predictingthepast

Mark known-length gaps with ? and unknown-length gaps with # in your text, then run the restoration task with the --restore flag. The model returns ranked beam-search candidates with scores, and the top prediction is shown with restored characters highlighted.

Aeneas processes Latin inscriptions while Ithaca handles Ancient Greek texts. Both perform restoration, geographical and chronological attribution, contextualization, and embedding, but use language-specific alphabets, preprocessing rules, and training datasets.

Model checkpoints, datasets, and retrieval embeddings are downloaded once from Google Cloud Storage on first run, so initial setup needs internet access. The skill otherwise runs inference locally and explicitly forbids web search or external tools during analysis.

Restoration time scales roughly linearly at about 10 seconds per additional ? character on a high-end CPU, so 20 gaps take around 5 minutes. Texts over 750 characters are also split into overlapping windows processed independently, compounding the runtime.

Input must be at least 25 characters after preprocessing; pad shorter texts with dashes. Texts exceeding 750 characters are automatically split into overlapping windows, with attribution scores averaged across windows.

Pass the --contextualize_exclude_test_valid flag when running inference. This filters contextualization results on the internal numeric id, excluding entries where id modulo 10 equals 3 (test) or 4 (validation).