The Ptolemy Project

A proposal to find, image, translate and grade what survives of the human record, and to release all of it into the public domain.

1. The problem

The record of the ancient world is incomplete, and the part that survives is largely unread. The failures differ by category.

1.1 Texts

Material that exists has never been examined. Around half a million papyrus fragments were excavated at a single Egyptian site from 1896; roughly five thousand are published. Some 500,000 cuneiform tablets have been excavated and approximately nine in ten have never been published. About 1,800 carbonised scrolls from one villa at Herculaneum were unreadable by any method until 2023.

Where digitisation has happened the systems do not connect. Each programme maintains its own identifiers, metadata and viewer, holdings in one are not discoverable from another, and most publish a viewer rather than a data interface. Archival-resolution files are commonly withheld even where the images are free to view, which places a large body of already-funded capture beyond reuse.

1.2 Artefacts

No inventory exists. There is no global catalogue of museum holdings, the number of objects held worldwide is not a known quantity, and the proportion in storage rather than on display is not recorded centrally.

Context is being lost faster than it is recorded. Findspot data for much excavated material survives only in unpublished notebooks held by individuals or departments. Objects removed without an excavation record cannot be stratigraphically dated or located, and that information cannot be reconstructed afterwards.

Licensing is inconsistent between holders of the same material. Some major collections release object data and images without restriction; others apply non-commercial conditions to reproductions of objects thousands of years old; others publish a viewer only.

1.3 Constructions

Dating is circular in a significant number of cases: structures are dated by associated finds, and those finds are dated by the structure. Attributions frequently rest on style or proximity rather than on any contemporary inscription.

Documentation is uneven and in many cases postdates damage. Where structures have been destroyed, the surviving record is often photographs taken for other purposes. Reconstruction drawings have in several instances become the standard image of a building and been reproduced as though attested.

1.4 Translation

Two populations are affected. For several ancient languages — Akkadian, Sumerian, Hittite, Coptic, Syriac among them — the number of people able to read the material is in some cases a few dozen worldwide, which is the rate-limiting factor on publishing the backlog regardless of how much is imaged.

The larger volume is modern. Two centuries of German classical scholarship, French Egyptology, Italian papyrology, Russian Central Asian archaeology, and Chinese and Japanese work on East Asia remain substantially untranslated, and the principal Chinese and Russian scholarly databases are subscription-bound and outside most Western institutional agreements.

1.5 Cities

Settlement sites are stratified: successive occupations sit vertically in the same footprint, and the vertical position is itself dating evidence. Troy carries nine principal levels. San Clemente in Rome stands over a fourth-century basilica, over a Mithraeum, over a first-century house.

That sequence is published as separate excavation reports, one per campaign and often one per layer. No standard record exists for a footprint through time, so the layers of a single site cannot be read as a sequence without assembling the reports by hand.

Reuse is unrecorded for the same reason. The Theatre of Marcellus became apartments; the Pantheon survived by conversion; Fort Qaitbay is built from the material of the Pharos. Reused blocks in later walls record what was demolished and when, and are noted in site reports rather than held as a field.

1.6 Territory

A place holds three identities that change at different rates. The physical site does not move. The name changes with whoever holds it: Byzantium, Constantinople, Istanbul; Akko, Ptolemais, Acre. The polity claiming it changes faster than either.

Records commonly carry one current name and one attribution, so a settlement appears as several unrelated entries across periods, or as one entry that silently drops its earlier names.

Boundaries are recorded as lines when the evidence supports a zone. Ancient frontiers were in many cases a set of controlled towns and tribute relationships rather than a demarcated edge, and rendering them as hard borders states more than the sources hold.

1.7 Excavations and events

Excavation archives are not held by any digitisation programme. Notebooks, plans, section drawings and photographs sit with universities, museums and individual directors, and are frequently the only surviving record of a findspot.

No count exists of campaigns against publications, so the size of the unpublished excavation backlog is asserted rather than measured.

Located events are recorded inconsistently. Battle sites are in most cases located by inference from a place-name in a text, and the inference is not distinguished from an identification. Where battlefield archaeology has been done it has in several cases revised the written account.

