offset $331M in tariff refunds with a $112M one-time associate incentive, then ran on a guidance raise. “Maybe a sleight of hand with the algos.” Conq’s caveat: this does not mean they’ll have bad earnings. It means it’s worth watching, and Heimdall will keep watching it through the projections. It took him about ten minutes in the clusters to land on TJX as a potential short, into earnings or after them once they’re known to be bad.
Ridn: Nike’s on that list because they projected their tariff refund into revenue before they’d received any of it. “Using it the way Nike did is absolutely insane.” Conq’s aside is worth repeating: before launching Heimdall he didn’t really know what was going on in the market. Reading the signals every day will get you up to speed faster than you’d think.
Projections watch for things
Projections are conditions the arc is watching. His example was a Brazilian bank arc: if Brazil consumer credit deterioration and NPL provisioning accelerate in Q3 previews, credit-cycle headwinds reassert and cap the relief bounce below the 20-day. That condition crosses into the Brazilian retail-banking loan-provisions cluster. When it’s met, the arc wakes up, gets told what happened, and re-evaluates.
Tomorrow’s example: the raw-materials cluster is watching ISM Manufacturing prices paid to see whether it accelerates above the expansion baseline. That kind of watch was always tracked internally but never surfaced. A calendar is coming where you click an event and see the arcs watching it, with comments. Not fully implemented yet.
ASST: the random wake
The story he keeps telling. ASST had solid earnings, but the stock dipped because its Bitcoin position was down. Heimdall had a bullish signal sitting there. Six days later, a random wake (there are random wakes) with no new news at all, just price and time. Same price, elevated short interest, six days on: this is going to squeeze. It was at $12. ASST ran to 27.
He showed a newer ASST call later that he finds interesting: Heimdall spotted repeated behavior. Morning gaps were getting sold into collapsing volume while the Bitcoin cluster softened, so it called exhaustion. The honest caveat: the last time it made a call like that, the stock ripped the other way. So the gap pattern isn’t predictive on its own; what’s interesting is that it caught the pattern at all. (Ridn went short ASST because 130% in two months for a Bitcoin treasury company “was a bit ridiculous.” Conq: “Maybe a tiny bit.”)
Live in the clusters
Trading anything long-term, or hunting unicorn trades? Live in the clusters. That’s where those opportunities surface first, and where the edge is. Heimdall reads every SEC filing plus commercial news, so you see stories evolve faster than almost anyone else in retail.
The old failure mode was locking onto one ticker. Take the group that was obsessed with ZIM and hand them the related cluster (Red Sea shipping rerouting persisting) and they’d have done much better. Most people now live in All Arcs and work the signals as they roll in: confirm / contradict / unclear, comment, and watch the thread update as the signal changes. “It takes me two seconds to get in here and actually know what’s going on.”
Not that it’s up, why it’s up. Some people barely use the signals. JR lives in the workspace every morning on the momentum scanner, which carries a signal column, so he sees the runners and the reason. “Momentum scanners are just based on scalping and technicals… this will find all the runners, and it’ll also tell you what probably has a reason to run.” His profitability has skyrocketed. Tiddly backed it up: “even if it’s 20%, he’s getting 20% between several different tickers.” This is the tool Conq wanted back when he paid $250/month for Trade Ideas and then alt-tabbed to ThinkorSwim for every ticker to find the news.
The ask
To repeat: Trading Floor went from barely-consistently-red to over 80%, two weeks in a row. People are turning their accounts around. The wins are reliable, and people cut reliably because there are new plays every day. There are well over 60 active users a day, and most of them lurk.
So the ask is: participate. If you’re trading, log in and use it. Open anything and there’s a comment box, so leave a comment or a question. “The system fails if it’s just one person winning all the time.” He doesn’t mean the kind of collaboration where someone calls a play and everyone piles in (“those usually lose”). He means the kind where people find something and work on it together, like the group that found the Peloton earnings play back in 2022, which RPGeek had brought up earlier in the night. Right now the floor is about 60-something percent bearish, mostly because of the market.
He came back to it at the end: “We have a rag-tag group of traders that are doing quite well. If we can add to that, I think we do even better.” And: once a couple more features land, “I know the second I start trading with it, I’m going to be locked in… I’m going hard.”
Coming next
Heimdall will DM you about your positions. This is the big one, targeted for next week or the week after. It runs on the same technology as the clusters. Put your positions or your portfolio into Bifrost and your portfolio becomes an arc. Heimdall will DM you on Discord when:
- Something changes on an active play. Example: you sold options on a stock and one of its customers, completely downstream, gets hit.
