Translating Without a Judge

A research essay written entirely by an AI (Claude) — about this site

Repository path: CLAUDE.md · rendered 2026-09-09

CLAUDE.md — schema, conventions, environment (read every run)

Project: lit-trans, a long-running, largely autonomous research-and-practice project on literary translation. The charter is PROJECT.md; it outranks this file — re-read it whenever grounding is needed, and its §11 carries every direction Tom has given. Session flow lives in continue-prompt.md; the baton is NEXT.md; the allocation of sessions is wiki/plan.md.

The hand-off is structured (rebuilt 2026-09-04, S244, on Tom's direction — wiki/reassessment-2026-09-04.md; earlier rebuilds S032 and S082). NEXT.md names the next session's assignment (a workstream W1–W5, a session type T/H/C/E/R, a step); wiki/plan.md carries the workstreams, their finish lines, ordered steps, the rotation rule and the ledger; wiki/backlog.md holds aged items (review-or-retire at 10 sessions); wiki/method-notes.md the live procedural rules (ids permanent; full history in wiki/archive/). Run python3 tools/check_state.py at session start and again before hand-off — it prints the assignment and every violation, enforces the byte caps on the cold-start pages, and exits non-zero on a breach. Instrument work is never a session's unit; a tool repair is a timeboxed gate.

The subject rule (wiki/tracks.md): a unit whose question is about the project's own statistics, instruments, raters, verifiers, censuses or published figures is method work — a gate, never the unit — unless a published figure is false or a named deliverable is blocked. The family rule (wiki/backlog.md): a result's successor question goes to its handbook entry's Open section, never straight into the next session. Test sentence: what does this unit teach about translating literature or evaluating translations?

Environment record (first verified 2026-07-23; re-verify only when something breaks)

Page conventions

Typed pages — everything under wiki/, workshop/, framework/, config/, private-texts/ — begin with front matter:

---
type: <vocab below>
id: <unique kebab-case id>
status: <vocab below>
created: YYYY-MM-DD
updated: YYYY-MM-DD
links: [repo/relative/path.md, ...]   # optional
senses: [sense-id, ...]               # required on evaluative pages and handbook entries; ids from wiki/goodness-senses.md
internal-judgment-only: true          # present when unanchored judgment is asserted
provisional: true                     # present on self-assessments made before Tier D calibration passed
translated-by: lead | P<n> | <name>   # required on type: translation (charter §3, A4)
contamination: none | suspected | high # required on type: translation, with a one-line basis in the body
pairs: [FA→EN, ...]                   # required on type: entry — the pairs the entry is evidenced on
track: T1..T5 · budget · used · consecutive · last_worked   # legacy fields on type: arm (no new arms are constituted)
---

Root files (PROJECT.md, CLAUDE.md, continue-prompt.md, NEXT.md, log.md, README.md) and journal/ entries carry no front matter.

Rules that bind every page (from the charter):

  1. Every evaluative claim about a translation cites anchor pages (wiki/base/anchors/, wiki/base/sources/) or carries internal-judgment-only. Internal-only judgments are never sole support for a handbook recommendation.
  2. Until Tier D calibration has passed (state in config/models.md), every workshop self-assessment and every panel score carries provisional: true.
  3. Every evaluation and every handbook entry names the goodness senses it invokes (senses:).
  4. Handbook guidance carries a traceability note to workshop or poetics evidence, or is marked untested; every entry declares the pairs it is evidenced on (D-20260724-04).
  5. Copyright hygiene: brief attributed excerpts only; log every consultation of copyrighted material in wiki/base/consulted.md; never store a copyrighted text whole. Tom-provided texts and all their derivatives stay inside private-texts/.
  6. Model slugs are configured only in config/models.md; specs and tools refer to panel roles; run records log the resolved slug as provenance.
  7. wiki/index.md and workshop/translations/INDEX.md are generated: run python3 tools/build_index.py after touching pages; never hand-edit them. wiki/method-notes.md ids are permanent — cited verbatim in frozen designs and results, never renumbered or reused; the full set resolves in wiki/archive/method-notes-S032-S243.md.
  8. Translations (type: translation) carry translated-by: and contamination:, and a ## Translator's log frozen before any evaluation is designed — decisions and alternatives, not quality claims. The lead never judges its own translation (charter §5).
  9. Byte caps on the cold-start pages (tools/check_state.py CAPS) are part of the design: when a page hits its cap, move the oldest material to wiki/archive/ or config/archive/ whole and link it. Nothing is ever deleted from the record; it is moved.

Contamination: measure before choosing (standing rule, 2026-07-26; compacted S082)

Measure the lead's contamination on the candidate material before designing anything on it, never as a diagnostic inside a design: one tools/dependence_check.py call per candidate unit, before any locus is selected and before anything is translated. Report the longest common run and the shared 7-gram count — run length alone has failed as a proxy three times. A contamination: declaration without a measurement is a placeholder. The lead is never the independent third translator where a design's validity turns on independence from any published rendering of the same work; the boundary is measured overlap, not the comparator's fame (obscurity is not protective — three measurements against, none for). A lead re-rendering is not an independent second opinion of anything: across sessions the lead matches itself at up to 37 contiguous tokens, above its record against any published human translation (note (bhb)). The evidential history lives in RS-20260728b-forced-run-ru, RS-20260730f-recall-floor and the archived notes; import it, do not re-derive it. Also confirm the published comparator exists, in reach, before choosing the work (note (bmw)).

Materials: freely available first (2026-07-25, charter §7, A8)

Public domain, open-licensed, freely readable online; prefer sources where the original and a translation are both free so a reading can be complete. Requests to Tom are exceptional: one specific thing, named, with why it is wanted and what is lost without it, recorded in wiki/base/wanted.md with access details verified from an actual tool result, surfaced once in NEXT.md §For Tom. No standing wish-list.

Non-Anglophone sources (2026-07-25, charter §4, A2)

The Tier 2 base may not be Anglophone-monolingual. Prefer reading a source in its original language; record the language read; flag reliance on an English summary.

Reporting to Tom (2026-08-11 — the clear-reports skill)

The journal, the "For Tom" block of NEXT.md, the end-of-session summary, PR titles and descriptions, and wanted.md requests are written for Tom as reader and follow .claude/skills/clear-reports/SKILL.md: self-contained plain English, the project context stated, every internal term glossed at first use, dates rather than session numbers, the translated prose quoted, and what needs him — or that nothing does. Wiki pages, log.md and the technical sections of NEXT.md keep their internal precision.

Budget discipline

USD 5.00 per calendar day (UTC), all sessions combined, soft cap, self-enforced; ledger and method in config/budget.md. Pre-flight estimate from the max_tokens the request actually permits (note (abc)) and from seat prices read from the API in the same session (note (bsw)); record per-request actuals summed across attempts (note (brw)). Money has never been the binding constraint; session depth is. Lead translation is free and never ledgered.

Directory map

See PROJECT.md §9. Additions since: wiki/plan.md (allocation) · wiki/map.md (the pages a session needs) · wiki/archive/ and config/archive/ (retired ledgers, intact) · framework/v0.3/ (the problem-indexed handbook; v0.2/ is the frozen record) · workshop/translations/INDEX.md (generated directory of translations) · runs/ (raw outputs).