TRACK RECORD

Every signal the radar fires here is graded by the stream that produced it: the token's price at the first trade one, five, fifteen and sixty minutes later, against the signal's own fill price. Nothing is fitted after the fact. Net subtracts the 2.5% round trip set on your copy plan; a hit is a grade at or above +10%, the bar a bonding-curve round trip needs to be green.

A logistic regression over the radar's graded record: what was known about a signal the moment it fired (the wallet's score and the sells behind it, the size, how far up the curve, minutes since launch, the hour, how busy the wallet and the mint were) against whether its five-minute grade cleared +10%. Trained on the older part of the record and judged only on the newer part it never saw, in time order. Its verdict is one of three words, and edge needs its top quarter to beat following every signal by two standard errors. It needs 200 usable graded signals before it says anything at all. Nothing here is a trade instruction; a probability on a signal is a guess, and only the forward record can earn it trust.

Every scan pass writes down what it saw and what it cost. Later passes supply the price that resolves it, so this is a forward test — nothing here is fitted after the fact, and no observation is scored against a price that existed when it was made.

The column that matters is lift: a score band’s return minus the average of everything this scanner listed over the same window. In a rising market every band goes up, and reporting that as skill would be crediting the market to the model. Intervals resample whole scan passes, never individual rows — twelve tokens seen in one pass share a market and a direction, and treating them as twelve independent trials narrows the interval by an amount nobody earned.

The ledger lives in this browser only. Nothing is uploaded. Observations older than 30 days are dropped.