Repository path: wiki/close-out.md · rendered 2026-09-09

Close-out — where the project ended (2026-09-09, S257)

Occasion. Tom wound the project up on 2026-09-09, verbatim: "I want to wind up this lit-trans project. I have turned off the Routine, so that won't fire anymore." He asked for the loose ends of research and framework development to be tied off, for a plain-language essay on what was learned (published with an HTML mirror of this record at docs/, deployed by GitHub Pages), and for the practical framework distilled into a document a person can hand to an AI agent together with a literary work. This page records the state every workstream was left in, what was finished at the close, what was left open, and what a successor would need to know. It is the last page written under the charter. The scheduled Routine is off; no session follows this one.

1. The record, by the numbers (2026-07-23 → 2026-09-09)

2. The state of each workstream at the close

W1 — the handbook (deliverable 1). Five families were written in the ordinary way, each applied to a fresh passage with a followability log, and set active: A change of language (S244), G footing and address (S246), I realia (S249), H register (S252), K telling the reader (S255). The nine remaining families (F mimetics, E sound in prose, J emphasis and typography, L long works, M evaluating, B rhymed prose, C radif and refrain, D meter and line, P the pipeline) were consolidated at the close, in this session, by parallel agents following framework/v0.3/README.md §Procedure, each entry then independently verified against its cited sources and repaired. They carry status: draft, because the procedure's step 6 — a fresh 300–800-word passage translated under the entry with a followability log — was not done; their Standing paragraph says so, and their §5 Application reads "not applied". Family Z (pair coverage, changelog, closing statement) is written in framework/v0.3/README.md. The finish line's remaining condition — the handbook applied end to end to one whole prose work (W3 step 6) — was not reached.

W2 — the jury (deliverable 2). Tier D calibration is EXHAUSTED, permanently (S250, RS-20260906-tierD-verdict-v3; config/models.md); every panel score in the repository is and stays provisional. Step 4 (judging filed translations on provisional labels) ran one installment (S253, RS-20260907-panel-judging-2); installment 2 was next and was not run. Step 5, the regime comparison on prose (R06 vs R04 vs R03), was never run, and R03 (the multi-agent regime reserved since run one) was never specified. HB-evaluating (family M) records the exhaustion and the honest guidance that follows from it.

W3 — practice (deliverable 3). The Persian/change-of-language thread closed at step 4 (S254). Step 5 chose the long prose work — Laza Lazarević's «Вертер» (1881, Serbian, 14,293 words), with six spans and a cadence declared in workshop/translations/verter/plan.md — at S256, the last Routine session. Step 6 (span A) was never begun: no word of «Вертер» was translated. The plan page stays as the record of the selection; its status is set to retired at the close. (A correction to that page and to NEXT.md: «Вертер» would not have been "the project's first work in a South Slavic language" — Lazarević's «Први пут с оцем на јутрење» (S162) and Aleko Konstantinov's «Бай Ганьо в операта» (S168) preceded it.) Step 7 was not reached.

W4 (evidence) and W5 (typology) were on demand and had nothing pending. No decision is open. The backlog is empty. No request to Tom is outstanding.

3. What was finished at the close, in this session

1. The nine draft handbook entries (§2) and family Z.
2. This page; NEXT.md rewritten as a closed baton; wiki/plan.md's ledger marked closed; the verter plan retired; one line in log.md; the journal entry for the day.
3. tools/build_site.py, which renders every releasable markdown file of the repository to docs/wiki/ as an HTML mirror (1,500+ pages; private-texts/, scripts and raw run outputs are excluded per charter §7), and builds the essay and the framework pages from site/.
4. The essay Translating Without a Judge (site/essay/, docs/), about 14,000 words in fifteen sections plus an appendix on method and a glossary, written for general readers, linking to the mirror. Checked mechanically at the close: every quotation in double quotes verified verbatim against the record, every number located in the record; the planned adversarial fact-check by independent agents did not run (the session's usage budget ran out), so the essay's claim-by-claim support rests on the digests it was written from and on the author's own reading.
5. The framework (site/framework.md → docs/framework.md and docs/framework.html): the project's guidance distilled into a document an AI agent can follow with a user — a structured consultation about purpose, senses, problems and regime, then the translation itself, with the translator's log and the register. Trial before release (S257): four simulated users, each played by an agent with a persona (a book-club reader with Chekhov's «Злоумышленник»; a graduate student with Hafez ghazal 4; a parent reading Miyazawa's 「注文の多い料理店」 aloud; a reader wanting Lazarević's «Вертер» in installments for his children), each talking to a fresh agent following the document. All four consultations reached a confirmed brief; the agents followed the steps, refused to grade themselves, and handled questions the document had not anticipated. The dominant defect was length — agent messages of 1,400–2,800 words, remarked on by the users — and the document was revised on it (standing rule 10, proportion) and on the users' other asks (serial-work register as a diffable file; a second human reader; publication and copyright; the meter question; spans cut at seams). The translation phase and the planned judge pass did not run (usage budget exhausted), so Phase 2 is untested. Transcripts: wiki/archive/framework-trial-transcripts-S257.md.

4. What is left open

• The handbook's end-to-end application to a whole prose work (W1 finish line; W3 step 6–7).
• Nine entries at draft for want of a translation limb; the pair-coverage table shows how thin the coverage is on most pairs.
• Every quality question. The project can say what translators do and what a choice costs; it cannot say which translation is better, and the framework says so on every page.
• The regime comparison on prose (W2 step 5) and the multi-agent regime R03.
• The successor questions listed under Open in every entry's §6.
• The reception-record problem: affect has two documented reception records in the whole project; a third was never found.

5. For a successor

Start with PROJECT.md (the charter, with every direction from Tom in §11), then this page, then framework/v0.3/README.md and its fifteen entries, then wiki/goodness-senses.md. The three course corrections (wiki/reorientation-2026-07-25.md, wiki/reassessment-2026-08-01.md, wiki/reassessment-2026-09-04.md) explain why the machinery has the shape it has and what it failed to see. The contamination rule (CLAUDE.md §Contamination) and the standing method notes (wiki/method-notes.md) are the lessons a successor would otherwise pay for again. The published essay and framework are at docs/; the essay's Appendix A describes the session mechanics for a reader outside the project.