§5 Method Checkpoint 0 · Solana
Method
Four steps, each one started by an event and never by a clock. There are no rounds, no countdowns and no deadlines anywhere in this project: a packet of fees arrives, a piece of training is bought, a checkpoint lands, a vote closes.
- 1. The dataset. The keeper reads Solana coins from Birdeye's token list and Jupiter's token API, keeps name and ticker, and marks a coin ALIVE once a reading of its market cap has been at or above $30,000. Junk, links and a list of blocked words are dropped, and the same filter is applied to what the model generates.
- 2. The packet. Creator fees from $SPI are forwarded to the keeper's wallet. Anything above the 0.05 SOL it keeps for network fees counts as forwarded, and every 0.5 SOL of it is one packet: 60 per cent buys GPU time, 30 per cent buys $SPI back and burns it, 10 per cent pays for RPC, hosting and Hugging Face.
- 3. The checkpoint. The GPU runs the number of steps that money bought. The weights go to Hugging Face, their sha256 goes into a memo on Solana, and the checkpoint publishes its loss and ten names with tickers, sampled with a seed and a temperature that are printed next to them.
- 4. The vote. A wallet holding $20 of $SPI signs one sentence for one of the ten names. It is a signature, not a transaction, so it costs nothing. When the next checkpoint lands, the vote closes and the winner is written into a memo and into the Hall of names.
Table 7Where a packet goes
Per 0.5 SOL of forwarded creator fees.
| Share | Part | SOL per packet | What it pays for |
|---|---|---|---|
| Training | 60% | 0.300 | GPU hours at the provider, a fixed number of steps for that money |
| Buyback and burn | 30% | 0.150 | $SPI bought through Jupiter, which routes to the bonding curve or to PumpSwap, then burned, with both signatures printed |
| Work | 10% | 0.050 | RPC, hosting, Hugging Face storage |
§5.1The model, and how it grows
A character level transformer in the nanoGPT shape. It reads a row one character at a time and predicts the next one.
Before launch the run is small and honest: a model of about 0.35 million parameters trained on a CPU, so the curve and the first ten names are on this site from the first minute and cost no SOL. The first paid packet moves the run to the 10 million parameter model on the rented GPU. After that the model grows with the market cap of $SPI, which is an event and not a date: above $100,000 the run continues at 30 million parameters, above $500,000 at 100 million. Each move starts a new line on Fig. 1, and the log says plainly which model each checkpoint belongs to.
| Stage | Parameters | Starts when |
|---|---|---|
| Warm up, on a CPU | 0.35 M | Before launch, already done |
| The run | 10 M | The first 0.5 SOL packet arrives |
| Second model | 30 M | $SPI market cap above $100,000 |
| Third model | 100 M | $SPI market cap above $500,000 |
§5.2What can go wrong, said plainly
Written here before it happens, so nothing has to be explained away later.
| Risk | What actually happens |
|---|---|
| The GPU provider fails | The piece of training is bought again at the backup provider. The keeper holds the SOL until a checkpoint it can verify comes back, and it re-scores every returned checkpoint on the held out rows with its own code before publishing it. |
| The model writes something offensive | The dataset filter and the output filter are the same code. A flagged name never reaches a vote, and the count of flagged samples is printed under Table 1. |
| The loss stops falling | Fig. 1 shows it, with no smoothing. Growing the model is what gives a new descent, and the log says when that happened. |
| Bot votes | The $20 threshold and one wallet one vote. A vote wins nothing but a line in the Hall of names, which is the cheapest possible thing to attack. |