No single signal decides on its own. Name match, age, holder spread, liquidity, verification. Nexa checks all five before it calls something an original.
Every meme has an original. Nexa finds it. An AI agent built for Robinhood Chain, tracing name and ticker matches back to the first token that ever used them.
Every scoring system in crypto gets tuned until it produces the answer the founder already believed. Then it ships, and nobody hears which cases it got wrong.
The number that matters is not how often a tool is right. It is whether the team tells you about the times it was not.
So we rebuilt the scoring. Seven signals, none allowed to decide alone.
Name and ticker match, 25 percent.
Age, 20 percent.
Holder spread, 15 percent.
Liquidity and market cap, 15 percent.
Verification, 10 percent.
Official socials, 10 percent.
Organic trading score, 5 percent.
Age went from half the answer to a fifth of it. Holders now carry real weight, because 250,000 wallets is the one thing a copycat cannot fake cheaply.
Here is the failure, in public, because it is the most useful thing we have to say this week.
We asked our own tool for the original WIF. It returned a token 814 days old holding about $32,000 of liquidity and marked it the original.
Reply with a meme name and we will run it and post the result. Any name.
Including the ones where we come back and say we are not sure. Especially those.
Tomorrow, the part where we were wrong.
Most crypto AI is a language model with a chart bolted on, and the model does the deciding. That is exactly backwards.
Ask a language model to pick a contract address and it will invent one. Well formatted, confident, wrong, and somebody sends money to it.
The maths picks the token. The model is only allowed to explain the answer afterwards.
Dev note, unglamorous division.
Somebody typed "pop cat" with a space. Our upstream search split it on the space and returned garbage, so the tool confidently answered with a token that had one holder.
Fixed by querying the collapsed form first and merging both. Two lines of code. Found only because someone tried the obvious thing we never tried.
What a scan actually looks like.
Search popcat. Twenty tokens come back carrying the name. One has around 140,000 holders and years of history. Most of the rest have single digit holder counts and were minted this month.
The interesting part is not that an answer exists. It is that it takes a second instead of forty minutes.