The immediacy and reach of the internet are reshaping scholarly communication. Alongside traditional venues such as professional meetings, conferences, colloquia or panel discussions, the internet can help create new communities of interest, and offers existing communitities new forms of communication including email lists, discussion groups and blogs.

Digital information technologies have equally dramatic implications for the permanent published record intended to communicate to both present and future scholars. My intention is to define as clearly as possible the assumptions and reasoning behind technical work at the Center for Hellenic Studies in support of its electronic publication initiative.

It is vital that we define explicitly what we mean by "scholarly publication," independent of any possible media for publishing, and what we mean by a "library of scholarly publications," independent of any possible media. Otherwise we risk falling into the trap of reshaping our work to the forms of a particular medium, rather than asking how different media enable us to pursue our goals. This is a serious risk for a discipline like Classics, because decisions we make today about the form of our permanent scholarly record will have consequences reaching far into the future.

Normal English usage complicates the job. The word "publication" can refer to the process of publishing something (e.g., "the future of scholarly publication"), to a single act of publishing ("publication is expected by January of next year"), to a product of that act ("The library has Austin and Bastianini's publication of the Posidippus papyrus"), or to the content of that product ("Petrie's publication first developed the idea of seriation"). For the sake of clarity, I will use the term "publication" only to refer to the process of publishing; I will refer to specific products of that process as "published works," and where necessary further qualify that with expressions like "published works in print media" or "digital published works." I will use the phrase "scholarly work" to refer to the content of a published work without implying that the work is published in any specific form, or even that it is published at all.

I begin this introductory section with an example illustrating why it is necessary to separate definitions of basic concepts from any reference to specific media. The following sections then set out a working definition of "scholarly publication" consideration of what that definition implies for the design of scholarly published works in digital form a working definition of "scholarly libraries" consideration of what that definition implies for the design of a digital scholarly library

Finally, a postscript briefly points to analogies between the production of hardware and software on the one hand and the production of scholarly published works on the other.

In considering such a broad topic as "digital publication for digital libraries," I hope that classical scholarship's unusual experience can provide a useful perspective. Classical scholars have seen the practice of recording primary scholarly materials migrate from papyrus rolls to codex manuscripts to printed materials; that experience should help us now as we migrate from print materials to collections of bits capable of being represented in many physical media (magnetic disk, RAM, and laser discs such as CD ROMs, for example).

In the transition to printed materials, Classics' track record has arguably been superior to most disciplines, as an important example will illustrate. Consider the consequences of including page numbers in printed works. This would have been pointless in hand-copied codices (since pagination might vary with each copy), and page numbers are not regularly found in fifteenth century incunabula, which tend to perpetuate the physical form of hand-copied texts. In the course of the sixteenth century, however, printers and readers came to recognize the value of labelling pages that would be identical across every copy of a given edition, and page numbers came to be used as a standard navigational device for working with printed books through new inventions like indices and tables of contents.

The temptation that most authors and readers succumbed to was to use this convenient navigational device as a reference system. Page numbers became the normal way to cite a passage in a printed text. This remains today a severely limiting convention, because it ties the logical act of reference—as I will argue below, perhaps the most fundamental scholarly activity&emdash;to a specific physical edition or version of a work. How frustrating this limitation is will be familiar to any teacher who has had to use a new edition of a textbook with nearly identical content, and found last year's notes out of sync with the current edition.

For the most part, Renaissance classicists and biblical scholars resisted this temptation. They recognized that their texts would be read in multiple versions&emdash;different editions and translations&emdash;that could all use a single canonical citation system. Phrased differently, they devised a logical reference scheme independent of any specific version of a text, and, because of their long experience with multiple versions of their central texts, avoided confusing the physical form of a new printed book with its logical structure. So today, we can still cite "Herodotus 1.5" and be confident that it refers to the same passage (book 1, chapter 5) in any version of Herodotus we prefer to read.

There is an audible echo of this decision in a discussion that has developed in the decade since the introduction of the World Wide Web. "Uniform Resource Locators," or URLs, are addresses on the internet providing a navigational scheme, and it is tempting to use them as a form of reference. Like page numbers, however, they are valid for only one version of a document, and in the accelerated time of the internet's world, the limitations of this navigation scheme as a form of citation are already familiar to all users of the WWW who run into dead-end links ("404 limbo"). In many cases, the desired document may still exist, but at a new address.

