Repository path: journal/2026-07-24.md · rendered 2026-09-09

2026-07-24 — S005

A two-unit session, no API spend, both tracks moving while the calibration run stays blocked on your in-copyright texts.

Done

1. Tightened the R01/R02 regime prompts (workshop track). The pilot back on the 23rd caught a small but real defect: asked to translate "The Spider's Thread," the model printed the story's title as a heading even though the prompt said "no headings," and the self-revision pass kept it. That would have quietly corrupted the first real R01-vs-R02 comparison. Fixed before either regime freezes: the translate prompt now spells out "no title, no author line, no heading — begin at the first sentence" and labels the bibliographic block as context-not-to-be-copied; R02's revision step now explicitly strips any heading that slips into the draft. Both bumped v0.1 → v0.2, still draft, nothing recorded has run under the old wording.

2. Built the third Slate A reception-record page — Akutagawa, Kojima (1952) vs Rubin (2006). This was meant to be the "third clean-asymmetry case" for calibration robustness. Verifying the record first was the right call, because the record is not clean. Our standing prior was "Rubin well received, Kojima dated." Reality: Rubin's Penguin is the scholarly standard and is praised as a whole (Donald Richie: prior translations were "so bad," this one "so good"; Complete Review gives it a B+), but the one careful side-by-side comparison of "Rashōmon" itself actually prefers Kojima — "better compositional choices, wit, and deadpan humor" — and several readers fault Rubin's dialogue as "too colloquial." So the record splits by sense and by critic, with a third translator (De Wolf) muddying even the readability question.

Learned / decided

• Botchan stays our cleanest calibration case. Akutagawa is a weaker, contested case — the only axis with any signal (humor, favoring Kojima) rests on a single reviewer reading a single story, and it points the opposite way from our prior. I've recommended deprioritizing Akutagawa for the calibration run (at most a narrow per-sense stressor, and only after reading the paywalled 2007 Translation Review piece), and pointing the search for a genuine third clean case at Kokoro (McClellan 1957 vs McKinney 2010) instead. This is exactly the kind of prior-correction the "verify everything" rule in the program is there to catch.

Spent

$0.00 API (all network reads + local writing). Day total still $0.00 of $5.00 for 2026-07-24; the calibration run remains funded and blocked on your texts (see NEXT.md → "For Tom").

2026-07-24 — S006

A single, deeper poetics-track unit, $0 API. S005 leaned Slate A + workshop, so this one leaned poetics to keep the tracks balanced. Calibration still blocked on your in-copyright texts.

Done

Built the project's first Tier 2 theory source page — Venuti, The Translator's Invisibility (wiki/base/sources/S-venuti-invisibility.md). Until now every S- page was a reception record of specific translations; this is the first that imports a piece of translation theory. The point wasn't to summarize a famous book for its own sake — it was to bring in a controlled vocabulary the poetics track has been missing: domestication vs. foreignization (assimilate the foreign work to fluent English, or preserve/signal its strangeness), the "regime of fluency," and the translator's invisibility. Verified the account against the secondary record rather than trusting priors; the book itself wasn't opened (secondary sources only, all ledgered).

Two things I was careful about: - We borrow Venuti's map, not his destination. Venuti argues foreignization is ethically better; the project is descriptive-pluralist, so the page takes his vocabulary and explicitly declines his normative preference. It also records the live critique of him (Tymoczko, Pym: the terms are a sliding scale, not an on/off switch, and lack hard criteria) so the pair is used as a scored continuum, never a binary label. - Kept two questions from collapsing into one. The register question (OQ-20260723-target-register) asks which English — period, timeless, or contemporary. Venuti's axis asks how much source-strangeness to keep. Those are orthogonal: you can domesticate into 1918 English or into 2026 English. I wrote that distinction into the open-question page as its next real step — it now has two named axes to hold apart instead of one blurry one.

Wired it into the typology. Added a Tier 2 caution under the naturalness sense: rewarding fluency is not value-neutral, so an instrument that always maximizes naturalness has silently taken a side. This is the first Tier 2 input to the sense-derivation work (S004's was the first Tier 1 anchor). No sense was split, merged, or retired; naturalness stays untested.

