7. What translators do — the prose problems
The handbook the project left behind is organized by translation problem: fifteen families, each with an entry that says what the problem is, what published translators did when they met it, what the project's own practice found, what the options are and what each costs, and then gives guidance twice — as numbered instructions for a person and as steps a pipeline could execute, with the places a person may step in. Every instruction is tagged evidenced, with the language pairs it was measured on, or untested. Five entries were finished in the ordinary way before the close, each applied to a fresh passage with a log of whether its instructions decided anything; nine were consolidated at the close and are marked drafts. This section reports the five finished prose families; the numbers are the entries' own.
Footing and address
Russian has ты and вы; Japanese has a morphology of deference; Bengali grades its verbs by politeness. English marks almost none of this grammatically. The one rule the project's framework had admitted, in August, addressed exactly this: where the source marks a relation by a form the target lacks, "the absence of a same-category counterpart is not the absence of the marking" — render the site again under a brief that requires the marking, and let it fall wherever English does mark such things.
The entry on footing and address records what happened when that rule was tested across ten language pairs. Where the marking rides in a slot English fills compulsorily with one form — above all the pronoun — "the marking does not reach English, and is not compensated elsewhere either" — "Six measurements, four pairs, all at the floor." But the relation mostly transfers anyway: Garnett, translating Chekhov in 1922, deleted the deferential clitic -с at all seven places it occurred, and three outside readers recovered the speakers' standing from her English with a mean error of 0.048 on a seven-point scale. Across five pairs the project "has not produced a population of sites at which a competent English rendering loses a grammatically marked relation." The instruction that survives is almost the opposite of the one admitted: expect most sites to need nothing, and spend the effort where the marking is all the source has — a bare «Что-с?» with no content to carry it. No predictor of where a lost marking lands survived a test, in four attempts, and carrying everything is mostly bulk: an exhaustive rendering of a Genji chapter grew by a fifth, and the perceived grading "follows local device density, not the source" — carrying every mark manufactures a hierarchy the original did not draw.
Culture-bound words
The realia entry asks what happens at a word that names something the target culture does not have. Its finding is that the choice does two jobs at once, and no measured strategy does both. Substituting a familiar target-culture item makes the prose sound native and, in the same stroke, relocates the story: with a domestic substitute at one site, a translated page read as set outside England in only one of seventeen judgments. The effect saturates at a single word — "Three guineas!" for "Three zlaté!", one substitution in 245, moved every judgment. Transferring the foreign word keeps the setting and places the prose the other way. And at a proper name the rule inverts: Malá Strana rendered as "the Lesser Town" read as more foreign-set than "this quarter."
The published record was consistent in one respect: nobody holds one policy. Glenn Shaw handled one class of Buddhist-afterworld terms four ways in a thousand words of Akutagawa; Garnett handled two dogs two ways in one sentence. The entry's instruction is not to read that as incompetence — "coherence is a decision to make deliberately" — and to decide, site by site, whether the prose or the world is what the passage needs.
Register
The register entry is the record of a general recommendation refused four times, each refusal replaced by a sharper reason. Nonstandard spelling, the traditional device for lowering register, placed the prose: a judge shown an' named it "US Southern/rural." The "placeless" arm of one experiment placed itself anyway, on words its rule had not caught — apologise, trodden, for a song. Measured properly, the claim that English prose cannot avoid marking its nationality was withdrawn: under an explicit rule the translator produced 0.51 marks per thousand words, below the least-marked of twelve published narratives (1.24). What the twelve books showed instead was that "'Placeless English' is not a book with no marks; no such book was found. It is a book that delays, then commits, then stays committed."
On the source-ward side the sharpest number is Andersen's: at sixteen compound words in one tale, all three published hands carried exactly three; an explicit rule carried sixteen. "A framework line of the form where the source compounds, compound would be a recommendation against the entire published record on this pair." And the reader's side produced the entry's most useful caution: a form-carrying rendering read as fidelity only when readers could see the source. Shown the translations alone, twenty of twenty-one preferences went to the flattened arm; shown the source, twenty of twenty-one reversed.
A change of language inside the source
Sa'di's Persian quotes Arabic; Dante's Italian quotes Latin; Tolstoy's aristocrats speak French. An English page cannot change language in the same way, and the entry counts what twelve published hands did at eighty such places. Where the embedded language is a quotation the Victorian reader could read, nobody deleted it: Latin was kept in 70 of 84 cells across four English Vita Nova translations, and marked by contrast of type in 69 of the 70. Where it is a code-switch — a character choosing to speak French — the same readership's translators collapsed it into English in 81 of 96 cells, kept it in 15, deleted it in none. Three-quarters of a century of tradition treated the author quoted Latin and the character spoke French as different acts, and the entry's first instruction is to classify each locus before opening any other translation. The project's own rule, do not supply a signal the author did not give, "put it outside the entire published tradition twice": every one of the four Dante hands distinguishes the Latin they keep.
What the translator tells the reader
The last finished prose entry, on glosses, notes and prefaces, sorts what a translator prints about a translation into four kinds that behave differently. A refusal — Preston, 1850, declining to reproduce al-Hariri's rhymed prose because it would give "an air of flippancy" — is checkable, and held: measured against the chance rate of his own vocabulary, Preston's clause-ends chime above chance in three cells of four, while Chappelow, who refused in 1767, sits at or below chance in all four. A promise — Eastwick, 1852, undertaking to preserve Sa'di's wordplay — was found empty at every one of fourteen enumerated places. A comparative self-report — Walter Leaf, 1898, claiming to use Hafez's rhyme-word "rather more sparingly" — held, against the project's own registered prediction that it would fail. And a disclosure of a source fact moved what blind readers preferred toward the rendering that carried the source's device, while showing them the Arabic itself moved nothing. The entry's own application is small and telling: translating Miyazawa Kenji, the lead drafted a footnote for the title's pun, then struck it "on the check that the ambiguity is present on the page without it."
The shape of the five
Read together, the finished entries share an instruction and a temper. The instruction is the first item of four of the five: enumerate before opening any other translation. The temper is subtractive: do not invent an archaic pronoun; do not treat respelling as placeless; do not expect the free slot; do not add the note until the draft has been checked. And on every page: "What the hands did, never what is better."