Bibliography

Provenance here is graded rather than binary. A file with a sidecar beside it and a transcript that can only be traced back to the scan it was made from are both traceable, but they are not the same claim, so they are counted separately.

The 9 unresolved files sit in 7 entries: Daily Northern Islander (2), Northern Islander (2), cihm 07807 (1), indiannamesofpla0000kelt (1), indiannamesofpla00kelt (1), NOAA bathymetry (1), newspapers com (1). They are listed rather than rounded away.

A year against a source here means one of two different things, and the catalogue keeps them apart rather than averaging them into one column. For 441 works — the newspaper runs and the retrieved pages — it is the span actually held, read off the files themselves, so a paper that ran for a century may show five years of it. For 56 works it is the year of publication as the holding institution states it in its own catalogue record. Hovering a year says which.

130 works have no year at all: geospatial layers and archived websites mostly have none to have, and many scanned items carry no parseable date in their identifiers. Where a year exists only in prose — one Internet Archive note reads "1885 reprint with supplement (project holds the 1854 printing separately)" — none is recorded, because a machine cannot tell which of those two years is the item's, and a wrong year looks exactly like a right one. Sorting by year gathers the undated at the end under their own heading; they are never sorted as though they were the year zero.

Each source above carries a cited by figure. It counts the archive's own knowledge-base articles whose citations point at a file belonging to that work — 9,931 citations across 7,030 articles, joined on the local_path each citation records. An article citing forty issues of one newspaper counts once, because it has used one source.

It is deliberately not a full-text hit count. A hit total is a property of whatever you typed into the search box; it would rank these six hundred works differently for every reader, and a column labelled "hits" would be a number nobody could check. This one does not move when the filter above is used.

119 of 627 works are cited by at least one article. The other 508 are held and not yet written about — 414 of them single newspaper pages retrieved because they mention the island, and 81 of them books. They show cited by 0 articles rather than nothing at all: an archive that holds more than it has read is the normal condition of an archive.

It is a measure of use, not of importance, and the catalogue's grouping rules differ in granularity — one work covers every State of Michigan layer, while a single Chronicling America run is its own work. A large count can mean a source was heavily used or merely that it was carved coarsely. The table below gives the composition of the largest figures, which is what tells those two cases apart.

4 entries in that table — State of Michigan curated layers, GLO survey field notes and land-entry records, Alan Lomax, Michigan folk-song recordings, Beaver Island (1938), GNIS names, variants and superseded names — are cited mostly by document/ records the knowledge base seeded one-per-feature from the layer itself. That is a real count of real citations, and it is not the same kind of use as the Beaver Beacon's, which is cited by narrative articles about events and people. Both numbers are true; only one of them is scholarship.

Of 63 distinct source hosts, each asked once by HEAD request, at least two seconds apart, and retried before any failure was believed: 618 answered, 2 redirected to a different domain, 0 returned nothing at the recorded address, 0 did not answer at all, and 3 were not contacted, because their robots file asks not to be crawled.

The categories are kept apart on purpose. A host that returns 404 is telling us the document has gone; a host that stays silent is telling us nothing, and the fault may be at this end. Earlier drafts of this survey merged the two, did not retry a dropped connection, and probed search endpoints as though they were documents — and so reported three living island websites and an entire federal land-records service as dead. None of them were. Those results were caught by checking the hosts by hand rather than trusting the tool, and the corrections are the reason the counts above are as dull as they are.

A status describes the host on the date shown. It does not promise that a particular document behind that host is still there, that its address has not changed, or that a reader without a subscription can see it.

4 entries record no source URL to check — mostly material supplied directly rather than retrieved from a website.

The working tree and the consolidated archive hold overlapping copies of the same material. They are not derived from one another, and they have drifted. This is measured here and deliberately not repaired.

18,061 files are present in both. 1,840 are in the working tree only. 7,537 are in the consolidated archive only — of which 7,519 are expected there by design (the knowledge base, the archive's own documentation, and transcripts withdrawn as unfit), leaving 18 unexplained. Both trees are written by live pipelines, so these counts are a snapshot taken when the catalogue was generated. Drift is measured here and deliberately not repaired; see CLAUDE.md.