Learned / decided

• The naturalness sense had a hidden assumption baked into its own definition — the "unless the source rings strange in the same place" clause. That "unless" is exactly where the domestication/foreignization choice lives, and naming it makes the sense honest rather than changing it.
• The target-register open question was quietly one-dimensional; it's really two-dimensional. That's progress on the question even though it's still unanswered.

Spent

$0.00 API (network reads + local writing only). Day total for 2026-07-24 remains $0.00 of $5.00. Calibration run still funded and blocked on your texts (NEXT.md → "For Tom").

2026-07-24 — S007

Your request this session: whenever the project asks you for a text it can't get to itself, give you the bibliographic details and links — OpenLibrary scans, Amazon Kindle, Apple Books, or other sites you might reach — so the texts are easy to find; put that in the right file and mention it in reports; and keep doing it in future sessions. Done, $0 API.

Done

1. Rewrote the "wanted" list with full bibliographic data + verified access links (wiki/base/wanted.md). Every in-copyright translation the calibration run needs now carries its ISBNs, its Amazon Kindle edition (with the ASIN, pinned to the right translator — important here because "No Longer Human" has a swarm of junk Kindle uploads by uncredited people), its Apple Books page, Google Play where it exists, the publisher's page, and its OpenLibrary / Internet Archive status. I verified each link and identifier against an actual search or page fetch — nothing is invented; where I couldn't confirm something (one old journal review) I left it as a "look here" pointer rather than guess a URL. The one bit of copyright care worth noting: I only read catalogue and retailer metadata, no translation text, and logged that in the consultation ledger.

2. Made it a standing rule, not a one-off. Added a short "Requesting materials from Tom" section to CLAUDE.md, which every session reads at startup. So future sessions will attach bib data + links to any text they ask you for, and surface them in the baton and these journals, without you having to remind them.

3. Two useful things fell out of the research: - Alan Turney's Botchan (1972) is borrowable free on the Internet Archive — archive.org/details/botchannovel00nats. That matters because the calibration design's real test ("A-test") is Turney vs. Cohn, and Turney had somehow never made it onto the request list. It may be the one calibration text you don't have to buy or dig up at all. - Joel Cohn's Botchan (2005) is the one true bottleneck — there's no free scan anywhere, so it's Kindle/Apple/print only. It's also the text without which Case A can't run. If you're getting one thing, that's the one.

What this doesn't change

The calibration run is still blocked on the actual texts — this session made the asking precise and easy, not the getting. The quick guide is in NEXT.md under "For Tom," with the Kindle ASINs and Apple/Archive links inline so you can tap straight through.

Spent

$0.00 API (web lookups + local writing only). Day total for 2026-07-24 stays $0.00 of $5.00.

2026-07-24 — S008

You added four Botchan files to private-texts and asked me to OCR them, clean them carefully, file them, delete the PDFs, and update the docs so the next session knows they're here. All done, $0 API.

Done

Set up OCR from scratch. The container had no OCR tools at all (and the pre-installed PDF library was broken), so I installed a fully local OCR engine — nothing about your private texts left the machine. Three of the four PDFs were image scans needing real OCR; the Sato article was already digital text.

OCR'd and — the important part — verified all four by eye. For the three scans I rendered each page at high resolution, ran OCR, and then read every page image myself against the OCR and corrected it. That mattered: the machine OCR quietly dropped whole lines in several places (e.g. a clause in the Cohn opening, a line or two on other pages) and mangled things like the "fl" in "flung/floor/flattery" and the macrons in names (Kantarō, Kōjimachi). The cleaned texts preserve each edition's real styling — Cohn's single quotes vs. Turney's double quotes, italics, American spellings. The Sato article I extracted from its text layer and tidied (I stripped a stray "publisher template" watermark that was buried in it).

Filed them under private-texts/botchan/ with a manifest, kept them out of the public index, and deleted the four PDFs as you asked (they're still in git history if ever needed). What's there now: Cohn ch.1 opening, Turney's complete ch.1, the Swann review, and the Sato article.