The Online Computer Library Center developed "Persistent URLs" PURLs, see http://purl.oclc.org to implement the notion of Uniform Resource Names, or URNs, articulated by the World Wide Web consortium (W3C). URNs are intended to describe logical names rather than specific locations on the internet. The confusion of logical name and physical address has become pervasive enough that the W3C maintains a page clarifying how terms for logical and physical reference are used in different senses in different parts of the W3C's documentation. http://www.w3.org/TR/uri-clarification/

In at least one respect, however, the discussion of how to refer to resources on the WWW has not yet recognized a significant implication of classicists' traditional practice of citation by logical reference: namely that the idea of a logical reference scheme follows naturally from a notion of a logical text. A canonical reference scheme citing Herodotus by book and chapter is a logical or notional reference in the sense that the organizational scheme is independent of any specific version of the work. By the same token, reference to a text called is itself purely logical in the sense that it refers to a notional text independent of any specific version. While URLs and URNs, on the other hand, both assume that documents may move their physical address, their concern is really with a single document. URLs and URNs have no concept of a notional document that would allow readers to treat a Greek text and English translation of Herodotus as two versions of a single logical text. Some applications using URLs approximate this kind of behavior. If a Web server and browser both support content negotiation, for example, the server can recognize a browser's preferences (for example, for one language over another) and serve one version of a document rather than another. This behavior is application specific, however: there is nothing in the URL protocol supporting it. For a classicist, there would have been no incentive to develop a regular, canonical scheme of book/chapter/section citation for one document called by Herodotus if references in that scheme only applied to a single edition or translation of the work. The power of the logical reference scheme lies largely in its application to an abstract notion of a work encompassing different editions and translations. That is why the first project undertaken by the Center of Hellenic Studies' technical work group defines a protocol supporting logical citation schemes for logical texts.

This extended story about how to cite a work should serve first as a warning. We still live in a period of experimentation in digital media comparable to the experimentation we see in fifteenth-century incunabula, when early printed works tended to mimic the physical form of hand-copied manuscripts. Our conventions are not yet settled, and the very real risk is that we will uncritically allow the new media to shape those conventions, just as most disciplines uncritically came to use page numbers as an easy substitute for a logical reference scheme. At the same time, the story is a challenge: classicists can draw on the long experience of their discipline to inform their thinking about today's digital incunabula.

We start with a functional definition: the distinctive role of publication in the scholarly world is to serve as the permanent record of reference for scholarly work. By this I mean that in contrast to other vital but more ephemeral forms of communication, published works constitute the sole record of scholarly work that other scholars, present and future, can reliably refer to and cite.

We habitually imagine this function in terms of published works in print: volumes of ink on paper, physically produced by publishers who market the volumes to scholars and libraries. We may even assume that features of published works in print that directly contradict the aim of a permanent record of reference are not artifacts of a particular system of technologies and institutions, but are intrinsic to scholarly publication. It is easy to assume, for example, that difficult or technical scholarly works will be accessible only in a few major research libraries&emdash;because after all it's not feasible to print more than a very small run of such material. Humanists whose work depends on published photographs and drawings will "know" that those published images will be limited in number because of the expenses of "publishing" them&emdash;although there is no scholarly rationale for this limit, which is antithetical to the goals of publication. We sometimes accept these limitations on the scholarly record as inevitable, but before we can hope to reimagine the permanent record of reference in terms of a networked world of digital information, we need to set aside assumptions that follow from our well established system of print publication. If the essential function of scholarly publication is to serve as a permanent record of reference, we need to identify what features of form are essential to scholarly publication in any medium.

As a first essential feature, I would argue that the form of a scholarly published work must be identically replicable. This follows from its function as a work of reference: scholars must be able to find identical content in identical copies of a publication. Of course, citation of archival material such as manuscripts inevitably leads any scholar to the same content, but the requirement that a scholar consult a single physical copy is perhaps the most salient distinction between archival material and published works.

