How Model Zero works, and what it can't do

Model Zero studies Pump.fun launches (and StonkFun launches). For each launch it records a snapshot at fixed ages (1m, 5m, 15m, 30m, 1h) made of 36 numbers: the price path, volume and its acceleration, buy versus sell flow, how many wallets traded and how many still hold (estimated from trades), how concentrated those holdings are, what the creator wallet bought and sold, how many tokens that creator launched before, and whether socials were set at launch.

Launches are picked every minute by a hash of their address, so the dataset includes the many launches that go nowhere, not just the famous ones. Every trade is read from the Solana blockchain itself. So far Model Zero has tracked 123,036 launches and 12,536,718 trades.

It also records the whole market every few minutes: which coins are trending across every Solana launchpad, how many coins with the same name are launching, and how hot the market is. Each launch is compared only with what was trending at that moment.

It does not know the future, who is behind a wallet, private group chats, off-chain coordination, news, or a project's intentions. It cannot tell you whether to buy or sell, and it never will.

Holder concentration in the model is estimated from trade flow (tokens bought minus sold per wallet). It misses transfers between wallets and can't link wallets controlled by one person. Wallet labels such as “insider” or “smart money” are not used, because the evidence for them would be weak. When Model Zero mentions wallets with a good record, it means a positive realized result across launches in its own sample, and says so.

The launch model learns from a coin's first hour. Coins older than two hours are covered by a separate established-coin model, trained on the market lists and what happened to those coins afterwards; when neither applies, the analysis shows live facts and no estimate.

Every statement on the Analyze and Talk pages carries one of these labels.

To compare a token, Model Zero takes launches in its dataset at the closest snapshot age, puts the main setup measures on the same scale, and picks the nearest ones (about 15% of the pool, at least 10). It then reports what share of them migrated, doubled, fell more than 70%, or were still trading. These are frequencies in a sample. A setup that ended badly 60% of the time still ended well 40% of the time.

A model that sees the future during training looks brilliant in testing and is useless in practice. Every snapshot has a hard time boundary. Features may only use trades at or before it, price candles that had closed by it, and creator launches that happened before the token launched. The code that builds features throws away anything later itself, and the database refuses to store a snapshot whose latest input is after its boundary.

What happened next — the price path, migration, creator exits — is stored separately as labels and only uses data strictly after the boundary. A label stays empty until its whole horizon has been observed. Automated tests feed the feature builder future data and check that the output doesn't change.

Live analysis uses the same feature code as training, so the model sees the same kind of inputs in both places.

Launches are split once, by a hash of their address: 70% training, 10% validation, 20% locked test. All snapshots of a launch land in the same split. Test launches are never written into a training dataset. Once enough of them have observed outcomes (at least 30 snapshots), they are frozen into a locked test set with a content hash. Benchmarks are scored only on launches that are active at the snapshot (at least 10 trades so far and some volume in the last 5 minutes — known at that moment). About three quarters of launches stop trading within minutes, and “a dead token stays dead” is too easy to count as skill, and every model version is scored on the same set. When a larger set is locked later, every earlier version is re-scored on it so the history stays comparable.

Scores are AUC for yes/no questions (0.5 = coin flip) and balanced accuracy for trajectories (≈33% = chance), with 90% bootstrap ranges computed over launches. There is no single “AI score”. A new version is deployed only if its average across the benchmarks it shares with the live version is at least as good; otherwise the old one stays live and the regression is recorded. Failed runs and worse versions are kept forever.

Datasets, benchmark sets, ledger entries, benchmark results and finished versions can't be edited or deleted; the database enforces it. Given the same dataset, feature version, configuration and seed, training produces a byte-identical model file.

The Model Zero token earns creator rewards from trading on pump.fun. Rewards are recorded in the ledger only after their on-chain transaction is verified. They accumulate in the training treasury. When the treasury reaches the training target ($50.00 currently) and enough new outcomes have been observed, one training generation starts: compute is purchased, a new version is trained and benchmarked, and the result is published — better or worse.

Rewards don't trigger training trade by trade; they are batched into generations so each run is meaningful and each cost is visible.

Pump.fun launches are noisy. Two launches that look identical at 15 minutes regularly end in opposite places. The honest output is a base rate for setups like this one, with the size of the sample it came from. Model Zero reports those and labels them; it does not promise outcomes.

See the ledger, training runs and activity log.