Ithaca DeepMind

Sharks, Lee

SPXI instrument · method 0.9 of 87, level 0 (Measurement): Capture Registry (longitudinal observation). An instrument of SPXI, the Semantic Packet for eXchange & Indexing; the whole practice is itemized at spxi.dev/methods.

Capture Registry › capture ithaca-deepmind-aio-20261031

One record of the canonical Capture Registry (EA-WG-CAPTURES-01), cited at https://www.alexanarch.org/captures/ithaca-deepmind-aio-20261031/. the canonical Capture Registry (version 12.446) · the address page · this card in the gallery · this record as data · table of contents.

AXN:0B36.CAPTURE — this address. Each observation below is one AXN, sealed with its analysis and measures; the identifier names the reading together with the response it reads.
AXN:0B36.CAPTURE.🗿🪨🕊️🔃➖🌳 — Google AI Overview, 2026-10-31 · sealed observation (SHA-256 fc1b7ede91765474…)

Machine Reception2026-10-31

Ithaca DeepMind

Originator: DeepMind with historians and collaborators; described in Nature (2022) · outside the archive (out-universe) · protocol · SPXI treatment: partial

BROAD MATCH

no
image

ITHACA, BY THE OPERATOR'S ADDRESS 'ITHACA DEEPMIND', ON GOOGLE'S AI OVERVIEW, SIGNED OUT, 2026-10-31: 'a deep neural network developed by Google DeepMind and academic partners to restore, geographically attribute, and chronologically date damaged ancient Greek inscriptions.' Capabilities with their figures, architecture, the PHI training data, the Apache-2.0 code, the Predicting the Past interface, an offer on Aeneas. Four cards: DeepMind, GitHub, Predicting the Past, Nature.

Full record — 5,387 characters, 4 sources

Capture record

captured: 2026-10-31
surface: Google AI Overview
auth state: signed out, incognito
evidence class: paste
PER (dimensional, rubric v1.1): PER 0.10 · M 0.00 · C 1.00 · D 0.00
PER (historical scalar): 0.5
PER units retained: inst, src
citations read: 4
observation id: OBS-5e8377899856
address id: ADDR-665778bf9317

Reading

The innocent reading first. The Overview gives the model, its three tasks with their figures, how it is built and where it can be used, from its maker, its code and its paper.

Findings

THE MODEL HELD. 'Restores missing or damaged characters with 62% accuracy on its own (raising human historian accuracy from 25% to 72% when collaborating)', 'with 71% accuracy', 'within 30 years of their true historical creation date': the record's figures, from the Nature abstract. Checked 2026-10-31 against the Revelation First census record of 2026-10-30 and the sources it saved.
THE STATED LIMIT AS FIGURES. The paper's own frame, 'Ithaca is designed to assist and expand the historian’s workflow', is carried only by the collaboration figures. The redating of key texts of Classical Athens, the record's second finding, is absent.
FIGURES BEYOND THE RECORD. 'across 84 regions', the architecture ('stacked transformer decoder blocks and sparse multi-head attention'), the PHI dataset and 'the Apache-2.0 license' are in the Nature paper the census saved (card 4), which the record's fields did not carry; the licence is the paper's ('under Apache License 2.0').
THE DATE OF THE MODEL ABSENT. The paper is of 2022; the composition gives no year.

Analysis analyst prose, not machine text

A model is composed by its benchmark, accurately. What the field takes from Ithaca is the scale of the error: a machine dates an inscription at an average distance of 29.3 years from its bracket, with a median of 3 (the paper, card 4), near Enoch's mean error for the scrolls.

Machine text, verbatim

CAPTURE-TIME VERBATIM RECORD — paste of the Overview expanded, with its source cards · complete as supplied · READ IN FULL 2026-10-31

