Tagging Without Black Boxes in the LLM Era: A Faceted, Rule-Based Protocol for Bibliographic Annotation

Researchers annotating an evolving bibliographic corpus with a faceted vocabulary re-tag the same items as the vocabulary stabilises, and probabilistic annotation makes each pass drift in ways the analyst cannot reconstruct. Here, we present a deterministic, rule-based protocol that publishes the annotation function itself, f(item) → tags, as editable CSV tables that detect cues in the title-and-abstract surrogate, map them to controlled tags, and derive further tags from those already placed. A linter certifies the tables before each write, a triage routes every departure to a repair or an inscribed analyst verdict, and versioning keeps every change revertible as a single line. Applied to a 402-item mutual-aid and crisis bibliography across fourteen dimensions and 155 controlled values, the function returns a SHA-256-attested identical tag set on every pass under frozen tables, so re-passes differ only where the analyst intervened. A coverage diagnostic, scored against a single LLM reference reading, characterises the surrogate scope, and construct validity remains a separate argument. Dictionary content analysis dissolves design judgment into the lexicon and probabilistic annotation migrates it into learned weights, whereas this protocol keeps that judgment a first-class, versioned object of the published record. For evolving corpora in the digital humanities, inscribing the annotation function keeps the deltas between passes readable as deliberate analytical decisions.