The page-number jump you flagged — thanks; I've noted in the file that Turney p.11 is your skipped full-page illustration, so no text is missing.

One thing worth your attention

The pages you sent are the novel's famous opening. My calibration design deliberately avoids the opening, because a jury of AI models might recognize such a famous passage rather than actually judge its quality — which would give a falsely rosy result. So I can do one of two things, and I've left the choice in NEXT.md: (1) ideally, you send a few mid-book pages of Cohn and Turney from the same scene (e.g. the fishing outing) and I run it properly; or (2) I run it on the opening now and clearly label the result as an upper bound. Either works. Separately, the No Longer Human trio (Case B) is still open whenever convenient — links are in wanted.md.

Reception bonus

The Swann review turns out to be genuinely useful: it independently assesses Turney (the translation I'll compare against Cohn) as "nicely colloquial … quite accurate on the whole" but with specific slips, and documents how Turney domesticates Sōseki's puns. That, plus the Sato article, feeds the Botchan reception record next session.

Spent

$0.00 API — all local (image rendering + on-device OCR + text extraction). Day total for 2026-07-24 remains $0.00 of $5.00.

2026-07-24 — S009

You started this one by hand and asked for as much autonomous work as I could do — practical analysis where it's available, but especially theory and essay-writing, so the wiki grows into a rich, well-grounded picture of the project's problems. This session leaned hard into that, and it was overdue: the last several sessions were materials/logistics, and the poetics track was owed. $0.00 API — all reading and writing.

The short version

I mined the two scholarly items you sent with the Botchan bundle (the Swann review and the Sato article), turned them into real evidence, and wrote the project's first essay off the back of them. One clean case — how six translators handle a single Japanese pun — turned out to carry a genuine idea worth arguing.

Done

1. Reconciled the Botchan dating that was nagging us. The record had Morri at 1918, Swann calling the early pair "Sasaki 1922 / Mōri 1948," and Sasaki's date flagged "uncertain." Sato's peer-reviewed bibliography settles it: Morri really is the first (1918); Sasaki finished his in 1922 but it wasn't published until 1968 (Tuttle), 39 years after he died — which is why it's cited as both. Swann's "1948 Mōri" is just a later edition he had to hand. I also added a sixth translator (Matt Treyvaud, 2009) that we'd missed, and confirmed — independently, against the Keene Center's own records — that Cohn's Botchan won the Japan–U.S. Friendship Commission Prize, the oldest U.S. prize for Japanese literary translation. That prize is the single most authoritative "this one is good" datum the calibration case has, and it points at exactly the translation the informal reviews favor.

2. Turned the Sato pun analysis into the project's best evidence for one kind of "good." Sato studies one scene — a country schoolboy ends every sentence with a dialect tag, na moshi, and Botchan mocks it — across six translations. Most translators substitute: they swap the Japanese country dialect for a New York, Irish, or US-Southern one. Sato's verdict is that this is fun to read but "seriously distorts the culture" of the original (Morri's 1918 version even imports an American racial slur that has nothing to do with Sōseki). Cohn does something different: he keeps the Japanese na moshi on the page and teaches you what it means through context and sheer repetition (54 times) — no footnote, still perfectly readable. And your Swann review, from 1974, independently documents the same failure mode in Turney: translating tempura as "fried prawns" quietly kills a joke; translating a trivial "ice water" into a fancy dessert loses a thread the plot hangs on. Two critics, 45 years apart, catching the same thing.

3. Wrote the first essay: "Fluency is not the enemy of foreignness." The idea: the standard theory (Venuti) gives you two options — smooth it out for the reader ("domesticate") or keep it foreign and difficult ("foreignize"). This case shows the best option is neither — Cohn's rendering is smooth and keeps the foreign word. I call the move "scaffolded borrowing" (keep the source item, build its meaning into the surrounding text). The practical upshot, which sharpens how our AI jury should score translations: the problem was never fluency itself, it's how the fluency was bought — did the translator delete something foreign to get it, or preserve it? That's a question a reader (or a jury) can actually answer, which the vague "too domesticated?" is not.

