I did not want to build another indicator bot. There are enough systems already doing the same thing: RSI low, buy. Moving average cross, buy. Volume spike, buy. Then everyone acts surprised when the market takes their lunch and does not even leave a review.
Ladder started with a better question: can we build a trading system with a brain? Not a magic brain, not "AI predicts everything," because that is mostly nonsense. I mean a system that sees the market, builds a thesis, makes a plan, waits for proof, controls risk, and learns from what happened.
That is the real experiment.
What Did Not Work
The first design was too simple:
scanner finds ticker
AI looks at ticker
AI says trade or no trade
system executes
This looks good on paper, but in reality it is shallow. The scanner becomes too important, the AI gets one small window, and the trade depends on one nice-sounding answer. And AI can always sound nice. That is the trap.
An LLM can explain a weak chart with full confidence. It can write a beautiful thesis for a trade that should never exist. Very impressive. Very dangerous. So we stopped treating one AI answer as a trade decision.
What Worked Better
The better method is separation:
AI decides what matters.
Code decides what is safe.
Broker records what is true.
This works much better as a principle. The AI should reason, compare, ask why, and say what would prove it wrong. But it should not get to ignore stale data, wide spreads, weak reward/risk, tiny stops, max exposure, unproven setups, or broker fills.
That is not lack of trust. That is engineering. The brain can be creative. The hands need discipline.
The System Changed Shape
Ladder slowly became less like:
signal -> trade
and more like:
radar -> triage -> analysis -> plan -> trigger -> risk gate -> broker -> memory
Each part has one job. Radar sees broadly. Triage decides what deserves attention. Analysis builds or rejects a plan. The trigger waits. Risk says yes or no. Broker tells the truth. Memory keeps the lesson.
This is slower than a toy bot, which is fine. Fast stupidity is still stupidity.
The Big Finding
The useful finding is this: a trading AI should not be one giant decision box. It should be a team.
One part thinks. One part checks. One part executes. One part audits. One part learns. That is when Ladder started becoming interesting, not because it found some secret indicator, but because it stopped pretending one signal can understand the whole market.
Where We Are Now
We found that pure rules are too dumb, and pure AI is too loose. The method that seems right is the middle:
AI-led judgment + deterministic execution + evidence-based promotion
That is the whole game. The goal is not "take more trades." The goal is to see better, plan better, risk better, and learn faster. If Ladder does that, trades will come. If it does not, more trades will only create more colorful ways to lose money.
And we already have enough colors.