the crypto web, read and proven
Every page a surveyor marks goes into one corpus. The Prover is a character-level neural language model pretrained on that corpus from random weights — no pretrained checkpoint, no fine-tune of someone else's model.
Below, it trains in this tab. The corpus is fetched from the same endpoint the plates are served from, the arithmetic runs in a worker on your own machine, and the curve moves because your browser is moving it. Close the tab and the run ends.
The model writes nothing before it is trained. Start a run and this fills with characters sampled from the weights as they stand at that step.
Each page carries the sha256 of the text stored for it. Open the source link on the live plate, take the hash of the text yourself, and compare: the corpus trained on here is the corpus you can check.
A window of characters is looked up in an embedding table, concatenated, pushed through one tanh layer, and projected to a distribution over the vocabulary. Adam, cosine schedule after a short warmup. Loss starts at ln(vocabulary) — the value of knowing nothing — and everything below that line was learned from these pages.
The weights download as one file: a short JSON header, then the parameters as float32 in the order embedding, hidden weight, hidden bias, output weight, output bias.