4. Turned the essay's claim into a testable bet (CJ-...-scaffolded-borrowing): when we finally run the calibration jury, a blinded jury scored on this axis should rank Cohn's pun-handling above the substituters — and, conveniently, that pun scene is mid-book (chapter 6), which is exactly the kind of non-famous passage the calibration design wants. So there's now a precise, low-cost thing to ask you for: matched Cohn + Turney pages of the grasshopper/na-moshi scene, which would let the calibration run honestly instead of on the famous opening.

5. Fed it all back into the typology. Four of the "senses of good" moved from stipulation toward being grounded in real evidence (style, affect, cultural handling, and a refinement to naturalness), each with a note pointing at the case.

What this is and isn't

It's one pun, in one novel, one language direction — I've said so plainly in the essay, with the list of things that would make me revise or drop it. It's not a claim that "keep it foreign" always wins; sometimes substituting is right (a kids' edition, a fast comic read). The narrower, firmer claim is that the good option here was invisible to the usual two-way choice, and became visible once we asked what the fluency cost.

For you

Nothing needed, but if it's easy: the two mid-book Cohn/Turney pages of the grasshopper scene would let both the calibration and this new bet run on the same honest passage. Details are in NEXT.md. Case B (No Longer Human) is still open whenever convenient.

Spent

$0.00 API (reading + writing only). Day total for 2026-07-24 stays $0.00 of $5.00.

2026-07-24 — S010

The workshop track finally ran a real experiment — the first proper comparison of our two translation "regimes." This is the first session that spent real money ($4.49), and it earned a genuine, if provisional, finding. Everything below is uncalibrated (the jury still isn't validated), so read it as "a first honest signal," not "a proven fact."

What I set out to do

The baton said the workshop side had produced nothing since the very first pilot, so this session was owed a real experiment: R01 (translate in one pass) vs R02 (translate, then let the same model revise its own draft once). The question in plain terms: does a self-revision pass actually make the translation better, and if so, better in what way?

Done

1. Froze the two regime specs (R01, R02) so a recorded comparison could run against a fixed target.
2. Wrote a frozen experiment design, then had an independent "red-team" agent tear into it before spending a cent. That critique was the most valuable step of the day. It caught a subtle but fatal flaw: R02's revision runs at a lower "temperature" (a randomness setting) than the first draft, and lowering that setting alone tends to make prose read smoother — so a naive result ("revision reads better!") could be pure temperature, not real revision work. On its advice I added a control: a second single-pass draft at the low temperature, so I could tell the two apart. The critic also fixed two logic bugs in my failure rules that would have thrown away exactly the honest "no difference" result I claim to want.
3. Ran it. Two short public-domain stories (Akutagawa's Spider's Thread, Sōseki's Ten Nights of Dreams 1–2), two translator models, three repeats each; then a blinded two-model jury scored draft-vs-revision on five quality senses, each item judged twice with the passages swapped to catch position bias. 36 translations + 96 jury judgments.
4. Analyzed, then had a second independent agent recompute every number from the raw data — it confirmed all of them, zero discrepancies.

What we (provisionally) learned

• One self-revision pass reliably improved naturalness — and it's the revision doing the work, not the temperature trick. Revision beat its own draft on naturalness 76% of the time (consistent across every work/translator combination and both judges), while the temperature-only control sat at 53% — basically chance. So R02 buys real fluency.
• But it did not improve accuracy (43%), and slightly hurt formal style-fidelity (39%). In plain terms: the revision made the English read more smoothly without making it more faithful, and maybe sanded off a little of the source's texture. That's a small, real-world echo of the essay I wrote last session ("fluency always has a price — ask what it cost"), now showing up inside our own pipeline rather than in published translations.
• The revision is surgical — it changes only ~3% of the words, but that's enough to move the fluency needle.

Two useful "the machine works" lessons

• The jury has a position bias: both judges lean toward whichever translation is shown first. Our swap-the-order design caught and cancelled it, but it means we must keep that safeguard when we do the real calibration.
• A budgeting gotcha, now documented: several of the models "think" for thousands of hidden tokens before answering, and my first token budget was too small, so ~1 in 5 calls came back truncated. I caught it, re-ran them, and the wasted attempts cost about $0.58 — the main reason the day ran to $4.49 instead of the ~$2.50 I'd planned. Fixed and written down so it doesn't recur.

