Summary

  • Four documents printed four counts of the same population in two and a half years: about 600, some 900, approximately 1,100, over 1,200. Only the last carries an as-of date in the same sentence as the number, and it was completed by Peter Deutsch in November 1993.
  • The Fall 1992 figure of approximately 1,100 carries Alan Emtage's byline, prints no measurement date, and cites his Winter 1992 conference paper in the place where a date would go. That paper's body sits on pages 93 to 110 of the proceedings and was not reached in readable form for this article, so its table is reported here as unread, not as missing.

A counted list, hedged once, dated by the page

Start with the earliest removable doubt. The archie manual page carries a date line of 20 May 1991 and is attributed at its end to Alan Emtage and Bill Heelan of McGill University. In the middle of it, plainly: "The archie database subsystem maintains a list of about 600 Internet ftp(1) archive sites." The same page adds the two operational facts that make the number mean something — each site is updated approximately once a month, and only UNIX sites are included. (archie manual page, 20 May 1991)

That is a document date, not a measurement date. Nobody on the page says the 600 was counted on 20 May; the page is dated, the number is hedged with "about", and the population is restricted to UNIX sites in the same breath. Read it carefully and you already know the figure is a floor built on a filter.

Nine hundred, in somebody else's question-and-answer file

A year later the number is larger and further from the people who ran the service. RFC 1325, the FYI answers file for new Internet users, printed: "The system curently makes available the names and locations of some 1,500,000 files at some 900 archive sites." The typographical slip is in the original. The bylines are G. Malkin of Xylogics and A. Marine of SRI, and the document's own date is May 1992. The same passage puts nine archie servers in service. (RFC 1325, May 1992)

Note what has changed. Six hundred was a first-party statement about a list archie kept. Nine hundred is a third party describing what the system "curently" makes available. The word "currently" is doing the work a date would do, and it is anchored to whenever the authors were drafting, not to a day anyone can check.

Eleven hundred, undated, and signed by the man himself

Then come the Fall 1992 pages that most of this literature leans on. Michael F. Schwartz of the University of Colorado at Boulder, Alan Emtage of Bunyip Information Systems, Brewster Kahle of Thinking Machines Corporation and B. Clifford Neuman of the University of Southern California's Information Sciences Institute published "A Comparison of Internet Resource Discovery Approaches" in Computing Systems, Volume 5, Number 4, beginning at page 461. In section 3.3 they write that archie "maintains a list of approximately 1,100 UNIX anonymous FTP archives world-wide", and in the same sentence they cite Emtage and Deutsch 1992. They add about 150 gigabytes of information, approximately 2.6 million files and 13 replicated archie servers, and they describe the collection method: recursive directory listings at each site once per month. (Computing Systems, Vol. 5, No. 4, Fall 1992)

Here is the whole problem in one sentence. The figure most often attributed to the 1992 paper is printed on a different page, in a survey, by a byline that includes the paper's first author; and where a careful writer would put a measurement date, the survey puts a citation. Approximately 1,100 is a real printed figure. It is not a dated one. The extracted text this reading rests on shows some broken word spacing, so the digits themselves are read with that caution, and the absence of a date is not an artefact of the extraction.

Twelve hundred, and finally a day

The only count in the set that arrives with its own clock is the latest. RFC 1689, published as FYI 25 in August 1994 but assembled from templates contributed by the projects themselves, states: "The initial implementation of archie tracks over 2,100,000 filenames on over 1,200 sites around the world (as of November 1993)." The archie template in that report was completed or updated on 1 November 1993 by Peter Deutsch of Bunyip Information Systems, and the appendix, checked or updated on 1 March 1994 by the same hand, says "there are currently over 1,200 anonymous FTP sites on the Internet" with about 27 archie servers. (RFC 1689, FYI 25)

The date sits inside the sentence with the number, which is the whole of the technique. Note also that the count is bounded — "over 1,200" — so even the dated figure refuses to be a precise denominator.

