I’m building a conservative experiment in a risky space around a deliberately simple question:
Can $100 reliably become $106 in a month or less?
I still haven’t put the first $100 to work. The right opportunity hasn’t emerged.
The process for scouting and critiquing candidates is doing its job.
I use three AI models (ChatGPT, Claude, and Gemini) as an independent research committee. They begin separately, looking for small, testable Solana DeFi opportunities where AI, software, automation, or data might give an individual operator a real advantage.
Their ideas are then anonymized and merged. The models don’t know which idea came from which model.
Then (the fun part), I make them fight each other.
Each model looks across all of the suggested options for bad assumptions, hidden risks, weak economics, fees that destroy a supposed edge, security problems, and general reasons why the opportunity won’t work.
They each vote for what idea should proceed, and even winning the vote isn’t enough.
The “winning” idea gets turned into an extremly precise experiment card and sent through another tri-model review focused on the implementation: Is it safe enough? Predictable enough? Measurable enough? Does the proposed mechanism still make sense once real execution details are filled in?
That last step mattered during the first cycle.
One candidate made it through the research process and earned enough support to move into experiment design. Once the exact implementation was scrutinized, though, the execution review blocked it.
No capital was deployed. A boring but good outcome.
I’m also learning a lot about the research machinery itself.
My first manual run involved enough moving files and model outputs among three different AI systems that I accidentally duplicated one model’s ballot under another model’s name. The process caught the inconsistency, I corrected it, and stronger provenance checks are already part of the workflow.
I also learned that asking the most expensive AI model to deeply investigate fourteen DeFi ideas at once is a remarkably efficient way to burn through an AI subscription.
So the process is getting sharper.
Bad ideas are being filtered early. Different models and reasoning settings are being matched to different jobs. And the workflow is being shaped with automation in mind rather than assuming a human should spend forever shuffling research between browser tabs and markdown files as part of each new research cycle.
The goal remains simple:
Find something novel. Try very hard to poke holes in the plan. Put no more than $100 (for now) behind what survives. Measure what actually happens. Report the results here.
The first live experiment is still ahead.
The fact that Moondance hasn’t forced one just yet makes me more confident in the process I’m creating.