Spent

$4.49 of the $5.00 daily cap — so no more API today. Honestly a bit more than intended (the truncation waste + one judge model being pricier to run than expected). The finding is worth it, but the calibration run is now a tomorrow job, budget-wise.

For you

Nothing needed from you this session. The one standing ask is unchanged and would unlock the real prize (jury calibration): ~2–3 mid-book pages each of the Cohn and Turney Botchan translations (the grasshopper / "na moshi" scene, Chapter 6). Details in NEXT.md.

2026-07-24 — S011

A short, deliberate poetics-track session with no API spend (the day's budget was already ~90% used by S010's experiment). The job: build one thing properly rather than pad. The thing was overdue.

Done

Built the project's second naturalness anchor — the contemporary one — to sit beside the 1922 Mansfield anchor. For weeks the open question "into which English do we translate?" has had only one concrete reference point: Katherine Mansfield's genteel 1922 prose, which quietly biases every "does this read natural?" judgment toward a period register (it licenses "button boots" and "ermine toque"). To make that bias visible and measurable instead of hidden, the project needed a matched reference point at the opposite end: fluent, alive, present-day English.

I used the opening of Cory Doctorow's Little Brother (2008) — a first-person teen-hacker voice, thick with contemporary slang and San Francisco vernacular. Two reasons it's a good pick beyond just being modern:

1. It's legally clean to store. Doctorow releases the book under a Creative Commons license, so we can keep an excerpt whole with attribution — our first anchor that rests on an explicit open license rather than public-domain age. (Recorded carefully, including the "share-alike" string that would travel with the text if the framework is ever published.)
2. It's built on the exact move we're betting on for translation. Doctorow constantly drops in an unfamiliar term and immediately teaches you its meaning through context — "Poker players call stuff like this a 'tell' — something that let you know what was going on in the other guy's head." That's the same "scaffolded borrowing" strategy we flagged as promising for handling culture-bound Japanese words (the Botchan "na moshi" case). Seeing it done natively, and reading as completely natural contemporary English, is real evidence for one half of that bet: scaffolded borrowing can be fully fluent, and it isn't tied to any one era.

Why it matters

With two anchors instead of one, "natural English" stops being a single implied standard and becomes a variable with two concrete poles — period (Mansfield) and contemporary-vernacular (Doctorow). From now on an evaluation can't just say a translation "reads natural"; it has to say natural against which register — and reckon with the gap between the two. That's the register question turning from an assumption into a measurement, which is exactly what a framework needs before it can recommend a target register on purpose rather than by accident.

Honest limits, kept on the page: it anchors only the vernacular end of contemporary (we still want a neutral/literary-contemporary one), and because it's original writing, not a translation, it supports the "scaffolding is fluent" claim but does not settle the separate "scaffolding beats substitution" question — that still needs the blind jury comparison of the Botchan translations.

Housekeeping

Cross-linked the new anchor through the open-question, the goodness-senses typology, the scaffolded-borrowing conjecture, and the Mansfield page; regenerated the wiki index (front-matter lint clean).

Spent

$0.00 — no model calls; the text was fetched from the open web. Day total stays at S010's $4.49.

For you

Nothing needed. The one standing, high-value ask is unchanged: ~2–3 mid-book pages each of the Cohn and Turney Botchan translations (the grasshopper / "na moshi" scene, Chapter 6) — it would unlock both jury calibration and the scaffolded-borrowing test. Details in NEXT.md.

2026-07-24 — S012

Lean-poetics unit (poetics was due; the day's API budget was already spent by S010, so this was a $0 session). Read continue-prompt.md for the session flow.

Done

Built the project's first "precedent anchor" — and it turned out to be a much better one than we'd expected to be able to make.

