I trust a small working thing more than a long explanation of the thing I have not built. The explanation can be tidy and still omit the constraint that shows up in the first afternoon of implementation. Building is how I meet that constraint.
Eric Ries describes a loop of building, measuring, and learning, with the product idea held as a hypothesis. I use that as a discipline, not as a ceremony. The useful version is short. What do I believe will happen? What did I put in front of someone, or in front of a dataset? What happened that I did not predict?
A prototype is allowed to be incomplete. It is not allowed to be incapable of being wrong. If every result can be narrated as progress, I did not run an experiment. I ran a demonstration.
This is also how I want to treat trading research. A simulator or a backtest is a build. It teaches me what my assumptions were. It does not, by existing, prove that a strategy makes money. Own Research is my public place for analytics and simulation tools. I am not publishing a result from it in this note.
Focus20 Labs is where I build practical AI systems for real workflows. The same rule applies. A workflow that only works in a demo has not been learned yet. It has been performed.
The limit is cost. I should not build a large system to answer a question a conversation or a spreadsheet could answer. Building to learn is a choice about the cheapest contact with reality, not a reason to write more code.