How neural networks restore, place, date and contextualise ancient texts — explained from first principles, then put to work across the corpora.
The building blocks made tangible — embedding, attention, the real 384-dimension vector, the position code, beam search — each you can poke at.
The two long roads that had to meet: the machine (RNN → word2vec → Transformer → BERT → T5 → Pythia) and the data (I.PHI, LED, DDbDP).
How a neural network restores ancient Greek, worked end to end on one real I.PHI inscription.
Contextualising Latin with a generative network, and how Ithaca grew into Aeneas (retrieval, vision, T5 + RoPE).
The merged narrative in four acts — Concept Lab → Lineage → Ithaca → Aeneas — with live concept-map jumping.
One neural torso trained jointly on Greek inscriptions (I.PHI), Latin inscriptions (I.Sicily) and Greek papyri (DDbDP) — restores, dates and locates offline, in your browser.
The cross-evidence index (~669k inscriptions, papyri, literature and editions): search, restore, attribute, connect — and co-work grounded research notes.
The first leakage-free comparison against Ithaca, Aeneas and a commodity LLM — reporting calibration (ECE) and selective prediction (AURC), not just accuracy.
What we can and cannot claim with numbers — the held-out eval, abstention patterns, and the identity/concept-bridge correction.