writing down what i know
summary: the experiment teidedb is built on. eighteen years of intuition into a spec dense enough that something else can deliver the implementation.
Core idea: the rayforce2 specification work made tacit database-engineering judgment explicit enough that another executor, human or agentic, could test it.
The experiment I have been running for the last few months: take eighteen years of database engineering - the parts that have stopped feeling like decisions and started feeling like reflexes - and write them down densely enough that someone without my background can deliver the implementation.
Most of what I know is not propositional. It is the shape my hands make when I sit down to write a query engine. No malloc on the hot path. Morsels of 1024 elements. State between morsels, not inside kernels. Predicate pushdown before fusion. None of those are obvious. None of them are wrong. Until I had a counterparty that would actually fail to execute when I left them out, I had no reason to write any of them down at the level required to enforce.
The agent is that counterparty. When the spec is dense enough, the result matches what I would have written by hand. When it is not, the agent gets stuck and surfaces the gap - which is exactly the part I did not know I knew. The gap goes into the next iteration of the spec, and the iteration after is denser.
The transcription is the value. The agent is the test of whether the transcription is complete.
I would not have written down what I know without something forcing the question.