Replication also directly serves as the best guarantor of the permanence of a work. If there is any lesson that is beyond dispute in the thousands of years of transmission of classical texts, it is that no single Library of Alexandria (or Library of Congress) can ensure the survival of a work: a fortunate copy anywhere can perpetuate a work beyond the lifespan of any single collection, however imposing. The more widely spread copies of a work are, the greater its chances of survival. If we hope for the preservation of classical Greek literature as far into the future as its creation is distant from us in the past, we should promote its replication, not just its archiving in a few massive repositories.

A second essential feature of scholarly publication is the alienation of the work from its author. Its form is fixed and outside the further control of the author once it is published: like evidence in a court case, a published work can no longer be withdrawn or altered once it has been accepted into the record. An author can of course change views or correct errors&emdash;but such changes can only enter the scholarly record through a further published work (whether a revised edition, or an altogether separate work). Our first two features together imply that a scholarly publihsed work must have a form that can be irrevocably and identically replicated.

Our third essential feature of a record of reference is that it must be citable in a fixed version. Our system of print publication handles this requirement well: editions of a work are generally unambiguously identified, and reference to a specific section of a specific edition of a printed work leads any reader to identical content.

Putting these features together, we can summarize the form of scholarly published works as works possessing an explicitly identified edition and explicitly identified citation scheme, that can be irrevocably and identically replicated.

With these working definitions of both the purpose of scholarly publication and the form of scholarly published works, we can now ask what a published work might look like in digital media. The defining functional characteristics, permanence and reference, have both been elusive in the internet's digital environment. By focusing first on the formal features of scholarly published works, we can treat replicability, alienation and citation in fixed versions as requirements that our technological specifications must satisfy.

A repository might comprise nothing more than a collection of published works, but a scholarly library is not just a repository. The function of a scholarly library is to provide access to a collection of published works in a setting that makes it possible to use them for scholarly work. A library providing access only to printed works will still include some kind of catalog for finding works, and will set aside space, such as a reading room with tables, where readers can bring together and work with multiple published works simultaneously. In brief, we define a scholarly library as a setting enabling scholarly use of a collection of publications.

John Unsworth has suggested that for the humanities, at least, this use is founded on "scholarly primitives." His contribution to a symposium on "Humanities Computing: formal methods, experimental practice" sponsored by King's College, London, May 13, 2000, and entitled "Scholarly Primitives: what methods do humanities researchers have in common, and how might our tools reflect this?" is available at http://www.iath.virginia.edu/~jmu2m/Kings.5-00/primitives.html As a provisional list for consideration, Unsworth suggests that scholarly primitives include the actions of discovering, annotating, comparing, referring, sampling, illustrating, and representing. Unsworth intentionally leaves unspecified exactly what might constitute the direct object of all of these verbal expressions: it is after all central to his argument that these primitives cross disciplines, and therefore will be applied to the distinct subject matter of each.

The distinct subject matter of different disciplines will still be represented in the published record, however. An editor of a text presents a citable version of that text to the scholarly world, an archaeologist who publishes finds from an excavation creates a citable representation of those finds, a reviewer of a unique dramatic performance might create a citable representation of the performance. The meaning of these representations will vary widely from discipline to discipline, but their representation in the scholarly record suggests that we can further qualify Unsworth's idea: scholarly primitives, as they appear in published work, operate on other published material. Even if an archaeologist compares two newly discovered artifacts, the publication containing the comparison will have to present (and represent) those artifacts before comparing them, so it remains true that any reader of the publication can consider the comparison as working with two objects citable in published material.

If we accept the suggestion that in scholarly publications, we can only compare, annotate, illustrate, etc., objects represented in the published record, this suggests that Unsworth's categories of discovering and referring are functionally and logically prior to his other "primitives." The specific functions the scholarly library must support are first and foremost the actions of identifying and citing published work.

Given our previous specifications for digital published works, how should we design a library for their use? First, we can note that the digital library does not need to be a single physical location. The collection of digital published works it makes available for scholarly work might be distributed; many physical collections might even be treated as a single virtual library.

While it would be easy enough to imagine a distributed digital repository where we store digital works, the more difficult question remains: how should we enable scholarly use of these published works? How should a digital library support activities like John Unsworth's scholarly primitives of discovery, referring, or comparing?

