Does your confidence match reality?

Think of the last time you were sure — "I'm 90% certain we'll close this by Friday." Now think about how often that kind of "90% certain" actually pans out. For most of us, honestly, it's a lot less than 90% of the time.

That gap stays invisible for one reason: we don't keep track. After the fact, memory quietly rewrites itself — "I always knew that was a stretch" — and the lesson evaporates. The only fix is almost embarrassingly simple: write the prediction down, with a number, before you know the answer. That's the whole mechanic here.

This is the one idea everything else rests on, and most people have never had it spelled out. Picture a weather forecaster who says "70% chance of rain." What would make that a good forecast?

Not whether it rains today. A "70% chance" isn't a promise that it will rain — it's a claim that on days like this, it rains about 7 times out of 10. So she's exactly right if, across all the days she says "70%," it rains on about 70% of them. The dry days aren't her being wrong — they're her forecast coming true. Three days in ten, it was supposed to stay dry.

So no single prediction can be "calibrated" or not — only a whole track record can. That's why Ivyra needs a little history before it can tell you much, and why it gets sharper the longer you use it.

When a prediction resolves, we measure one thing: how far your confidence was from what happened. Say 90% and it happens — you were barely off. Say 90% and it doesn't — you were badly off. Average that across all your predictions and you get your score.

Lower is better, like golf. A score of 0.25 is the "I'm just guessing" line — exactly what you'd get by shrugging "50/50" at everything. Beat it, and your confidence is carrying real information about the world.

Your score says how good your calls are. Your bias says which direction they're off — the gap between how confident you felt and how often you were actually right.

"You run 19 points overconfident" means that, on average, reality came in about 19 percentage points below your stated confidence. The opposite reading means you were too cautious, and things happened more often than you claimed.

The calibration curve puts your confidence along the bottom and how often things actually happened up the side. If you're spot-on, every dot lands on the diagonal — your 70%s happen about 70% of the time.

Dots below the line mean overconfident: things happened less often than you claimed. Above the line means underconfident — you knew more than you let on. Flip between the shapes below to see it.

Here's the part almost everyone misses. There are two separate ways your confidence can fail, and most people only know the first.

One: your numbers can be dishonest. You say 90% and things happen 60% of the time. That's the overconfidence we've been talking about, and the curve catches it.

Two: your numbers can be empty. Imagine answering "60%" to everything. You're never badly wrong — but you're never really saying anything either. You've made yourself safe and useless.

Boldness measures whether your confidence levels actually tell your outcomes apart. Do the things you call 80% really happen more often than the things you call 55%? If they do, your numbers carry information. If everything you say clusters near 50%, they don't — however honest each one looks on its own.

Confidence that's honest and decisive at the same time moves to high and low values as the evidence earns it, and those values match how often things actually happen. Confidence that stays hedged near the middle is safe but says nothing — a hedger's 80% calls land no more often than their 55% ones.

Your insights page gives you two different things, and it's worth knowing which is which.

The verdict is the one-line summary at the top — "you lean overconfident," "calibrated and bold." It only ever describes what's true about your track record. Fixed math produces it; it states a fact and stops there.

The insight goes further: it explains why, and what to do differently. This is the one place AI helps — it reads your own predictions and the words you wrote, names the pattern behind the numbers, and points to where you can adjust. It never touches your score.

You can look at your whole history or just your recent calls — and the two can honestly disagree. Your lifetime read might say "overconfident" while your last stretch says "calibrated," because you've been getting better.

That's the point, not a glitch: recent form shows whether the training is working, while lifetime shows the deeper habit. Neither one is the "real" number — they answer different questions.