Notice the mental models in what you read
Metaphoria is a tool that notices when a piece of writing is leaning on a mental model — sunk cost, survivorship bias, correlation mistaken for causation — and points it out while you’re reading. Not as a quiz, and not as a verdict on whether the writing is any good. It just says: this argument is doing this thing, here.
The point isn’t the tool. It’s that after enough of these get pointed out to you, you start catching them yourself.
What is a mental model?
A mental model is a reusable shape of reasoning. Sunk cost is one: we’ve already put two years into this, we can’t stop now. Survivorship bias is another: looking at the companies that made it and concluding their habits are what made them. Once you know the shape, you see it everywhere — in earnings calls, in postmortems, in your own drafts.
They aren’t errors, exactly. Most of them are shortcuts that work often enough to be worth keeping. The problem is that they run whether or not they apply, and they run quietly.
Why is it hard to notice them yourself?
Because reading and evaluating are different jobs, and reading wins. When a paragraph is well written, you follow the argument rather than audit it. People who can name every bias on a list still miss them in live text — knowing the definition and catching it mid-sentence are separate skills.
The gap is worst in the interesting cases. When an argument announces itself — “this is a classic sunk cost fallacy” — nobody needs help. What’s hard is the paragraph that uses the model without ever naming it, where that reasoning is holding up the whole argument and nobody says so.
How does Metaphoria work?
You give it text. It returns which mental models are present, where in the text they show up, and how strongly. The interesting range is the middle: models that are genuinely there but never named. Catching those is the actual bar, and it’s the one the system is measured against.
Measured is the operative word. Every change to the detector is scored against a set of labeled examples — real passages, with the models they contain known in advance — so “this version is better” is a claim with evidence behind it rather than an impression.
What are you building?
A browser extension: highlight a passage on any page, see the models at play in it. That’s the delivery surface. Most of the work right now is on the detector underneath it and on the evaluation set that says whether the detector is any good.
There’s nothing to sign up for yet, and no date to announce. The blog is where findings from building it will show up — including data from running the detector over real text. More on the reasoning behind the project is on the about page.