The question is much simpler for a world of print publication: if the only means of using published works requires physical access by a human reader, then reading rooms with generous tables will solve some of the most basic problems. But we have emphasized in our design of digital works the importance of exposing their semantics for reuse by other software. To create an environment for working with digital published works, a digital library will have to enable digital works to build on previous digital published works by interacting with them through these interfaces.

John Unsworth's concept of "scholarly primitives" is helpful in thinking about this challenge, although I believe his analysis of how we should design software to address scholarly primitives is misdirected. His valuable insight is that higher-order scholarly activities depend on prior primitives. Unsworth imagines implementing these primitives in software that an end-user would consult, such as a web page. A human intermediary is of course always necessary when working with print media, but the promise of a digital library is that these primitives can also be implemented as methods for digital publications to interact with each other.

If we try to imagine a broad architecture for permitting interactions among potentially distributed objects, we need look no further than the internet. The internet depends for its most basic functionality on protocols defining how software should transfer information from one location to another. The fundamental job of converting names to addresses in the internet's numeric format, for example, is carried out by a service, a program following a protocol for exchanging information between a client and a server. A "domain name service" fields queries about names like "chs.harvard.edu" and returns information like a numeric address of 207.188.245.132.

We can organize our digital library similarly as a network of services permitting interaction among objects using specified protocols. Services corresponding to Unsworth's primitives will be foundational for other, more complex operations. First and foremost, a digital library's services will have to include services permitting the discovery of published works and their citation schemes, and reference services enabling retrieval by the defined citation scheme. Other services will be built on top of this foundation. A "difference service" might take two references and apply an appropriate algorithm (specific to the type of objects being retrieved) for defining how they differ. The results of a difference service might in turn be used by higher-order functions, say a statistical analysis of results from a series of "diffs". Human users could of course inspect results at any stage of this pipeline, as they must do in print, but the distinctive power of the digital library lies in its potential for augmenting this intervention with interaction among automated services.

As an initial specification for our scholarly digital library, we can say that it should:

Beyond this initial list, experience with digital libraries may suggest higher-order scholarly activities that need to be supported on top of the minimal primitives in the form of additional services. These services could be a mix of further generic services and discipline-specific services: since the library is after all a virtual construct, it would even be possible for different digital libraries to provide different kinds of scholarly use of an overlapping set of published works, provided that standards for discovery and reference are accepted.

The American Precision Museum in Windsor, Vermont, tells the story of the remarkable changes that overtook production of machinery in the early nineteenth century, most vividly illustrated by the rapid changes that took place in the production of firearms.

An eighteenth-century gunsmith might devote an entire month to producing a single rifle. This highly skilled craftsman would have to fashion the individual parts, from the wooden stock to the numerous metal pieces of the lock mechanism, and then, judging by eye, carefully work each piece until it fit precisely with the other individually shaped pieces. The resulting masterpiece of craftsmanship was a valuable and unique artifact&emdash;unfortunately for hunters or soldiers in the field, who could neither interchange parts among weapons, nor themselves repair the custom-crafted rifle.

The revolution in gun production depended on the use of machine-made parts. At the London Crystal Palace Exhibition of 1851, the Robbins and Lawrence Armory of Vermont could exhibit a series of rifles made of interchangeable parts. Each of the separate parts of the rifle was manufactured to a highly precise specification; at the Vermont factory, machine operators who needed only modest skills could manufacture the parts in quantity, and in some cases even operate two or more machines simultaneously. To the amazement of visitors to the London exhibition, the rifles could then be assembled from the perfectly interchangeable parts with only a screw driver. In like fashion, soldiers in the field needed no more than a screw driver to swap out damaged parts with parts from other guns. Interchangeability of precisely specified parts effected a revolution both in how guns were manufactured and how they could be practically used and maintained. Mass production did not eliminate the need for highly skilled labor, however: it organized it and distributed it differently. The entire range of the old gunsmith's craft had to be analyzed, each part precisely specified, and machines capable of producing each part to the required degree of precision had to be designed and built. The engineers who accomplished these impressive feats had a different set of skills from the old gunsmith, but developed and perfected a set of talents at least as rare as the gunsmith's. The product of the gunsmith's month of specialized labor could serve one hunter or soldier (at least, until the gun failed); the product of the nineteenth century inventor and machine builder could keep a factory running around the clock producing rifles.