1.8 Why the work is not being done

None of the above is caused by restriction of access. Institutions publish what they have funded; the constraint is that reading, cataloguing and translating the backlog is unfunded and is not attached to any career path.

The scope is global by necessity rather than ambition. The record of any single civilisation is distributed across many jurisdictions: the Dunhuang manuscripts are held in five countries, Egyptian material in every major European collection, cuneiform in Baghdad, London, Berlin, Philadelphia and Istanbul. A national programme cannot assemble a national record.

What has changed is that capture and reading are now tractable. Flat imaging, reflectance transformation imaging for incised surfaces, photogrammetry for objects and structures, and tomographic imaging where an object cannot be opened, are all within reach of ordinary equipment and trained non-specialists. Machine translation and restoration are producing published results in exactly the low-resource languages where the reader population is smallest.

Applied to a unified corpus rather than to isolated collections, those methods support a fuller reconstruction than any single archive permits: the same object, text or structure seen against everything else recorded from the same place and period. That output is the subject of section 18, and is released without restriction.

2. Texts

Four operations: locate, acquire where at risk, capture at archival resolution, publish. Unit costs are documented. The Endangered Archives Programme has funded over five hundred projects in more than ninety countries and digitised over sixteen million images since 2004, funded by one private foundation.

Its terms address the sovereignty constraint: material does not leave its country of origin, a primary copy is deposited with a local repository, and projects include local training.

The programme does not distribute high-resolution files, referring those requests to the original holders. Sixteen million images are captured and the archival-resolution files are not publicly available. Releasing them is a licensing decision on completed work, at no additional capture cost.

Acquisition is uncovered. The programme funds digitisation in situ and not purchase, and no equivalent facility exists for at-risk material in private hands. The Timbuktu manuscripts were evacuated in 2012 and 2013 without institutional funding.

3. Artefacts

An object without a recorded findspot can be described but not stratigraphically dated, located, or associated with other material. Removal without excavation record destroys the contextual data on which evidential value depends.

Reproduction income is a material constraint on adoption. Institutions that fund conservation from image licensing would lose that revenue under the licensing terms in section 15, and the proposal does not currently replace it.

3.1 The register

Each object requires a persistent identifier and a tamper-evident record of its scan and its stated provenance. The holding mechanism is unresolved and is listed in section 21.

A record states that a claim was made; it does not establish that the claim is accurate. The claim and the evidence for it are stored as separate fields and graded separately, on the same scale used for texts.

4. Constructions

Fields: structure, location, date, evidence for the date, commissioning authority, and survival state — standing, ruined, relocated, lost, or attested only in description.

Dating is weakest in this leg. Structures are commonly dated by associated finds, and those finds are in some cases dated by the structure. No contemporary inscription records the commissioning of the Sphinx; attribution to Khafre rests on proximity.

The register records derivation as a field. The Pharos does not survive; its material is on the sea floor and in Fort Qaitbay. Thiersch’s 1909 reconstruction became the standard image and subsequent buildings were modelled on it, propagating a disputed reconstruction as an attested form.

The works table on Egypt before the Ptolemies is the working prototype for this schema.

5. Cities

A fifth record type, distinct from the others in that it describes a place through time rather than an object or a structure.

Fields: footprint, layer, elevation and stratum, date range, what stood there, the evidence for it, what replaced it, and what was reused. The unit is the location, and the entries are read downward.

Elevation is not metadata here. In a stratified site the vertical position of a layer is part of the dating evidence, and a record that carries only a date discards it.

Continuity is a separate field from sequence. A layer may be demolished, built over, converted, or quarried for material, and those are different relationships. The Mithraeum under San Clemente was not destroyed; it was filled and built upon, which is why it survives.

Layers within one footprint are not equally attested and are graded individually. A dated foundation inscription, an excavated wall in a recorded stratum, a structure named by a contemporary writer, and a street line inferred from later property boundaries are four different grades. A rendered model that draws all four at the same solidity states more than the record supports.

The Esquiline summit in Rome is a worked example of the record type. The sequence on one footprint runs from sacred groves through a temple of 375 BC, a street of houses, a fourth-century hall, and a basilica of 432, with a first-century villa six metres below the present floor.

