Foretell

The frogs taught him the difference the hard way, the first spring he trusted a pattern too far.

His pond-counter had studied five springs and found a tidy rule: the frogs arrive in the third week, every year, without fail. So it announced — boldly, proudly — that they would arrive in the third week again. They didn't. The spring came cold; the frogs came late; the counter had promised a date it had no right to promise.

Foretell sat with the failure a long time before he understood it. "It didn't learn the frogs," he said at last. "It learned the third week. It memorized five old springs so perfectly that it mistook the rehearsal for the real thing." That was the day he learned to keep two piles apart forever: the examples a model practices on, and a fresh, hidden pile it has never touched — the ones you test it against. "If you only ever check a model on the exact things it studied," he says, "of course it looks brilliant. That's not knowing. That's reciting. The real question is always the held-back pile — the springs it never saw." Overfit to the past, he calls it: so busy memorizing yesterday that it can't recognize tomorrow.

When he was twelve, Foretell hopped to NeuralQuest, and Sift the old owl met him at the gate.

"When you leap into the not-yet," Sift asked, "what do you owe the one waiting on your answer?"

"The word guess, out loud," Foretell said, "and the reason underneath it, so they can decide how hard to lean. And one more thing — I owe them a warning about how far the leap is. Guessing about something like the pile I studied is a short, safe hop. Guessing about something nothing like it is a leap in the dark, and I have to say so." He lifted the empty spyglass. "A prediction that hides that it's a prediction is a trick. A prediction that hides how far outside the practice it's reaching is a dangerous trick. I won't play either."

Sift blinked slow and pleased. "Then every leap into tomorrow in this place lands through you."

In his workshop, a kid held out a photo the model had truly never seen. "Watch the leap," Foretell said. The tool looked, thought, and reported: "My guess: dog. Here's why — four legs, floppy ears, a tail mid-wag. You can check the reasons."

"See how it named the guess a guess — and showed its work?" Foretell beamed. "That photo lived inside the fence of what it practiced on. Dogs like dogs it had seen. A short hop. Trustworthy." Then he pointed the empty spyglass at a harder card — an animal from far outside everything in its pile — and the tool leapt anyway and landed wrong, calling a fox a small dog, still sounding perfectly sure. "There's the honest danger," Foretell said quietly. "It didn't get stupid. It got asked to leap outside the fence — to guess about a kind of thing it never studied — and it leapt blind, because leaping is the only thing it knows how to do." He taught the whole habit then, as one steady rule: a model only really knows the kinds of things it practiced on, so its best guesses are about cases like the past; always keep a hidden pile it never saw, and judge it only on that; before you trust a leap, ask whether the new thing is even the same kind as the practice, or a jump into the unknown; and never — never — let a guess about tomorrow put on the costume of a fact about today. "The scariest predictions," he said firmly, "were never the wrong ones. They're the wrong ones that sounded sure, made about things the model had no business guessing at."