About
One person, one application cycle, seventy-eight thousand lines.
FlintWorkbench came out of writing to sixty professors and losing the research twice. Not the writing, which is the part everyone worries about. The research: read on a Tuesday, gone by the following Monday, and no way to answer where a number had come from.
What happened
The first workspace had a faculty roster with a hundred names on it. Five of the strongest matches were listed as affiliate, which on that roster meant tenure track, and acting on the label would have skipped every one of them. That is the kind of thing you only learn by getting it wrong, and it is most of what is encoded here now.
Why it never calls a model
Partly because the person using it is already paying for a better one than we would embed. Mostly because a tool that runs the model is a tool that quietly substitutes a cheaper one the month margins tighten, and there is no way for a customer to tell. Taking ourselves out of that position was worth more than the feature it costs.
Where it actually is
One product is built and running. Three are described and not written. One person uses it, the author, for a real application cycle, which is how every real bug in it was found. Nobody is paying yet. None of that is on this page by accident: a product whose argument is that claims should be checkable cannot round its own down to confidence.
What happens next
Grad is one season and then you are done with it, which is exactly the retention problem it looks like. It is the cheapest of the four to run and the proof that the rest works. The money, if there is money, is in the three that are bought by people with budgets.