LABELING_GUIDE.md

Labeling guide

A decision guide for the human reviewer in the review UI. See README.md for pipeline mechanics; this doc is purely about how to label.

Verdicts

• match — the CCE entry is the U.S. copyright registration for the work the MARC record describes. Same work, same edition, same language.
• no_match — the pair is not the same registration. Capture the why in the note field if anything about the call is worth flagging.
• unsure — there's genuine ambiguity even after looking. Prefer this to guessing. Use the note to record what the doubt is.
• skip — keyboard s / space. Use sparingly; the pair stays in the queue. Prefer unsure so the doubt is recorded.

Different language → no_match

The most common false positive: same title, same author, different language. A 1953 Spanish translation and the 1925 English original are different copyright registrations with different registration numbers in the CCE — they share an intellectual work but they're not the same record. Mark these no_match and mention "translation" or the language in the note so the pattern is recoverable in later analysis.

CCE tells the labeler the registration is for a translation rather than the original:

• Phrases in the CCE title / note like "translated by", "translation of", "[Translation]"
• Author field on the CCE side names the translator, not the original author
• Claimant is the translator or translating publisher

When you genuinely can't tell whether the CCE entry registers the original or a translation → unsure, note the ambiguity.

Series and sets — asymmetric rule

When MARC and CCE describe different scopes within the same series or set, the verdict depends on which side is the broader entity. The question we're answering is "does this CCE registration cover the work the MARC describes?" — not "are these the same bibliographic entity?" Coverage flows from the broader registration to the narrower work, not the other way around.

• MARC = series/set; CCE = single member → no_match. The CCE registration covers only the specific volume; calling the pair a match would imply the registration covers the whole series, which it doesn't. This is the common case (Princeton's MARC frequently describes whole multi-volume sets; CCE registrations are usually per-volume).
• MARC = single member; CCE = series/set → match. A series-level registration covers the members of the series, so the member-level MARC is genuinely covered by that registration. This is rarer but real. Capture "series-level CCE" (or similar) in the note so the inference is visible downstream — a future consumer needs to know this is a series-coverage match, not an exact-volume match.

When you can't tell whether a CCE registration is series-level or volume-level → unsure, note the doubt.

Country/publisher divergence (transatlantic editions)

Divergence in publisher and publication country alone is never grounds for no_match.

• If the connection can be affirmatively established — a documented imprint or co-publishing relationship, renewal data confirming the same translation or setting, or overwhelming content identity (unique title + author + year) — label match, and add the same_work_foreign_publication category when the CCE registration's publication event is in a different country from the MARC's.
• If the connection cannot be established, label unsure with [reasons: pub_differs].
• no_match requires a content-level disagreement: a different work, a different title, or an incompatible extent within the same publisher.

Evidence anchors:

• Same-year publication of the same title/author across countries is presumptively the same text. Simultaneous transatlantic publication was standard practice under the 1909 Act: its ad interim provisions pushed publishers of English-language works first published abroad to secure US publication quickly (this is the mechanism behind the AI registration class).
• Extent differences across different publishers are weak evidence of content difference. The manufacturing clause required US typesetting for full-term protection, so page counts routinely differ between printings of identical text. Extent is a strong differentiator only within the same publisher.

same_work_foreign_publication and different_edition sit at opposite ends of the same axis, and it is worth keeping them straight: same_work_foreign_publication means the same content under a foreign publication event (a match), whereas different_edition means the content itself differs enough to be a separate registration (a no_match or unsure). This is guidance, not a rule the tooling enforces — genuine edge cases exist, so let the evidence decide the verdict and use the category to record what you saw.

E-book reprint badge

When the card shows a yellow E-book reprint badge, the MARC record's extent field contains "online resource" — Princeton's record describes a digital reissue, not the original publication. Year and publisher in the MARC side belong to the digital reissue (e.g. a modern aggregator), not to the original artifact the CCE entry would have registered.

E-book reprints are filtered out at acquire time (#30). If you ever see this badge, the filter missed an indicator. Note the pair_id and mention "e-book" in the free-text note; we'll extend the filter.

The note field

The note is the only structured signal alongside the verdict. Optional — leave it blank when the call is obvious — but use it freely when something is worth flagging. Useful things to capture:

• What surprised you. "Author looks right but publisher is from a different country" / "title matches verbatim but the work is clearly different".
• What made the call ambiguous. "Could be a translation or the original — CCE entry has no language hint" / "two distinct works share this exact title in the same decade".
• What the matcher seemed to get wrong. "scorer underweighted the author here — Doe vs J. Doe" / "year window may be too tight — reg date is one year past the publication year".

Notes are not meant to be read individually after the fact; they get analyzed in aggregate to surface patterns the matcher and the labeling workflow should learn to capture. Be specific, but don't worry about consistency of phrasing — that's a job for the later analysis pass.