This seems to me to offer a close analogy to our current situation in scholarly publishing. The challenge of designing digital published works for digital libraries is precisely to break down the old tasks into constituent parts, specify them, and design machines (in this case, in software) to carry out those tasks. We could transform the hand-tooled publication process resulting in a unique piece of craft work into a system for producing published works from precisely specified, interchangeable parts. We will not need less scholarly expertise to accomplish this: but we will need to direct scholarly talent away from crafting unique objets d'art, and towards contributing to a system that can change both how we produce publications, and what kinds of secondary uses can be made of them. We will need no less imaginative creativity to accomplish this: to the imagination of the author, we will have to join the creativity of the programmer.

When I have presented some of the ideas in this paper to other humanists, one common response has been, "But I've never seen a digital published work that fits your definition." Often the implication is clearly that no digital work satisfying the definition offered here could possibly exist.

While I can only agree that I have not yet seen a digital published work in the humanities that meets all the criteria that I have proposed, the suggestion that it is impossible to create such works organized in digital libraries of the kind described here is disproved by the counterexample of a very conspicuous community: the global collaboration known as the free software movement. Every guideline we have suggested for the production of scholarly published works is represented by the best free software. Software is distributed in fixed versions, written to open specifications, with freely available source code distributed under licenses allowing full reuse. Projects can design collections of such published software that interact predictably&emdash;that is, they can create an environment for working with free software analogous to the role a library plays for published works.

Compare, for example, two "digital libraries" designed to create a working environment using the Linux operating system, the Knoppix project and the Gentoo project. The Knoppix project http://www.knoppix.org/ puts together an operating system and suite of programs on a single CD. Users can boot a computer from the CD and work with applications like Web browsers, office suites, and many other "published works." The Knoppix project creates a compact, portable working environment that a user can carry on a single CD.

The Gentoo project http://www.gentoo.org/ corresponds to what I have earlier called a "virtual library." The physical repository is distributed at mirror sites across the internet; you can "check out" a work to use on your local computer using a logical reference that might or might not include specifically requesting a particular "edition" or version. You "check out" a published work using logical references; the Gentoo system converts these to physical locations at one of a series of mirror sites, and the material is downloaded to your computer. The Gentoo system tracks "cross references", or dependencies, among different published works in the library, and acquires new "editions" as needed or requested, so if you wish to use a new version of a Web browser that requires "updated editions" of any further "published works," the Gentoo environment sees to it that the necessary "cited works" are available. The Gentoo project allows you to create a highly customized working environment tailored to your interests, and as up to date as you like.

It should not be surprising that the free-software community has evolved a "publication" system that so closely satisfies the needs of the scholarly community, for the operation of this self-organized community may more closely model the scholarly ideal than any other community anywhere. Published contributions can be seen by anyone, and the resulting rigorous scrutiny of the source code in free software projects is a major reason for the high quality of projects like the Apache web server, the most widely used web server on the internet. Status in this community can only be gained by submitting work that your peers value.

The work of the highly skilled (and often highly individualistic) programmers who contribute to this community could not be further from the routine of a machine operator in a nineteenth-century factory. They work with the gossamer fabric of ideas, rather than molten metal or wood, but the community's productivity has, like that of the nineteenth-century factory, experienced explosive growth as these software engineers have developed a carefully designed system of interlockable parts made to precise specifications. Engineering ideas in this way is as natural to a community sharing the scholarly world's values as it was to the engineers who transformed the building of hardware in the early nineteenth century.

Our digital incunabula are not yet recognizable as digital published works. Modern scholarship has seen the craft of a skilled manuscript copyist give way to the industrialized production of printed works. This mass production helped effect new uses of writing throughout society, and gave rise in turn to a new set of specialized crafts built around publishing in print. These crafts will now yield in importance to the a new set of crafts centered around the production of digital publications. It is our responsibility as scholars both to learn from this history, and to realize that there are already functioning examples today showing how we might begin to move beyond digital incunabula into an era of true digital publication for work in digital libraries.