reading the ledger
Same configuration as v0.65 (best so far on the headline benchmark), retrained on the newest dataset.
For each question, the signals the model leans on most. Measured by hiding one signal at a time on launches it didn't train on and seeing how much worse it gets; the percentage is that signal's share among the five shown.
The launch model above covers a coin's first hour. For older coins, Model Zero trains a second model on its own record of Solana's trending, most-traded and most-organic lists (every two to ten minutes) and the price of every coin in them afterwards. It answers: does it drop 30% in the next hour, halve or rise 50% within six hours, and is it higher a day later. Scored on coins it never trained on.
Scores from 0 to 1: pick one launch that crashed and one that didn't; the score is how often Model Zero rated the crashing one riskier. 0.5 is a coin flip, 1.0 is perfect. Direction calls are the share called right (a guess gets about 33%).
“Likely range” is where the score would land 90% of the time on a different sample of launches. Questions without enough examples yet say so instead of showing a number.
Both versions scored on the same held-out coins. “New” = a question the first model wasn't asked.
Versions are never overwritten or deleted, including failed ones and ones that got worse.