A configured assistant, or a theory that can be wrong

Custom GPTs and garden models both look like "a thing you make and share". One is a product configuration; the other is a claim about the world with a stated failure condition.

The short answer

A custom GPT is judged by whether people like using it. A model is judged by whether its predictions hold.

The distribution model differs, and that turns out to matter more than it sounds.

A custom GPT lives in a store, is discovered by popularity and rating, and its quality signal is engagement. That is the correct design for a tool. It is the wrong design for a claim about reality, because whatever number is displayed becomes the thing everyone optimises — and popularity is far easier to move than accuracy.

The garden has no ranking at all, deliberately: no stars, no downloads, no leaderboard. A model's card shows counts of claims made and graded, never a score. You find models by filtering on domain and what you can actually run, not by position in a list.

And a model is a repo, not a listing. Fork it, change the premises, keep what works — your copy starts with no record, because a track record belongs to whoever earned it and does not transfer by being copied.

Side by side

Custom GPTs	Model Meets Reality
Discovered by popularity and rating	No ranking of any kind — filter, don't sort
Quality signal is engagement	Record is counts of claims made and graded
Lives in one store	A git repo you clone, fork and edit
Configured behaviour	Premises, falsifiers, deletion clause
Popular ones stay visible	Refuted ones stay listed, with the record showing

What Custom GPTs does better

Far easier to build and share, with a real distribution channel and an audience already there. Custom GPTs can call APIs and take actions; a model is a document and does nothing on its own. If you want a working tool in front of users this week, build a GPT.

Where this stands today

Said plainly, because the whole point of this site is not overclaiming. The Model Garden is opening small — no models listed as of September 2026. Every record shown is self-graded: the author's own count of claims made and resolved, labelled as such on each card. Nothing here is ranked, and no independent resolution layer exists yet. Custom GPTs has things this does not, named above rather than omitted. Compare the designs, not the scoreboards — there is no scoreboard.

https://github.com/someone/their-model Help me use this

Paste that into any assistant. It reads the model — premises, falsifiers, deletion clause — and reasons through it.

The difference that is not on the table

Custom GPTs answers your question. It does not keep a record of whether the answer held. That is not a flaw — it is not what an assistant is for. But if you have watched one thing closely for years, that knowledge is already a model: premises, predictions, blind spots. Writing it down turns it into something an assistant can reason through, and something a date can settle.

Share your understanding →

Browse the Model Garden Publish your own