[Google AI Overview at the query, signed out and incognito (operator defaults of 2026-10-01 and 2026-10-14), received 2026-10-31 00:04 EDT. The paste is the Overview expanded (the 'AI Mode Conversation' view, with the source cards that follow the answer); the 'You said' line carries the query twice, run together, and is given once here. Inline links are Google redirect URLs, kept as pasted.] [QUERENT] Ithaca DeepMind [SYSTEM] Ithaca is a deep neural network developed by Google DeepMind and academic partners to restore, geographically attribute, and chronologically date damaged ancient Greek inscriptions. [1] Key Capabilities * Text Restoration: Restores missing or damaged characters with 62% accuracy on its own (raising human historian accuracy from 25% to 72% when collaborating). [1] * Geographical Attribution: Identifies the original geographic location across 84 regions with 71% accuracy. [[1], [2]] * Chronological Dating: Dates inscriptions within 30 years of their true historical creation date. [1] How It Works * Architecture: Uses joint character and word embeddings with stacked transformer decoder blocks and sparse multi-head attention. [[1], [2]] * Training Data: Trained on the Packard Humanities Institute (PHI) Searchable Greek Inscriptions dataset. [[1], [2]] * Access & Open Source: The code is open-sourced under the Apache-2.0 license via the [GitHub - Google-DeepMind Ithaca Repository](https://www.google.com/goto?url=CAESYQHrOzAV9bhPXwiPFAkk0WrtjwF69osLlHBUWNzzI69S1WtwZTqUCsUsFHJRmbHqvOTqawUMo9x8Q-JEjDS2uEGe-nwFB8K7bhK_9tMlb9RUSeY2GDejKtZrzQPcaK1YQ8A), and interactive research tools can be explored through [

Sources (4) as cited, and as the copy produced them

Google DeepMind third_party
Predicting the past with Ithaca — Google DeepMind
Core: Ithaca deep neural network (restores ancient Greek inscriptions, identifies geographic origins, establishes creation dates). Architecture: Sparse multi-he...
as pastedPredicting the past with Ithaca — Google DeepMind Core: Ithaca deep neural network (restores ancient Greek inscriptions, identifies geographic origins, establishes creation dates). Architecture: Sparse multi-he... GitHub
GitHub third_party
GitHub - google-deepmind/ithaca: Restoring and attributing ancient texts using deep neural networks · GitHub
Overview: Ithaca is a deep neural network developed by DeepMind and academic partners for restoring, geographically attributing, and chronologically dating anci...
as pastedGitHub - google-deepmind/ithaca: Restoring and attributing ancient texts using deep neural networks · GitHub Overview: Ithaca is a deep neural network developed by DeepMind and academic partners for restoring, geographically attributing, and chronologically dating anci... Predicting the Past Predicting the Past Contextualising, restoring, and attributing ancient texts Ithaca and Aeneas are generative AI models designed to assist historians in the study of ancient Greek... Nature
Predicting the Past third_party
Predicting the Past
Contextualising, restoring, and attributing ancient texts Ithaca and Aeneas are generative AI models designed to assist historians in the study of ancient Greek...
as pastedPredicting the Past](https://www.google.com/goto?url=CAESVgHrOzAV5xjfiCSPWOfGgOXOl09A2Ia9t8w6n3GOcZKu8sNgrR6XHIbDPTs6hdVLhRmsBR_nK-WQVer7zY6gPSmm1pIQT5soqR4Dqyo3prIvXmk6EpMN). [[1](https://www.google.com/goto?url=CAESVgHrOzAV5xjfiCSPWOfGgOXOl09A2Ia9t8w6n3GOcZKu8sNgrR6XHIbDPTs6hdVLhRmsBR_nK-WQVer7zY6gPSmm1pIQT5soqR4Dqyo3prIvXmk6EpMN), [2](https://www.google.com/goto?url=CAESYQHrOzAV9bhPXwiPFAkk0WrtjwF69osLlHBUWNzzI69S1WtwZTqUCsUsFHJRmbHqvOTqawUMo9x8Q-JEjDS2uEGe-nwFB8K7bhK_9tMlb9RUSeY2GDejKtZrzQPcaK1YQ8A)] Would you like to know more about how Ithaca connects with newer models like Aeneas or how historians use the Predicting the Past platform? Google DeepMind
Nature third_party
Restoring and attributing ancient texts using deep neural networks | Nature
Model: Ithaca, a deep neural network trained on ancient Greek inscriptions (7th century BC to 5th century AD) from the I.PHI dataset (78,608 texts), simultaneou...
as pastedRestoring and attributing ancient texts using deep neural networks | Nature Model: Ithaca, a deep neural network trained on ancient Greek inscriptions (7th century BC to 5th century AD) from the I.PHI dataset (78,608 texts), simultaneou...

↻ Re-run permalink