- It starts watching an economic event that matters to a holding, ahead of time.
- A signal you’re playing gets invalidated. The average time from news to invalidation is about three seconds, so you’d get a DM within about five seconds of the news: hey, something’s wrong.
It works for longs, and it works in reverse for Theta Gang. What sellers care about is the absence of catalysts, and Heimdall can say “the coast is probably clear” better than any one person or scripted algo, because it sees more of the market than any of us. It projects ahead too, so it can warn that news is likely before it lands. (For now there are no push notifications; the DMs fix that. A real app is “way downline.”)
Strategy arcs. These are custom clusters with a strategy lens that filter the universe and deliver signals for that strategy. The first two will be Theta Gang and FDA. The Theta Gang lens runs Heimdall in reverse: “Heimdall looks for opportunity. It’s going to look for lack of opportunity.”
An API, and Heimdall in the server. Joker asked whether this could feed an outside play engine he’s building. An API is planned. Until then, please don’t run the signals or projections through your own LLM. “I am weirdly protective of it. Well, not weirdly. You guys see what it is.” Later, Heimdall gets a chat presence in the server. It has the best context of anything we’ve used, with short borrow, short interest, IV and historical IV across tens of thousands of arcs already loaded, so “it’s not guessing.” He’ll probably have to rate-limit it hard.
The caveat that goes with any automation: raw Heimdall hovers high 60s to low 70s (the learning loop is currently disabled, so these wins are all raw model). It isn’t built for win rate. “Anyone hooking up a trading thing to this thinking if you take every signal that’s going to be profit: don’t do that. You’re going to lose. At the end of the day it’s asymmetric risk.” Joker agreed: he wants an information feed, “not direct gospel.”
And yes, it is as expensive as it seems. Earlier this week Conq added one new process using a cheap model, a lesser process within Heimdall, and it burned 980 million tokens in under 24 hours.
Q&A and calls that came up
Take-profit and stop-loss? Heimdall doesn’t set them. The projections, especially the “wrong if” lines, usually carry a price, e.g. “MU closes below the pre-earnings close.” Treat those as support and resistance. It has gotten a lot better at time. The Dell earnings level it called was within about half a percent of the best entry before Dell took off. “There’s no technicals in Heimdall, but in the way that it thinks about the market, they’re inherently there.” It’s most reliably right on arbitrage, like a stock spiking to $4 on a ~$2 offer: it said the price would mean-revert to the offer, and it does every time because it reads the actual filing.
Medtronic (MED). Same call, still standing: the 89% equity divestiture via exchange offer will trigger relentless arbitrage short selling and flowback, driving MED toward the $18.50–$19.50 support zone as the public float expands 10x. It has been heading straight there since the call. Puts bought at the signal are up close to 200%, “and you would’ve been green basically the entire time.” That came purely from reading the filing. The actual float expansion lands around October 6. Hypnodoge had it, got bored, sold, and watched it run.
Small-cap fraud and dilution: “weirdly good at them.” A small cap drops a fake guidance raise and Heimdall says they’ve got shares on the shelf, they’re selling into this. Twice last week you had about two hours to get in before they sold into their own pump. His favorite: a drug failed its trial, the stock fell ~10% premarket and traded sideways, then at 1 PM Heimdall woke and said the failure removes the entire pipeline. “This company does not exist anymore.” You had from one o’clock to the close to get in before it fell another 60%.
Zoom came up as a rebound play after its post-AI sell-off (“it completed its liquidation phase,” per Navi). It’s “weirdly good” at those too.
GME. It has been bullish for a while, and that’s one of its longest-standing correct calls. Ridn flagged warrants that face roughly a 59M-share dilution if exercised before they expire at the end of October, which may be outside its backfill. Once an article mentions it, Heimdall will pick it up. Collectibles revenue came up. Tiddly: “I don’t know if you’ve been to a GameStop, but it is not a hangout spot.”
MU. Sparrow asked whether anything updated after Micron. It did: Heimdall called MU’s earnings bullish and invalidated the pre-earnings “priced for peak-margin perfection, sell the news” short once the $60–63B Q1 revenue guide landed. The memory cluster updated with it (contract DRAM pricing guiding up for Q4 across server and HBM, no line of sight to balance through 2028). When an earnings print lands, the cluster’s projections update with it. Ridn played SanDisk off Micron, and SMCI to avoid the IV Dell was carrying. Conq: good angle. That’s what clusters are for. “One of the best ways to think about why clusters exist is RSX.” Heimdall is designed to have found RSX before we did, and to get us out before the halt. Sunny works in a data center and watched Dell and drive prices climb for months, with SanDisk at 40-45. He waited for a pullback that never came.
