korg · the next demo, in plain words

AI that can't lie about what it called.

Before we build it — here's the whole idea, as simply as I can put it. Read it, poke holes, and we'll tweak from here.

the trust layer for AI · proof you can check yourself · works with any AI · no crypto, no catch

01 the problem

AIs brag about predictions. "I called that crash!" But you can't check.

Did it really predict it — before it happened? Or did it quietly change its story afterward? Right now an AI can lie about its own past, and you'd never know.

02 the fix: a sealed envelope

Make the AI seal its prediction in an envelope before the event. Don't open it until after. If what's inside matches the sealed version, that's proof — it called it, and it couldn't have changed it. That's the entire idea.

🔒
1 · Seal
AI locks its prediction before the event. The contents stay hidden.
⏳
2 · Wait
The event happens. The AI can't touch the sealed envelope.
🔓
3 · Open
Reveal it. Anyone checks it matches the seal. No take-backs.

03 try it yourself

the AI's prediction
the public list (nobody can secretly edit this)
🔒 sealed · fingerprint —

04 why this is a big deal

Do this for every prediction and the AI builds a track record it can't fake. Now you can answer a question nobody can answer today: which AI is actually good at predicting — proven, not bragged.

Line up Claude, GPT, and Gemini, same questions, sealed envelopes, and let the receipts decide. And that's the first brick of something bigger: AI you can actually trust, hire, and pay — because it has a history it can't lie about.

To be straight with you: this proves the AI locked in its guess before the event and didn't change it — not that the guess was smart. When we can't prove something, we say so. We never fake it.

05 the one trick

It's one idea you already understand.

You can't change what's inside a sealed envelope. korg is just (1) the seal and (2) a list where, if anyone changes it, it shows. No crypto, no catch, nothing complicated — it's a plain file, and your own phone or laptop does the checking.

06 so what we'll build

A track record you can verify yourself.
  1. A scorecard page — one AI's predictions: each one sealed-before, opened-after, with a green ✓ you can check.
  2. The leaderboard — real AIs, real questions, provable scores. "the honest AI predictor rankings."

Everything above already works in korg (the seal + the list-nobody-can-secretly-rewrite are the same pieces behind the poker demo). This just points them at predictions.