The science · How ProfilEd thinks
Most AI responds every time you ask. ProfilEd is designed to stay quiet until it has earned the right to speak. This page shows exactly how it decides, including the real numbers, and the parts we haven’t built yet.
No account needed · about five minutes
The central principle
“The product earns the right to speak by the completeness and accuracy of its model of the person. Every output — email, signal, curriculum recommendation, check-in opening — is a claim. Claims require evidence. Evidence requires confidence. Confidence requires accumulated, cross-validated data.”
“A product that speaks rarely but precisely builds a different kind of relationship than one that speaks often and is sometimes off. ProfilEd’s moat is precision, not volume.”
Three questions
What you’re wrestling with now. Whether you’re in a state to be challenged. What you left open last time.
Not just which gap you have, but how you learn. The same gap needs a different approach for different people.
Different people return for different reasons. ProfilEd is designed to learn which kind of reminder actually helps you, and to stop sending the kind that doesn’t.
What it pays attention to
In 1868, F. C. Donders showed that how long someone takes to respond reveals the thinking in between.2 ProfilEd listens to more than your words.
None of these is a verdict on its own. A signal only becomes something ProfilEd says to you after it repeats, and passes the rules below.
How you learn
The curriculum doesn’t change. How it’s presented to you does: the difficulty of the opening question, whether you’re asked to explain, and how a next step is framed.
| You learn best when it… | So ProfilEd… | What doesn’t work for you |
|---|---|---|
| “Shows me it worked for someone like me” | Leads with evidence and real stakes | Generic “you should learn X” |
| “Shows me the milestone” | Makes progress concrete: “at this step you’ll be able to do Y” | Vague outcomes and no visible progression |
| “Shows me why” | Asks you to explain and connect ideas, Feynman-style | Surface-level content with no depth |
It only decides which of these fits you once it has seen the same signal in three or more sessions. Until then, it doesn’t guess.
What it knows, and how sure it is
Example: someone six sessions in
Every output checks these levels first. Low confidence means it doesn’t make that kind of claim yet.
What it’s allowed to say, session by session
At any session count: “You’ve resolved this” is never claimed from what you said alone. It needs evidence from your practice.
For the technically curious
These are the real thresholds each kind of output has to pass before it reaches you.
| Before ProfilEd says… | …it needs |
|---|---|
| Your check-in opening question | Built from your last session only if that session was within 48 hours. Otherwise, an open question. |
| Your daily question by email | Medium confidence or higher in what you’re working through, and the question has to be a genuine question. |
| An invitation to come back | Your last session had real substance: it has to clear a minimum quality score. |
| Naming a pattern in how you think | Seen in 3+ sessions, high confidence, and the AI’s output passes validation. |
| Your weekly arc | 4+ substantial sessions this week, good average quality across them, and the output passes validation. |
| Your two-week arc | High confidence, based on 10+ sessions, and the things you’ve been working through are genuinely different from each other. |
| “You keep circling this” | Your direction reads as circling, your recent sessions are more than 85% alike, and 10+ sessions in total. |
| A milestone prediction | Your pace has been steady for 3+ weeks, not just this one. |
| “You’ve resolved this” | Confirmed by your work in the Curriculum Graph. Never claimed from language alone. |
| A curriculum recommendation | Only once it knows how you learn best. |
say it⇔inputs are real∧output passes checks∧confidence ≥ what this claim needs
The quality gate
“If any answer is no: log the reason, send nothing, never degrade to a fallback.”
“It is not defensive code. It is the product’s epistemic standard — the code expression of ‘we only say what we know.’” — ProfilEd’s architecture spec, on the quality gate
What it keeps
Everything ProfilEd concludes about you can be traced back to what you actually said and did.
Not all mistakes cost the same
“Being wrong about a curriculum recommendation costs a session. Being wrong about a pattern insight costs the relationship.”
So the more personal the claim, the more evidence it needs.
Honest limits
The first two are the gaps our own architecture document names as the next things to build. The third is already under way.
“These two gaps [the first two above] are what turn ProfilEd from a system that makes intelligent guesses into a system that learns whether its guesses were right. The data wins over the theory.”
The standard
“This is the intelligence quest. Not ‘how do we send more personalized emails.’ But ‘how do we build a complete enough model of each person that everything the product says is earned.’”
“Do not say that for which you lack adequate evidence.”
Just your first name. Your first check-in is free, and you’ll leave with something to do tomorrow.
No account needed · about five minutes