Frequently asked questions.

You have probably heard of the Library of Alexandria, where knowledge was gathered onto papyrus scrolls.^{[^1][^3]} Less familiar is Pergamon, Alexandria’s great rival, which became traditionally associated with parchment as an alternative writing substrate.^{[^1][^2]} The lesson we take from that history is simple: knowledge becomes less fragile when it is no longer bound to a single medium. Project Pergamon takes its name from that principle. We are building an institution around the problem of preserving knowledge at historical timescales.

The Archimedes Series and ARCHI archival standard are named for the Archimedes Palimpsest. This artefact supports our thesis that preservation efforts must extend to material for which the means of interpretion have not yet been invented. You can learn more about how new technology decoded an ancient message through the Walters Art Museum's website on the project

We use “public” because “open source” has become a contested term in AI, especially in cases where model weights are publicly released but training data or other components remain undisclosed.^{[^4][^6][^7]} “Publicly released,” by contrast, is a verifiable fact: the rights-holder made the model available for public download on documented terms.

Pergamon is designed to record those distinctions rather than blur them. Our metadata profile distinguishes a component that was withheld — for example, when a rights-holder chose not to release training data — from one that is merely unknown, where public evidence was insufficient to establish provenance with confidence. That accession record gives future researchers a clearer account of institutional decision-making, including the present-day drift toward decoupling public weights from open training.

Online availability is not the same thing as durable survival.^{[^8][^9]} Publicly released models can be delisted, neglected, censored, fragmented across platforms, or rendered unusable when the surrounding technical context disappears. A live download link is a convenience left to someone else’s discretion, without any preservation guarantee.

Physical, independently held copies add custody, independence, and durability that networked distribution alone cannot promise.

The OAIS Reference Model (ISO 14721) is part of the foundation we build on, and Pergamon is designed to conform to it wherever it applies.^{[^10][^11]} But OAIS is a general preservation reference model, not an AI-specific archival profile. The information needed to interpret a generative AI model is not exhausted by its source code, nor by any other single technical artifact.

On this frontier, the work of future historians may resemble philology more than direct reference: inferring meaning from a bundle of materials without full access to the people and institutions that produced them. Pergamon’s metadata format is expressed as a profile of the widely used BagIt packaging standard (RFC 8493), and it maps onto OAIS concepts while addressing the distinctive preservation problem posed by generative AI.^{[^12]}

Because most existing software archives solve an adjacent problem, not the full preservation problem for public AI. Code archives are essential for preserving repositories, version history, and source files. But a generative model is not reducible to source code alone. Its historically meaningful record includes weights, architecture, tokenizer, provenance, documentation, and the terms under which different components were released or withheld.

Project Pergamon is designed to complement those archives, not replace them. The problem is not just storing files. It is preserving the conditions under which historically significant public AI systems can still be understood, verified, and reconstructed in the future.^{[^10][^12][^13]}

LOCKSS — “Lots of Copies Keep Stuff Safe” — is one of our aspirational models.^{[^13][^14]} Its core insight is social as much as technical: many independently administered copies, continually checking one another, are stronger than any single custodian. That is a standard Pergamon’s emerging governance model will need to meet.

But LOCKSS was built to preserve web-published, renderable records gathered by crawl. A frontier AI model is much larger, not ordinarily web-harvestable, and better understood as a technical system to be interpreted rather than a page to be rendered. Pergamon extends the LOCKSS logic to that new kind of artifact. Our packaging approach is designed to remain interoperable with adjacent archival systems rather than locked inside a single preservation silo.^{[^12][^13][^15]}

The Linear Tape-Open (LTO) format remains one of the proven, scalable, and low-energy choices used in working archives today.^{[^16]} It is not a magical forever medium, and we do not present it that way. Real preservation depends a disciplined archival format and workflow: the institutional challenge that no one medium solves once and for all. Our initial preservation horizon is measured in planned migration cycles, while our logo, a snake who sheds his old skin for the new, speaks to the growth of knowledge across materials. Over longer timescales, the archive may well move onto other media, including ceramic- or silica-based substrates, if and when they become viable.^{[^17][^18]}

We do not guarantee that the future computing environment will run today’s models exactly as they run now. What Pergamon preserves is the record future custodians would need to reconstruct, emulate, or meaningfully interpret them.

That includes weights, architecture, and the surrounding documentation needed to recover a model’s technical logic as honestly as possible. Future performance may in fact depend on a world we do not control; That is why we strive to improve the odds that reconstruction and emulation remain possible.

This concern is real. Publicly released models can be used against the public's interests, and Pergamon does not dismiss that risk. But preservation does not create public access where none existed before: the models we preserve are already publicly released.

Pergamon also does not position itself as a universal arbiter of which models are acceptable everywhere. Instead, we are designing an archival framework in which different jurisdictions and institutions can govern access to weights, training data, execution requirements, or other components differently. Just as important, those restrictions become part of the archival record itself.

Models with pubicly available training data are linked to that corpus by reference: a practical and legal decision. Practically, an LLM with final weights measured in gigabytes may have been trained on a corpus of tens of terabytes. And embedding every corpus inside every model's record would be redundant, since a single corpus is often shared across many models. The legal concern runs deeper: A training corpus is usually not the model-maker's to give away. Pre-existing works that belong to others and are held elsewhere (web pages, books, code, images) fall outside the responsible mandate of a public AI archive. Where a corpus is openly liscenced and seperately stewarded, we point to it as its own preserved artifact. Our records remain comapct, lawful, and honest: we never pretend to hold what is better served elsewhere.