We automated the process of discarding, not the process of creating EA

Last time I wrote a story about making AI search for edges. Searching for edges isbefore you createthe story about.

This time it’s the opposite—after you create, you discarda story.
In this article, I describe the principles for mechanically creating EA using an AI agent. I will freely distribute the blueprint, so please use it.

What costs the most in EA development

This is my personal belief: the one thing you must not do in EA development is

to pour time into implementation and optimization while not noticing the faulty logic.

If you turn it into code, run tens of variations of parameters, and create a performance chart, people will become unable to discard it. The more time you’ve spent, the harder it is to let go.

This is what is called the sunk cost fallacy.

And those who become reluctantadd conditions instead of discardingto avoid losing positions and run it again. Repeating this will yield the “100% win rate, explosive EA” mentioned in the previous article. It may sound exaggerated, but content close to this can easily occur.

So Iautomated not the creation process, but the discarding process.

We discarded four

With this mechanism, I validated turning famous methods into EAs one after another,discarded four in a row as failbacks.

It wasn’t that I couldn’t create them.It was that they didn’t meet the criteria I set in advance.

I mechanically prevented the “a little more persistence” for four candidates. (If they had persisted, one of them could have been raised to a level where it could be listed.)

And the ones that passed the same criteria have actually gone out. EA that passed this mechanism are being reviewed and sold.

This doesn’t mean that by automating the discarding process you will automatically pass.It is a proof that it is not a broken device that can only produce failures.We cannot build a system that guarantees passing.

Why hand the judgment to a machine

Not because it’s faster. Speed is a side effect.

The true reason isthat when you judge by yourself, the criteria move after you see the results.

Even if you had decided “drop below PF1.0 and discard,” when 0.98 comes up you think, “just a bit more.” I’ve done this many times. I didn’t intend to break the criteria, but every time a valid reason pops into my head.

If you run it through agents,the criteria must be written in advance to move.The order becomes forcibly reversed. I could say I created this mechanism just for this purpose.

What changed after introduction

The time to discard became shorter.Previously it took weeks to give up on a single logic. Now illogical ones are eliminated before writing the code.

The reasons for discarding remain, so bad logic can be excluded in advance.Instead of saying “it was bad,” the gate, the threshold values, and the criteria at which they fell remain. Thanks to this, AI now helps filter out similar but different logics before trying them.

Publishing numbers is no longer scary.The machine cross-checks whether the initial code values and performance chart align, so there is no structural accident of “the published numbers and the actuals differ.” This was the biggest change for sellers and developers alike.

Only one principle to write

There are five principles in total, but here I’ll write aboutthe one that caused the most actual harm.

Detect anomalies in the environment first.

There was a case where the test environment’s fees were set much higher than standard. Running tests without realizing it,the sensible logic looked wiped out.

So at the very start of the process,place a gate that doubts the environment before writing a single line of logic. This was added after the failure; it wasn’t obvious from the start.

What this mechanism cannot do

“If you just hand over the procedure, your AI will produce market-ready products”—this has not been demonstrated.

The results available now are those I obtained when I acted as site supervisor. There is no record of reaching passing status by handing over the procedure and swapping both people and AI.

Therefore,this mechanism is not an unattended device that makes products by itself.If you manage to create a passing EA in your environment, please let me know.

I’ve summarized the design principles into a PDF

The remaining four principles andwhere in the process the gates are placed (order of gates 0–4), how “failures” are emitted, andwhat to withhold from the machine and keep in human hands.

I’ve compiled this into a 6-page PDF and will provide it for free.I am not withholding.Putting gates that mechanically discard is beneficial for both EA buyers and sellers.

It’s free, but you need to apply.