The Arc is where the collection started: the RDF tinyshakespeare. Train a 2.86M model from scratch and, instead of scoring a final number, watch RDF structure appear in layers, each rung lighting roughly in order.
Every epoch, from the same fixed prompt and seed, the model is scored on six things that should arrive in sequence:
| rung | what it means | lights |
|---|---|---|
| L1 syntax | emits parseable Turtle at all | ~epoch 5 |
| L2 vocabulary | IRIs are real ontology terms, not gibberish | ~epoch 5 |
| L3 schema | hasType takes a PokéType; ranges respected | ~epoch 5 |
| L4 shape | a Species gets the canonical field set | early, climbing |
| L5 reasoning | weakTo is the chart's verdict on its own types | ~epoch 15 |
| L6 reality | knows pikachu is real and zarquon isn't | ~epoch 5 |
Grammar, vocabulary and schema finish within the first few epochs; the sample stops being
garbage bytes and becomes well-formed triples fast. Reasoning is the slow one. The
weakTo lines stay random long after everything around them is correct, and only
lock onto the type chart around epoch 15 (0.000 to 1.000). The words are right long before the
meaning is.
One detail worth noting: L6 (reality) lights early, with the vocabulary rather than the reasoning. Knowing which names the world contains is a property of names, and names are what byte-BPE learns first.
These are real samples captured during training, one per epoch, same prompt, same seed (temperature 0.7), so the only thing changing as you drag the slider is the model itself. The final epoch is not a picture: it is the same weights running live in the Minter and Reasoner. The Arc shows how they got that way.