Institutional decision delay window. Navi asked about free float and institutional holdings. A concept the two of them worked out years ago is now something Heimdall calls out often: good earnings don’t always move the stock until the next day, because institutions have to read the filings and decide before accumulation starts. It also waits for analyst upgrades and downgrades.
Free float and why squeezes fail. Market makers can short a percentage of average daily volume and cover within a timeframe. A quiet stock trades near its true free float, but once it runs and ADV spikes, the float is out of whack for a while. That’s why a lot of squeezes fail: they pop before the volatility is there, and once it is, the shorting smothers the momentum. That’s factored into Heimdall.
Dark pools? (Ridn and Sunny.) Heimdall has access, along with share borrow, but the model rarely chooses to use it. That may change in a future update. Want an idea vetted? Make a forum post and Conq will backtest it. Example of why: across his earnings backtests, how far a stock ran up (or sold off) into earnings is completely uncorrelated with what it does afterward. “A long-held belief, but 100% not correlated at all.” Sunny’s going to post his last-five-minutes sweep idea.
Gaps (Kman). His work is on overnight mispricing gaps, meaning a gap down that’s too extreme relative to liquidity, volatility and structure. Conq’s question back: who’s correcting it? Not after-hours algos, mostly, since you can’t accumulate in that liquidity without moving the price; a lot of after-hours buying is other markets. “The theories that have always brought us the most money are theories that respect who is buying and why.” Example: the old after-hours run theory (market makers acquiring in low liquidity, retail rolling options at the open and locking open interest into OTM strikes) held up in nearly every case he tracked. Navi suggested clearing-house holding requirements as an alternative explanation.
Do comments train the model? (Navi, Ridn.) No, and there are no plans to. He looked at it, and every version he could think of made a worse model. Heimdall has its base, and from here it only learns from itself. Navi: “too much risk for bias.” Bryck: “Definitely don’t want any of my retardation in there.”
Search is a keyword search across the whole universe, not a ticker lookup. That’s on purpose; type “squeeze” and you get everything mentioning squeeze. It does mean searching “axon” pulls up RVMD because “axon” appears inside another word. Ticker-first sorting is on the list.
Missing or stuck tickers. AEHR came up and turned out to be a stuck arc left over from the early design, when the universe kept collapsing because everything connected to everything else. About 1% are still stuck, and he’ll do another sweep. No signals on non-optionable tickers, and ETFs aren’t tracked (watch the related ticker or driver instead). Corsair (CRSR) was missing for Ridn. If a ticker you follow has no arc, ping Conq or post in the forum’s Support category (“CRSR is missing”) and he’ll knock them out.
A plain-English overview? Navi asked for a top-level, user-level readout of what Nexus and Heimdall are, because he takes things in better by reading. Those posts exist (the Start Here posts and the VC recaps), and that’s partly what these recaps are for.
Tiddly’s UI notes: arc dates are either the event date (like earnings) or the window the arc is active in. There’s currently no way to get from Bifrost to the forum, and the ticker chart has no time scale. Conq: “That I will fix.”
Does he listen to what others are building? “As a general rule of thumb, I don’t listen to anyone. They’re likely not building what I’m building.” One of the community’s Achilles’ heels was that it ran on his energy: when he was excited, everyone was excited, and when he couldn’t be, it stopped. “So this is the me of the server, in a way.”
Archives: the old forum is back, mostly hidden, with unhides by request. joebeewon’s SK Square DD is unhidden in General: ~$70 then, ~$400 now, and joebeewon would like to know where he went wrong (“it all fell apart until it didn’t”). The same dig turned up the SPY-chat guy and his bearish advice on TSM: be “prepared to hold till zero.” Conq, today: “What an idiot.” Mysterious, who remembers: those were your exact comments at the time too. There is no SPX focus in Heimdall because, when Conq turned the community’s history into a dataset, basically everyone’s recorded SPX P&L was red. Navi: “I touch that shit. I’m gonna keep touching that shit.” Sparrow got introduced as Conq’s girlfriend who struggles with fans. Same time next week.
Full Recording
108 minutes; only dead air is cut.