The evidence for each layer differs in kind: some excavated, some inscribed, most literary. A Hill of Mothers sets out the sequence, and Reading the Arch the surface decoration, including the restoration record that determines what any detail on it can be asked to support.

Existing work is partial and single-period. Reconstruction projects model one moment; lidar survey has recovered settlement plans under canopy without excavation; gazetteers hold stable identifiers for ancient places. None of them carries a footprint through its full sequence with the evidence attached to each layer.

6. Territory

A sixth record type. The unit is the place, and the entries are dated attributes of it rather than facts about it.

Fields: place, coordinates, name with date range, polity with date range, boundary state, and the evidence for each. A name is a dated field, not an attribute, so a settlement carries every name it has held with the period each was in use.

Boundary state is graded rather than drawn: controlled, tributary, claimed, disputed, unknown. The same territory can be tributary to one power and claimed by another in the same year, and both entries stand.

The Levant demonstrates the requirement. Coele-Syria is Ptolemaic, Seleucid after 200 BC, Hasmonean, Herodian, then Syria Palaestina after AD 135 — five names for overlapping but not identical areas, with the settlements inside them changing names on a separate schedule. Thirteen divisions of that one territory are set out on Who Ruled the Land.

Existing gazetteers hold ancient places with stable identifiers and some hold name variants. None carries polity and boundary through time with the evidence attached to each state.

7. Excavations

A record type in its own right, and the provenance layer under three of the others. Every artefact, city layer and structure entered the record through an excavation, a survey or a chance find, and the conditions of that recovery determine what the material can support.

Fields: site, coordinates, campaign dates, director and institution, method, what was recovered, where the archive is held, what was published, and what was recorded but not published.

Method is graded because it is not constant. Material recovered at Nineveh in the 1850s carries context that the methods of the time destroyed; material from a modern stratified excavation does not. An object record that cites its excavation inherits that condition, which is why the two are linked rather than merged.

The unpublished backlog becomes measurable rather than asserted. Counting campaigns against publications states the size of the gap directly, which no current source does.

The Esquiline case is representative. A campaign of 1966–71 recovered a building of 37 by 30 metres beneath one basilica and identified its owners from epigraphy; most of the footprint has never been dug, a catalogue of 14,000 roof tiles from the site is unpublished, and the excavation report is not available online. Everything in general circulation about the site is at summary level.

Site archives are the material at risk. Notebooks, plans, section drawings and photographs are held by universities, museums and individual directors, are frequently the only record of a findspot, and are not covered by any of the digitisation programmes listed in sections 11 to 14.

8. Battles and events

Located events, distinct from territory in that they occupy a moment rather than a period, and from constructions in that most leave nothing standing.

Attestation and location are graded separately. A battle can be securely attested and insecurely located: Qarqar in 853 BC is named in a contemporary Assyrian inscription and the site is not securely identified. The two claims are recorded as two entries.

Where material evidence exists it frequently qualifies the written account. Battlefield archaeology at Kalkriese and Towton produced distributions of weapons and casualties that revised the received narrative in both cases. Most ancient battle sites have no such evidence and are located by inference from a place-name in a source.

Fields: event, date, participants, outcome as claimed by each source, location with its own grade, and material evidence where any exists.

Sources conflicting is recorded rather than resolved. Both sides claimed Kadesh in 1274 BC in their own inscriptions, and the register carries both claims with their channels named.

9. Translation

Two distinct problems.

Ancient languages — Akkadian, Sumerian, Egyptian, Hittite, Coptic, Syriac, Classical Chinese — where the reader population is in some cases a few dozen worldwide. Machine translation is producing published results in this category.

Modern scholarly languages, which is the larger volume. Two centuries of German classical scholarship, French Egyptology, Italian papyrology, Russian Central Asian archaeology, and Chinese and Japanese work on East Asia remain substantially untranslated.

Unit costs here are the lowest of the eight legs: most modern-language material is already digitised, so the requirement is compute rather than capture.

Source text is displayed with every translation. A machine translation distributed without its source cannot be checked against the original.

10. The top-level system

Existing digitisation programmes are not interoperable at catalogue level. Each maintains its own identifiers, metadata schema and viewer, and holdings in one are not discoverable from another.

