11. What is not known

A record that says when it was wrong should also say what it never learned. These are the blank regions, in the order a reader of the findings is most likely to trip over them.

Nothing about what a human reader hears. Every "reader" in the evidence was a model seat, and the charter forbade collecting human judgments. The findings about readers — that a form-carrying rendering reads as fidelity only with the source in view, that an added italic is attributed to the author, that a disclosure moves preference where showing the source does not — are findings about how three or four outside models responded to instructed tasks. Their relation to a person reading a book is unknown.

Nothing about which translation is better. The jury was never calibrated; four attempts and one authorized redesign ended in the word exhausted. The handbook says what published translators do and what a choice costs, and it is silent on whether any option is the right one. A calibrated instrument might exist under a different design — the last verdict page says the failure was of "this specific instrument, this jury composition, this dose range and this no-source design choice" — but none is planned.

Thin pairs. Thirty language pairs appear across the fourteen handbook entries; only four — Persian, Russian, Japanese and Italian into English — are declared by six or seven entries, and sixteen pairs are declared by a single entry. Almost nothing was measured with English as the source, and the record itself notes that reversing the direction "does not reverse it — it replaces it."

One hand, one period. Almost every published translator in the evidence worked between 1767 and 1930, because that is what is free to read; period and translator are confounded throughout. The project's own practice is one translator's — one model family — and its finding that this translator is not an independent sample of itself applies to every comparison it made with itself.

Whether the drafts hold up. Nine of the fourteen handbook entries were consolidated at the close, from the record, by parallel agents, and then independently verified against their sources. They were never applied to a fresh passage with a followability log, which is the step that turned the five earlier entries from drafts into working guidance. The nine are marked draft for exactly that reason.

The long prose work. The charter's last direction was to return from verse to narrative prose and translate a whole serialized novel — a Serbian short novel of 1881 with no English translation anywhere — applying the handbook end to end. The work was chosen on the project's last scheduled day and not one sentence of it was translated.

The regimes never compared. The charter called comparing regimes "the workshop's basic experiment." It was run twice, on an uncalibrated jury; the multi-agent regime reserved on the first day — translator, editor, source scholar — was never specified, and no live human-in-the-loop protocol was ever written. What a person adds at each of the entry points the entries name is, on this evidence, unknown.

The mechanism of self-match. The project measured that its own hand matches itself at up to fifty-one consecutive words and that an opposed rule set collapses the match; it could not separate the possible causes — the model's weights, its decoding, the forcing of the source, the brief — and it could not measure what it remembers of any published translation.

Affect. The sense that most readers would put first — does the joke land, does the ending sting — is reachable only through documented reception, and the project found two records in forty-nine days. A translator's log cannot reach it, a model jury could not be trusted on it, and a serial regime "has no way of making a cumulative effect visible to the person producing it."

The unread. The reading program named what it did not reach: Qian Zhongshu's 化境 and Liang Shiqiu (in copyright), Mori Ōgai's practice, Benjamin, Goethe, Humboldt, Meschonnic, Borges, Chukovsky, Ortega, Fu Lei's 神似, Venuti's history chapters. The typology's claim to be derived from more than Anglophone taste rests on six primaries, read whole, and no more.

What the record does not count. No page lists the source languages or the language pairs; the log's running count is inconsistent; the project never published its own total spend. The figures in this essay for those quantities were computed at the close from the files, and are labelled approximate where the files disagree.

None of these is a reason to distrust what the record does say. They are the shape of the space around it, and the project's charter asked that they be written down: "write the null; when in doubt, under-claim."