The file counts run the wrong way

One check a reader can make without leaving the set. Three of these documents also printed how much material archie held: about 1,500,000 files in May 1992, approximately 2.6 million in the undated Fall 1992 survey, and over 2,100,000 filenames as of November 1993. Read as a series, the middle figure exceeds the later one. Nothing here is evidence of a shrinking index; the likelier reading is the boring one, that each figure was taken on its own basis, by its own author, for its own purpose, and none of them is a monthly measurement of the same object. Two and a half years of a growing service, and the numbers go up, then up a lot, then down.

That is the signature of snapshots, not of a series.

The paper everyone cites

The Winter 1992 USENIX paper "archie - An Electronic Directory Service for the Internet" by Alan Emtage and Peter Deutsch of McGill University exists, was presented in San Francisco in January 1992 in the Innovative Applications session, and begins on page 93. The digitised proceedings scan confirms all of that from its front matter, its table of contents and an author index that prints "Peter Deutsch 93" and "Alan Emtage 93". A February 1992 posting of the conference programme reproduces the same title, authors and affiliation. (proceedings scan, front matter and author index) (USENIX programme listing, February 1992)

What the scan did not give up in readable form is the paper's own body, pages 93 to 110. No table, no column heading and no printed date from inside the paper appears in this article. This matters because a widely repeated pair of numbers is attributed to exactly that table: 1,025 sites known, 886 indexed, said to be dated 30 October 1991. In October 2026 a BTW article on this subject stated those figures and the table they sit in. (BTW, 8 October 2026) On 21 November 2026 a second BTW article said the same table was not consulted for that piece and was not used as evidence there. (BTW, 21 November 2026) The honest statement of position is the uncomfortable one: two of this desk's own pieces have taken different positions on a table that neither the earlier piece nor this one has read. Until a page image or an intact text of pages 93 to 110 is examined, the pair 1,025 and 886 is a claim with a citation and no inspection behind it.

One more 1992 document cuts against the neat story. Peter Deutsch's McGill master's thesis, dated June 1992, describes the system in language that reaches this article only as an excerpt rather than as a document opened here: it "stores several million records gathered from several thousand sites". (Deutsch, Resource Discovery in an Internet Environment, June 1992) "Several thousand" and "approximately 1,100" cannot both be a count of the same thing on the same basis in the same year. Either the thesis is counting something else — records, or a wider universe of sites — or the two hands that signed one paper were working from different pictures of their own system. The excerpt does not say which, and this article will not decide it for them.

Where my arithmetic stops

What I can do is smaller than what the literature does. I can set the printed counts side by side because each is printed with its document. I can subtract 900 from 1,100 and say the survey's figure is larger than the RFC's, and I can point out that neither number is dated, which makes the difference a difference between two printed statements rather than a measure of growth. I can note that the only bounded count, "over 1,200", sits more than a year after the undated 1,100 and cannot be compared to it as a rate.

What I will not do: divide 886 by 1,025 to produce a coverage ratio from a table I have not read; convert any count of indexed sites into a share of the Internet, because no document in this set prints a denominator of sites that existed; turn 600 to 1,200 across two and a half years into a growth rate, because two of those numbers are hedged, one is bounded, one is undated, and the population is defined only as "UNIX anonymous FTP archives" in some sources and as "sites" in others; or treat the manual page's 20 May 1991 date line as the day someone counted 600 sites.

My arithmetic is mine and it claims exactly what the printed numbers allow: four different numbers, printed by four different documents, with one date among them.

That is the finding, and it is duller and more useful than a growth curve. The figure that travelled for three decades travelled as a number, and the thing that did not travel with it was not a caveat or an apology but the ordinary machinery of a measurement: the day, the unit, the basis of counting and the name of whoever counted. One page in the set has the whole machine in one sentence. It was written by the co-author, two years after the number everybody quotes.