The proposed system sits above them. It harvests from any programme permitting it, holds local copies, and ingests material no programme covers: private scans, newly captured imaging, translations and grades. Physical custody is unaffected.

An index resolves to external sources; if a source goes offline the material is lost to it. Held copies are not subject to that. Translations and grades have no external holder in any case.

The International Dunhuang Project demonstrates cross-jurisdiction harvesting under an unresolved ownership dispute. Some 50,000 manuscripts from one cave were divided between five countries by the collectors who removed them. The images were unified on a common standard; custody of the objects was not addressed.

Holding copies introduces two failure modes. Concentration is addressed by mirroring across continents under separate operators. Selection bias is addressed by an accept-all rule: material is not refused, it is graded, and the grade is published.

11. Where the data is — texts

Ninety-six sources are listed across sections 11 to 14, by leg. The lists are a starting point for the harvest, not a specification of it, and are not exhaustive.

Coverage is uneven by language and by state. Western European national libraries publish IIIF and in several cases bulk access; the Russian, Chinese and Indian holdings are portal-only or partial, which places a large body of material outside reach of any automated harvest.

12. Where the data is — objects

Wikimedia Commons and Wikidata are the only sources on these lists that are already aggregated, already licensed for reuse, and already exposed for bulk harvest. Wikidata additionally holds cross-collection identifiers, which makes it usable as a concordance between holders that share no identifier scheme.

Crowd-sourced imagery is a separate case. Photographs taken by visitors and in the field frequently show objects that the holding institution has not published. They arrive without accession number, findspot or provenance, and on general image platforms they arrive without a reliable source for the photograph itself. Material of this kind is admitted only where an identifier can be attached — commonly a visible object label — and is graded at the lowest level in all cases.

The specialist corpora are in several cases better structured than the museum portals they draw on. Arachne maps to CIDOC-CRM and exposes open interfaces; the Beazley database publishes bulk download. Both are single-domain and neither aggregates beyond its field.

Licensing is the variable. The Metropolitan Museum, the Smithsonian and the Rijksmuseum release object data and images without restriction; the British Museum applies a non-commercial share-alike condition; several major collections publish a viewer only. The same object type is therefore harvestable from one holder and not another.

Auction catalogues are in many cases the only public record an object receives, and they are withdrawn when the lot closes. Capturing them at point of sale is the only opportunity.

13. Where the data is — structures, cities and territory

Excavation archives are the largest gap in this leg and the least addressed. Site notebooks, plans and unpublished reports hold the findspot data on which dating depends, and are held by universities and individual directors rather than by any repository.

14. Where the data is — scholarship and standards

The Russian and Chinese scholarly databases are subscription-bound and outside the reach of Western institutional agreements, which is the practical form the translation problem in section 9 takes.

Holdings figures, licensing terms and interface availability change. Each entry is stated as published by its holder and is not independently audited here.

15. Public domain from capture

The underlying works are out of copyright by age. Where a right is asserted it is asserted in the reproduction — the photograph, the scan or the model — not in the work.

Precedent: the Staatliche Museen zu Berlin withheld its scan of the Nefertiti bust, and following a third-party release in 2016 published its own data under a non-commercial licence. The object dates to approximately 1345 BC.

Reproductions of public-domain works — photographs, scans, 3D models, transcriptions, translations and derived data — are released into the public domain. No non-commercial clause, no attribution requirement, no database rights asserted.

Each exception creates an enforcement point for a future holder. A non-commercial restriction excludes textbook, broadcast and model-training use, which account for the majority of downstream use.

16. The technology

Capture methods are matched to material, not to budget. Most of the volume is addressed by three low-cost methods; the expensive ones apply to a small number of high-value cases.

The three low-cost methods cover the bulk of the backlog. Flat imaging, reflectance transformation imaging and photogrammetry are within reach of ordinary equipment and trained non-specialists, which is what makes the volume tractable at all.

Storage is a material cost and is not estimated here. Archival-resolution capture produces large files, and a corpus at the scale described in section 1 implies petabyte-order storage with mirrored copies. No figure is offered because it depends on capture standard, retention policy and mirror count, none of which are settled.

Compute divides into three loads: text recognition, which is cheap per item; translation, which is cheap per item at volume; and model training, which is a periodic fixed cost. Translation and restoration are compute rather than capture, which is why section 9 is the least expensive leg.

