Patterns, stated plainly

A TWIN is a statistical description of how one public wallet has historically entered trades: the market caps it buys at, how old tokens usually are, how much liquidity is present, what narratives keep recurring, and how large the buys tend to be. It is a model of a pattern, not a person, and it is built only from activity that is already public on-chain.

For each historical buy, TWIN reconstructs what the token looked like at that exact moment. Anything that cannot be sourced for that timestamp is marked UNKNOWN. Today's numbers are never written in as if they were historical, because that would flatter the model.

Every value carries a level. FACT comes straight from the chain or the transaction. INFERRED is calculated or classified from other data. UNKNOWN means TWIN could not source it, and it stays visible rather than being quietly filled in.

The match score is deterministic. Each feature is compared against the model's historical bands and weighted; scores are capped when too few features are available, so a thin comparison can never claim a high number. Language models are used only to name a narrative or summarize context — they never set a score and never claim what a trader intends to do.

Trades below a minimum value are excluded so that spam, test transactions and dust cannot shape a model.

A model needs a minimum number of clean trades before it is treated as trained. Until then it is labelled LEARNING or LOW DATA and its validation numbers read LOW DATA rather than a flattering figure.

Every signal is recorded before the outcome is known, and misses are never removed. Accuracy on this site means behavioral-match accuracy — how often the real wallet later bought what the twin flagged — not investment returns.

Connecting a wallet means signing a free message that proves you control the address. No SOL is spent, no transaction is submitted and no spending approval is requested. TWIN never takes custody, never holds funds and never trades on your behalf.