Model Meets Reality · publish

You already model the world. Write one down.

Everyone carries working theories — about their industry, their city, their field, the thing they have watched closely for ten years. Most never leave the person holding them, and so never get tested.

You know how funding actually moves in your sector. Why the reorganisation failed in a way the post-mortem missed. Which planning applications get approved and which quietly die. Why your neighbourhood's traffic scheme did the opposite of what was promised.

That knowledge is a model. It has premises, it makes predictions, and it is probably right about some things and wrong about others — you just have no way to find out which.

An assistant answers your question and moves on. Nothing keeps a record of whether the answer held. Here you write down how you think something works, say what would prove it wrong — and the date arrives.

What you can't do anywhere else

Graded on a date, not rated

No stars, no upvotes, no leaderboard. Your model is checked against what actually happened — on a date you set in advance, provable from your repo's git history rather than taken on trust.

Be wrong in public without disappearing

A model graded wrong stays listed with its record showing. Everywhere else, being publicly wrong is a reason to delete the post. Here it is the most useful thing you can contribute.

Own it completely

Your model is a git repo under your account. No platform holds it, no algorithm ranks it, and nothing here can edit or delete it. You can walk away with it.

Stay anonymous if you want

No account here, no email, no tracking. Publishing stores a repo URL and a date — nothing about you. Use whatever GitHub identity you like.

The narrow field is the point

The most valuable models here will not be about markets or geopolitics. They will be about things almost nobody understands, held by the handful of people who do.

Someone who has run a small-town water utility for fifteen years knows how infrastructure decisions actually get made in a way no consultancy report captures. Somebody who has watched one niche industry through three cycles can say which signals lead and which lag. That knowledge does not scale, does not publish, and mostly dies with the career.

A narrow model that is right beats a broad one that is vague. Because nothing here is ranked, a model of one small domain is not buried under popular ones — people find it by filtering for the domain they actually need. Twelve resolved claims about something specific is a stronger record than a hundred about everything.

Not every model predicts

The word "model" here does not only mean a forecast. Most of what people actually know is a way of reading something, not a bet on it — and those are publishable too, as long as they can be wrong about something observable.

A forecaster

Says what happens next. "Applications of this type get approved within two cycles; these ones quietly die."

A classifier

Sorts things others lump together. "These three failures look identical and have completely different causes." Wrong when a case it sorted one way behaves like the other.

A tracer

Says how something travels — which decision moves which other decision, and in what order. Wrong when the chain runs the other way.

A tracker

Says which signals lead and which lag in a domain you have watched for years. Wrong when the leading one stops leading.

An adversary

Exists to attack a common belief and say what would refute it. A model whose job is to be the counter-case.

A model of models

Reads not the world but how other people read it — where a whole field is systematically blind. One level up, and the same rules apply.

Ten kinds in all — forecaster, classifier, tracer, finder, tracker, generator, attributor, adversary, mirror, timer — and three levels: modelling the world, modelling models, modelling the modelling. You declare which yours is; nothing stops you being the first of a kind here. The one rule that does not bend is the same for all of them: it has to be able to be wrong about something you can observe, on a date. That is what separates a model from an opinion, and it is the only thing this site checks.

What publishing actually involves

Write down how you think it works
A few sentences. Not a paper — the premises you would say aloud if someone asked why things go the way they do in your area, or how you tell these cases apart, or which signal you actually watch.
Name one thing that would prove you wrong
The hard part, and the whole point. Something observable, with a date. "If X does not happen by March, my second premise is wrong." A model that sorts rather than predicts still has one: name the case that, if it behaved the other way, would break the sorting. If you cannot write this, the model is not ready — and that is worth knowing on its own.
Say when you would give up on it
The deletion clause: the conditions under which you would retire the model rather than patch it. Almost nobody writes this, which is why almost nothing gets retired.

Push it to GitHub and submit the link

Three files — your model, a small card, a licence. Nothing is uploaded here; the repo stays yours.

MODEL.md      your premises, falsifiers, deletion clause
model.json    a short machine-readable card
LICENSE       so others may actually use it

Then wait, and find out
Your claims resolve on the dates you set. Hits and misses both stay visible. That record is yours, and it is the only thing on this site that means anything.

Try it before you commit to anything

You do not need a repo, an account, or any setup to find out whether your idea holds up as a model. Paste this into ChatGPT, Claude, or Gemini:

I have a theory about <your area>. Help me turn it into a falsifiable model:
premises, at least one thing that would prove it wrong with a date, and the
conditions under which I should retire it. Push back if my premises are
unfalsifiable or if my test could not actually come out the other way.

My theory: <write it here>

If the assistant cannot find a way for your theory to be wrong, that is the finding. Plenty of confident beliefs turn out to be unfalsifiable, and discovering that costs you nothing here.

What this asks of you, honestly

Two things, said plainly. First, it asks you to be publicly wrong sometimes — that is not comfortable, and no amount of framing makes it comfortable. What it buys is finding out.

Second, the garden is opening small. There is no audience here yet, no independent grading layer, and every record is currently self-graded — the author's own count, labelled as such. If you publish now you are early, with everything that implies. We would rather say that than pretend otherwise.

Submit your model The format, in full See what is already there