17. Grading

Recovery does not establish accuracy. Ithaca reports 62 per cent top-one accuracy unaided and approximately 72 per cent in combination with a historian. A restoration at that confidence is recorded as a restoration, with the figure stated.

Grading separates two claims that most registers combine. A contemporary inscription in a well-attested language with bilingual keys is not the same evidence as a contemporary inscription in Sumerian, which has no known relatives and was deciphered indirectly through Akkadian, or in Etruscan, which can be read aloud and largely not understood.

Object and reading carry separate grades. An entry may hold three for the object and one for the reading; the combination is common and is not a contradiction. The same scale applies to artefact provenance: excavated with a published findspot, collection history without excavation context, or an owner’s assertion.

Grading rules are published and versioned. A proposed change carries evidence, on the same standard as a claim.

18. The output

Every record in the corpus carries a place and a date. An object has a findspot; a text has a place of composition and a place of survival, which are frequently different; a structure has coordinates; a city layer has both; a polity has a boundary and a period.

The primary interface is therefore a map with a time control rather than a set of documents. Setting a date renders the state of a region at that date, in layers that can be shown independently: territory and boundary, settlements under the names then in use, standing structures, city layers at that stratum, and the objects and texts recorded from those places.

Each element carries its own grade, and the rendering reflects it. An excavated boundary marker and an inferred frontier are not drawn alike; a recorded findspot and an attributed one are not drawn alike.

The join that does not currently exist is object to findspot. Collection records give the holding institution; where an object was found is often held in an excavation report or is absent. Linking objects to where they were found rather than where they now sit is what makes a collection usable as evidence of a place.

Two consequences follow. Dispersal becomes visible: selecting a site returns everything excavated there regardless of where it now sits, which is the Dunhuang problem rendered as a map rather than a dispute. And absence becomes visible: a region with no markers in a period is not a quiet period but an unexcavated or unpublished one, which a list of holdings does not show.

Generated pages are produced from the same records: the register, the timeline, the links between entries, and the grade on each claim. Generated text is labelled as generated and versioned. The corpus is the evidence; the page and the map are derived readings of it.

19. What it would cost

No total is offered. Unit costs below are published figures.

• Archive rescue, digitised in situ: up to £60,000 per project over 24 months, material remaining in the country of origin.
• Endangered Archives Programme, first ten years: just over £6 million, single private funder, five hundred projects.
• Capture equipment: photogrammetry requires a phone and a turntable; reflectance transformation imaging requires a camera, a lamp and a reflective sphere.
• Reading prize, Vesuvius model: privately funded, results within twelve months of launch.
• Release of archival files already captured: a licensing decision, no capture cost.

Set against approximately eight million objects in a single museum and some 500,000 excavated cuneiform tablets, these are low unit costs. The binding constraint is funded labour, not capital or technology.

20. Income

The output is public domain and is not sold. Revenue is therefore raised against a free service rather than from it, on the model of publicly funded infrastructure.

The condition that holds across every stream is that no contributor obtains access another party does not have. A funder receives the same corpus as any other user, which is what allows states with unresolved disputes to contribute without contributing to each other.

Two streams do work the others cannot. Reproduction-income replacement addresses the obstacle identified in section 3: institutions that fund conservation from image licensing have a material reason to refuse open terms, and an offsetting endowment removes it. Contract imaging reaches holders with material and no capacity, and is the mechanism by which private and monastic collections enter the corpus at all.

The project is a cost centre with partial offsets. It is not a business and does not become self-funding at any scale described here.

Funding is accepted from any source on the licensing terms in section 15: all output public domain from capture. The Vesuvius Challenge was privately funded by two individuals and produced results within twelve months.

Ancient-language restoration is a low-resource sequence problem with no copyright encumbrance on the training data, and is therefore a viable research target. Ithaca was built in a commercial laboratory and published in Nature.

No exclusive right is created or transferred. A funder obtains the same access as any other party. Contributions are therefore separable, which permits participation by states with unresolved disputes between them.

21. What is not settled

Replacement of museum reproduction income; selection among the register options in section 3.1, or use of two in parallel; governance of the grading rules; siting and operation of the mirrors; and the governing entity for the system as a whole, which is not specified in this draft.

