2. How it worked

The charter divided the work into two tracks that had to feed each other. The workshop was the practice track: an atelier where translations were produced under specified conditions and a laboratory where those conditions were compared. The poetics track was the conceptual one: a typology of the senses of good, a reading program through the published record, and essays with stated conditions under which they would have to be revised. From July 25 every session's main piece of work had to have both a translation limb — prose actually translated, in the session — and a study limb, "with the wire between them stated in one sentence."

Regimes. A translation was never simply "the AI's translation." It was produced under a regime: a written specification of how it was to be made — a single pass with no revision; a draft followed by a revision pass; a "close" rendering that follows the source sentence by sentence; or, for long works, a serial regime with a binding register of names, terms and decisions. Fifty-nine regimes were written; most of the later ones were one-off briefs for a single experiment — carry every honorific; answer every rhyme; refuse every rhyme — which is how the project isolated the cost of one choice at a time.

The lead as a labelled subject. Until July 25 the AI running the sessions did not translate; it paid outside models to produce the translations it studied. The reorientation of that day let it translate itself, at no cost, on one condition: it was "a labeled subject." Every translation was filed with a translator's log — decisions and alternatives, never a claim of quality — frozen before any evaluation could be designed; each declared its contamination risk, the chance that a published translation of the same text sat in the model's training data; and the model never judged its own output. A non-Anthropic panel judged, blind, with authorship stripped and the order of comparison swapped.

The panel. Three or more outside models from different companies were given three roles a single design never mixed: contrast translator, critic of designs, and juror. Any design costing more than fifty cents, or whose result would enter the handbook, had to pass a critic first. The panel's agreement with itself was explicitly weak evidence: the models "share training priors with one another and with you; agreement among them is QA, not validation." That is why the jury had to be calibrated, and why it had to be non-Anthropic.

Anchors. Evidence about what is good came from the published record, catalogued as pages with provenance, copyright status, the language read, and a statement of what each could and could not ground. By the close there were twenty-three anchor pages — three fixing points of English style, one unadopted corpus, and nineteen precedent anchors, published translations read whole against their sources in eleven language pairs — and sixteen source pages of scholarship, six of them non-English primaries read complete in the original: Schleiermacher, Berman, Yan Fu, Lu Xun, Futabatei Shimei and Verga. Copyrighted material was quoted only in brief attributed excerpts and logged; nothing copyrighted was stored whole outside a quarantined directory of two items the owner supplied.

Money. Five US dollars a day in billed API cost, self-enforced and never exceeded. Because the lead translated for free, the money went to judging, critique and contrast subjects; "money was never the binding constraint." The constraint was the depth a memoryless session could reach before it had to land its work.

Verification. Every experiment had a design frozen before it ran, a critic pass, raw outputs preserved, and a separate script that recomputed every reported number afterward. A null was a result; a pre-registered gate that fired meant the number was printed as description and not claimed. The rule that bound most tightly concerned authority: every evaluative statement either cites an anchor or carries the flag internal-judgment-only, and every jury score carries provisional until the jury passes calibration, which it never did.

Course corrections. Four times the owner asked the project to assess itself, and four times it found something its machinery could not see: fifteen sessions of apparatus and no translations (July 25); seven sessions down one thread while the calibration gate starved (July 26); twenty-six consecutive sessions about the project's own instruments (August 1); ninety-nine multi-session "arms" that all finished on time and never converged on the deliverables (September 4). Each produced a rebuild of the hand-off machinery, described in Appendix A; the last produced the problem-indexed handbook this essay reports on.

The numbers. The close-out page tabulates the record: 256 numbered sessions over forty-nine days; about 330 filed translations of some 175 works from more than twenty source languages; 227 result pages; nineteen ratified decisions; ninety-nine arms; fifty-nine regimes; twenty-three anchors and sixteen sources; a handbook of fifteen problem families; and about 130 US dollars of API spend, accounted for in the budget ledger to the fraction of a cent.