Normalisation.
The first stage of understanding.

A letter says he, the St. Louis office, last Tuesday, Dr. Wattendorf. A reader understands these expressions by knowing what they stand for: a particular person, an office in a city, a day, an individual of the register. The normaliser of famulus-proof makes that knowledge explicit. For every expression of a sentence it writes a footnote: the individual or the interval it stands for, a status (fixed, open or background), a grade, and the grounds — the register entry, the letter’s sender, the sentence that names the person, each with its address.

The work runs as a score of named stages: the context is laid out, the Stanza analysis is read, every phrase is assigned an expression type, the type’s resolver reads it, critical questions are put to each reading, a verdict is taken, the footnote is written, and the keys for search are stored. The verdict is the grade of the weakest ground or assumption that carries the reading, never the strongest.

In the Kármán lab copy, 2,270 sentences carry 6,556 footnotes. The normaliser’s code has 5,653 lines; its rules are not in the code but in declared files — the expression types, the language lists, the context layers and the relevance model — so that a learning loop may revise them and a reviewer may read them. See the sentence record in the documentation.

Which earlier person does he take up? The normaliser does not answer this with rules that name a page. It scores every candidate the text before the pronoun mentions on declared criteria: how near the last mention is, the role it had in its sentence, whether earlier pronouns already took the candidate up, how often and how it was introduced, whether its sentences share the topic of the pronoun’s sentence, whether the kind fits, and — in a letter — whether the candidate is the writer or the addressee. The weights and the threshold stand in a file.

A reading is fixed only when the best candidate leads the second by the declared margin. Otherwise the footnote stays open and lists the scored candidates. The footnote of a fixed reading cites every criterion with its value and weight, and the alternatives that lost with their distance — the reader sees why one person won and by how much.

The weights are learned, and learning is guarded. A learner may only change the declared attributes; an edit that names a person, an identifier or a word of the probe sentences is rejected. A candidate is tested on frozen probes: the train case, a held-out case of the same kind, a book, and guards taken from audits. It is kept only when it loses no expectation that held before and gains beyond the case it was fitted to. So far 3 rounds were run and 0 kept. An earlier candidate that lowered the threshold decided many open pronouns, and the audit of whole letters found more than half of those decisions wrong. That audit became the guards. The synthetic tests of the model, with invented names in English and Latin, count 8 functions.

A letter has a writer, an addressee, a date and a series of letters before and after it. A book has none of these. The normaliser reads either carrier: for a book the unit is the document’s running text, and the criteria that need parties or a series switch themselves off — the footnote says so instead of faking them.

The test case is Euler’s E57, a Latin printed book, read in a lab copy of 1,036 sentences with 1,706 footnotes, without writing into the collection. Names stand there in small capitals and Latin cases (NEWTONUS, NEWTONO). They reach the register person through declared folds — small capitals folded, a declared case ending removed — and every fold that fires is a ground of the footnote. A Latin pronoun that may stand for a thing competes with the nouns that agree with it in gender and number. The book probes pass 5 of 6 expectations; the pronoun that still fails stays open rather than guessed.

The method and its first results are described in the drafts Famulus normalisation (2 pages) and The Famulus Probe Lab (19 pages): normalise, filter, classify — generating critical annotation for the items of a reference storage.