22. Sources

Reference codes read source.work.passage and resolve on the Master Source Register, which carries every source used across this series.

1. Excavation and battlefield evidence — the Kalkriese and Towton battlefield surveys are the standard cases in which material distribution revised a written account. The Qarqar site is not securely identified; the battle is attested on the Kurkh Monolith, BM 118884. Kadesh is claimed as a victory in both the Egyptian and Hittite records. Supports sections 1.7 and 8.
2. Places and territory — Pleiades supplies stable identifiers and name variants for ancient places; Pelagios and its map interfaces link datasets by place. Neither carries polity or boundary state through time. The thirteen divisions of the Levant referenced in section 6 are set out with their sources on Who Ruled the Land. Supports sections 1.6, 6 and 18.
3. Cities and stratified sites — the San Clemente sequence (fourth-century basilica over a Mithraeum over a first-century house) and the Troy level count are the standard published site sequences. The Forma Urbis Romae survives in fragments and its reconstruction is incomplete. Lidar recovery of settlement plans under canopy is reported by the surveying projects and is not independently checked here. Supports sections 1.5, 5 and 6.
4. Wikimedia Commons and Wikidata — freely licensed media and structured identifiers, both exposed for bulk harvest, Wikidata through a public SPARQL endpoint. Licensing on Commons is per-file and varies; the platform aggregates institutional donations alongside individual uploads, and file-level rights must be checked rather than assumed. Supports sections 9 and 11.
5. Object and inscription corpora — iDAI.objects (Arachne), German Archaeological Institute and the University of Cologne: approximately 1.7 million registered images, mapped to CIDOC-CRM. Beazley Archive Pottery Database, Classical Art Research Centre, Oxford: over 130,000 records and 250,000 images, with bulk download. Corpus Inscriptionum Latinarum volumes published before 1940 are open through Arachne. Figures are as published by each project and are not independently audited here. Supports sections 9 and 11.
6. Endangered Archives Programme, British Library, funded by Arcadia — project counts, country coverage, image totals, the £60,000 project cap and the first-decade cost are the programme’s own published figures. The provisions on material not leaving the country of origin and on local repository copies are from its published terms. Supports the callouts, §1, §4 and §10.
7. International Dunhuang Project — the dispersal across five countries and the digital reunification. Supports §2.
8. Vesuvius Challenge — the 2023 results on the Herculaneum scrolls, and the funding structure. This memo has not verified the readings themselves, only that they were awarded and published. Supports §1, §10 and §11.
9. Ithaca — Assael, Sommerschield et al., Nature 603 (March 2022). The 62 and 72 per cent figures are the paper’s own, on its own test set. Supports the callout, §1 and §8.
10. Akkadian translation — Gutherz et al., 2023. Supports §1 and §7.
11. Nefertiti scan — the 2016 release by Nora Al-Badri and Jan Nikolai Nelles, and the Staatliche Museen zu Berlin’s subsequent publication of its own data. The circumstances of the artists’ scan are disputed and are not established here. Supports §3.
12. Public-domain reproductions — Article 14 of the EU Copyright Directive (2019/790) provides that reproductions of public-domain visual works do not attract new copyright. Its transposition varies by member state and it does not bind the United Kingdom or the United States. Supports §3.
13. Lost work: Manetho, Aegyptiaca — commissioned under Ptolemaic patronage, lost as a book, surviving in quotation. The dynastic numbering still in universal use is his. Named here because the precedent for state-funded compilation is the same dynasty this project takes its name from.
14. Sphinx attribution — no contemporary inscription records who commissioned it. The Dream Stele between its paws was set up by Thutmose IV roughly a thousand years later. Supports §6.
15. The Pharos reconstruction — Hermann Thiersch’s 1909 drawing, and the coin evidence that disputes it, are set out in The Ptolemies. Supports §6.
16. Verification register. Carried as reported and not settled: the accuracy of the Vesuvius readings; the circumstances of the Nefertiti scan; the transposition of Article 14 outside the EU; and whether the machine-translation results cited generalise beyond their published test sets. No cost is offered for the project as a whole, and none should be inferred from the unit costs in §10.