Appendix B. Glossary

Terms are explained as this essay and the project record use them. Where the project fixed a definition, the page that fixes it is linked.

Accuracy. One of the eight senses of a good translation: the source's propositional content, imagery and stated detail carried over "without unlicensed addition, omission, or distortion." A choice the target's grammar forces — a compelled specification — carries no penalty. Typology.

Affect. The sense of a good translation under which "the joke lands, the dread accumulates, the ending stings." A whole-text property that the project could reach only through documented reception records, of which it found two.

Anchor. A page recording a piece of published human evidence read whole: either a naturalness anchor (original writing in the target language, fixing a register point) or a precedent anchor (a published translation read against its source, recording what the translator did and what it cost). Twenty-three were built. Anchors.

Arm. From July 26 to September 4, the project's unit of multi-session work: a declared question, a completion criterion, a step budget. Ninety-nine were constituted; "a rabbit hole is an arm that outran its budget without declaring it." Replaced by workstreams on September 4.

Baton. The file, NEXT.md, that each session rewrote for the next: the assignment, the state, what was blocked, the budget position, and a block written for the owner.

Binding register. In serial translation of a long work, the running document of names, terms, forms of address, voice rules and decisions that later spans must obey, revised only by numbered erratum, with a list of questions the next span must settle.

Chime. A rhyme or near-rhyme at the end of a clause or line; in Persian and Arabic rhymed prose (saj') the paired clause-ends chime. The chime slot is the position the rhyme word occupies.

Compensation. Putting a device of the translator's own where the source's device could not be carried — a sound figure elsewhere, a marked address noun for an untranslatable pronoun. The handbook's oldest positive instruction, and one the record narrowed repeatedly.

Consistency. The sense of a good translation under which "names, terminology, motifs, register do not drift without cause"; weight grows with length.

Contamination. The risk that a system's "own" rendering echoes a published translation absorbed in training. Measured before material was chosen: the longest run of shared words and the count of shared seven-, twelve- and fifteen-word sequences against a reachable comparator. Rule.

Cultural mediation. The sense of a good translation under which culture-bound items — realia, allusion, measures, names, titles — are handled "so the target reader neither trips on opacity nor is handed a flattened, de-situated world."

Declared purpose. Not a sense but a parameter: the reader and use every evaluation must state. Demoted from a sense (purpose fit) on August 1.

Domestication, foreignization. Venuti's terms, from Schleiermacher: moving the author toward the reader (fluent, assimilated) or the reader toward the author (marked, retaining the source's difference). The project used the map and not the politics, and found the binary too coarse for its best case. Source page.

Dose. In the calibration test, the number of damaged sites in a passage — eight ("heavy") or three ("light").

Enact / state. For a Japanese sound- or manner-word, rendering it with an English phonaestheme (shamble, thump, thump) or with a plain adverb (slowly); the record found these commit to different things.

Evidenced / untested. The tags on every instruction in the handbook: evidenced, with the language pairs it was measured on, or untested, meaning no run has executed it or no evidence reaches that pair.

Footing. The relation a source grades between speaker and addressee — by pronoun choice, honorific morphology, a deferential clitic, a respectful plural — that English mostly does not mark grammatically. Entry.

Frozen. Of a design, a translation, or a log: committed to the record before the step that could have influenced it (the run, the reading of a comparator, the evaluation).

Held-out arm. A control in the calibration test: two independent competent translations of the same passage, on which the jury should sit at chance.

Internal-judgment-only. The flag on any evaluative statement that rests on the project's own unverified judgment rather than on an anchor. Most of the record carries it.

Jury, panel. The outside AI models, from companies other than Anthropic, used as blind judges (the jury), as critics of designs, and as contrast translators. Their agreement with each other was treated as weak evidence.

Lead. The project's own AI agent — the model that ran each session, translated, designed, measured and wrote. The "I" of this essay.

Long work. A text translated serially across many sessions under the serial regime; six were finished.

Naturalness. The sense of a good translation under which it "reads as fluent, idiomatic prose of the target language," judged on the target text alone against a declared register; a low score is markedness, not badness.

Perceived source carriage. The sense, added August 2, under which a marked departure from target norms "reads as a deliberate attempt to carry over a feature or effect of the source" rather than as a failure. Depends on whether the reader can see the source.

Placeless. Of English prose: carrying no word a reader could place nationally. The record found no published book is placeless; it "delays, then commits, then stays committed."

Provisional. The flag on every quality score from the jury while the jury had not passed calibration — which is to say, on every quality score in the record.

Radif. In Persian verse, a word or phrase repeated identically after the rhyme at every rhyming position of the poem. Entry.

Realia. Culture-bound things — a dish, a coin, a garment, a district, a rank — that the target culture does not have a word for. Entry.

Register. The level and texture of the language — high or low, period or contemporary, marked or neutral. The record found that lowering a register has "no free device." Entry.

Regime. A fully specified way of producing a translation: which model or hand, what it may read, whether it revises, under what rule set. Fifty-nine were written. Regimes.

Sham arm. A control in the calibration test: an edit of the same size as the damage but neutral in quality, on which the jury must sit at chance — "what makes a positive result mean anything."

Specificity. In the calibration test, the requirement that when one sense is damaged, that sense's score fall further than the others. The jury detected damage at ceiling and failed specificity at the heavy dose.

Style correspondence. The sense of a good translation under which the source's marked formal features "receive functional equivalents rather than silent flattening."

Tier D, Tier P. The two calibration tiers: detection (can the jury detect deliberate, sense-targeted damage? — the gate) and peer discrimination (can it reproduce a documented human ranking of two good translations? — a certification). Neither passed.

Translator's log. The numbered record of decisions and alternatives kept with every translation, frozen before any evaluation is designed; "decisions and alternatives, not quality."

Voice. The sense of a good translation under which "the reader of the translation meets someone, and the question is whether it is the someone the source presents."

Workstream. From September 4, one of three lines of work with a finish line and ordered steps — the handbook, the jury, the practice — that sessions rotated among.