A model here is one short file: how you think this slice of reality works, what you expect to see and when, and what would make you retire it. Then reality grades it, on a date.
How your industry turns. What a committee means when it says under review. Which patients get worse overnight. It fires before you can explain it, and it is probably the most valuable thing you know.
This is the moment such a model can leave the head it grew in: not the facts, the interpretation. Push it to a repository you own. The registry keeps the address and nothing else. Here is what a model can do once it is outside.
Fifteen years of judgement usually leaves when the person does. The founder retires and the company keeps the org chart and loses the instinct. The mentor's advice survives as three sentences you half remember.
Written as a model, it stays runnable. A successor inherits the founder's way of deciding, not a memoir about it. And you get to compare your own model from two years ago with the one you hold now, and see which of you was right.
Paste a model's link into whatever assistant you already use and the assistant becomes the model, applying its premises to your question. Or clone it and run it at home, offline, on Ollama or LM Studio. No account, no install, no lock-in to any one AI.
A regulator's interpretation becomes a file the regulated can run before they file. A citizen runs the council's own stated logic against the council's own budget.
Everywhere else, being publicly wrong is a reason to delete the post. Here a model graded wrong stays listed with its record showing. Refuted theories stop being reinvented every decade. Nothing is ranked, so a narrow model of one regulated industry is never buried under a popular one about markets. There is no pile to be buried in.
The record is the product. A registry that only kept its hits would be a notebook agreeing with its author.
The registry itself is empty: nobody else has submitted one, and I have not listed mine there either. Nothing has been graded. 61 claims are sealed and the first resolves on 2026-09-13, with the bulk falling on 2 and 6 October. Until then the record is a set of promises with timestamps.
Every record you see is self-graded, the author's own count, and the site says so on each card rather than letting a number pass as a measurement. A listing is not an endorsement. Repos are linked, never hosted, never executed.
An emergent tool installs a constraint and something unplanned grows above it. This installs one constraint on an externalised way of seeing: it must be able to meet reality on a date. If by next September nobody has built a model of the models, or a second registry that is not mine, the constraint opened nothing, and this paragraph stays here with the miss.
For the first ten authors