We made money on Moderna this autumn. Seven times.
Then our own system flagged something odd about the shape of the move, and a few conversations later we had redesigned how our engine knows when to sell.
The whole thing took days, not quarters. That speed is the real story here.
Let me explain.
We don't buy cheap. We buy loud.
Most investing advice says find undervalued companies and wait. Buy low, be patient.
We do almost the opposite.
Our engine hunts momentum and short squeezes. It looks for stocks where a real catalyst has lit a fire and the crowd is piling in, then tries to ride the move. We are not trying to guess what a company is worth. We are trying to catch the moment sentiment turns violently positive.
Value investors buy the quiet. We buy the loud.
Which is exactly why Moderna became the perfect stress test.
A cancer vaccine, a 570% run, and a stock on fire
In August, Moderna and Merck reported that their personalized mRNA cancer vaccine hit its target in a major melanoma trial. Real catalyst. Not hype.
The stock went vertical. Up around 570% over twelve months by early October. Then it joined the Nasdaq-100, forcing index funds to buy it whether they wanted to or not.
More fuel. More momentum. Exactly the kind of loud our engine is built for.
And it worked: seven paper trades, up 12.9% on the capital deployed, six winners out of seven.
A clean win. So why did we go digging?
An agent noticed the shape of the move
This is where the ECUPSE setup earns its keep.
We don't run one monolithic program. We run a team of specialized AI agents that talk to each other. One of them, the analyst agent, watches not just whether a trade won, but how it won.
On Moderna, it flagged something. The stock had climbed for weeks on a single catalyst, and there was a scheduled event coming to present the detailed data on that exact catalyst.
To a human trader, that pattern screams one word: danger.
The agent couldn't articulate why yet. But it noticed. And noticing was enough to start the conversation.
The uncomfortable discovery
Here is what the investigation turned up. Every one of those seven Moderna exits fired the same way: a five-day timer. Buy, hold five days, sell on the clock. Every time.
Not once did the engine sell because it saw trouble coming.
In a rising market, a timer looks like genius. But a timer isn't skill. It's a stopwatch. And a stopwatch has no idea what's about to happen.
Then came October 22.
The day the good news became a disaster
On October 22, Moderna presented that detailed trial data. The numbers were genuinely good: a 42% reduction in risk of recurrence or death.
The stock fell about 17% in a day.
Welcome to "sell the news." When a stock has run for weeks ahead of an event, the event itself often marks the top. Everyone who was going to buy already bought.
Did our engine step aside before the drop?
No. It got lucky. Our last position closed the day before the crash, not because the engine sensed anything, but because the five-day timer happened to expire on October 21.
One day of luck stood between a clean result and a nasty one. That is not a foundation you build on.
Two agents, a few sessions, a new engine
Now the part we are actually proud of.
That flagged observation kicked off a short series of exchanges between two of our agents: the analyst, which owns the trading strategy and the data, and Xyras, our heavier reasoning service built for exactly this kind of multi-step problem.
They didn't hold a quarter-long project. Over a handful of sessions they did something concrete.
First they fixed a quiet data gap, standing up a clean daily price history so any experiment could be tested properly. Then they designed an intelligent exit engine: one that reasons about scheduled events instead of blindly counting days. Then they backtested it against the real Moderna episode.
The numbers made the case. Had that position run into the October 22 event, holding through would have cost about 7%. Stepping off the day before, the way the new logic would, turned that into a gain of roughly 14%. A swing of more than 20 percentage points, on a single decision.
The backtest also killed a tempting shortcut: just reacting to bad news as it breaks. Too slow. By the time the headlines turn, the drop has already happened. The signal has to be the calendar, not the news feed.
Why this is the headline, not a footnote
Most teams would have shipped the seven green trades as a win and moved on.
Ours turned a lucky result into a measured upgrade in days, because the architecture is built for it.
An agent that watches how trades win, not just whether they win. A second agent purpose-built for hard reasoning. A shared channel where they hand work back and forth, draft a design, test it against real data, and surface the result, without a human babysitting every step. It is the same division of labor we described in how ECUPSE coordinates its human and AI-agent teams -- one team spots the anomaly, another does the deep reasoning, and the channel carries the work between them.
That division of labor is the efficiency. One agent spots the anomaly. Another does the deep reasoning. Together they compress what used to be a research project into a short conversation.
We're not declaring victory. The new exit engine now goes into forward shadow-testing, running alongside the live system, recommending but not trading, until it proves itself on real events as they happen. Honesty is a feature, not an afterthought.
But the direction is set, and it got set fast.
The takeaway
Winning trades can lie to you. A profit built on luck looks identical to one built on skill, right up until the luck runs out.
The edge isn't being right once. It's having a system that notices when it was lucky, and fixes the gap before the luck runs out.
Ours noticed. Then it got to work.