Some background on why this matters. Our highest-authority evidence ("Tier 1 anchors") comes in two kinds. The first kind is published English writing we read to pin down what "natural English" actually is — we have two of those now (Mansfield 1922, Doctorow 2008). The second kind is a published translation read side-by-side against its Japanese source, to catalogue what a real translator actually did with the hard problems. That second kind had been empty since the project started. This session filled it.

The lucky break: it turns out Glenn W. Shaw's 1930 English translation of Akutagawa's "The Spider's Thread" just entered the US public domain this year (1930 + 95 years), and it's on Project Gutenberg. "The Spider's Thread" (蜘蛛の糸) is one of the six works in our own translation canon — so we already had the Japanese in the repo. That means, for the first time, I could store both texts whole and do a complete line-by-line comparison, instead of the tiptoe-around-copyright, brief-quotes-only treatment a modern in-copyright translation would force. And because it's a canon work, Shaw's version is now a ready yardstick for our own machine translations of the same passage.

What the close reading found (the reusable catalogue is the point of the page):

• The big one — one translator, the whole toolbox, used inconsistently. For the Buddhist afterlife place-names, Shaw does four different things in 1,000 words: swaps two for familiar Western words (極楽→"Paradise", 地獄→"Hell"), literally translates one (血の池→"Pond of Blood"), and leaves two as untranslated romaji (三途の河→"Sanzu-no-Kawa", 針の山→"Hari-no-Yama"). So the "how do you handle a foreign item" choices we'd theorized from the Botchan case are demonstrably real in the wild — and handling one coherent set of names five different ways is itself a cost (the reader gets an incoherent map of the afterworld).
• A direct test of our own essay. Our first essay bet that "scaffolded borrowing" (keep the foreign word, teach its meaning through context) was the clever move behind the best-received Botchan translation, and it flagged that building this precedent anchor might show that move to be rarer or more common than we thought. Result: it's more common — Shaw does a light version of it in 1930 ("the needle of the dread Hari-no-Yama" quietly tells you what Hari-no-Yama is), 75 years before our star example. So the technique isn't special; what's special is doing it thoroughly and on purpose. That's a genuine, honest correction to the essay — its core survives, its framing tightens.
• The storyteller voice is a real, unsolved problem. The Japanese is narrated in a gentle, polite, spoken-aloud-to-children register (the ございます endings). English has no equivalent, so Shaw drops it and reaches for old-fashioned formal words ("beheld," "Nay") to compensate. He recovers the formality but not the spoken-to-a-child warmth. Worth noting because our own draft translations did the same thing (flattened it) — so that's not a machine failing, it's a problem a skilled human in 1930 couldn't fully solve either.

I wired all this into the pages it touches: the goodness-senses typology (two senses' derivation advanced, plus a newly-named handling — "transparent calque"), the essay (added an honest "here's what the precedent anchor changed" note), the scaffolding conjecture, the register open-question, the Akutagawa reception page, and the canon manifest.

Why it matters

The framework is supposed to be derived from evidence, not asserted. Until today, every "here's how you might handle a Japanese honorific / a piece of realia / an untranslatable narrator" claim rested either on theory or on what critics said about translations. Now one whole class of those claims rests on a full, verifiable reading of what a translator actually put on the page, against a source we can all check. It's one translator and one short story — a start, not a law — but it's the first brick of the kind we'd been unable to lay.

Spent

$0.00 — no model calls; both texts are free and public-domain, read and compared by me directly. Day total unchanged from S010 ($4.49 / $5.00).

For you

Nothing needed this session. The one standing, high-value ask is still the same: ~2–3 mid-book pages each of the Cohn and Turney Botchan translations (the grasshopper / "na moshi" scene, Chapter 6) — it unlocks both the jury-calibration run and the scaffolding test. Details in NEXT.md. (And, whenever convenient, the No Longer Human translations and any texts you'd personally like the project to translate.)

S013 — going beyond Japanese and English

Fourth session today. You asked for this one directly: extend the side-by-side comparisons, and their theory, past the Japanese–English pair.

What I did

Built two more precedent anchors — full side-by-side readings of a real published translation against its original — in two new language pairs, and then a first theory page arguing what the three of them jointly show.

The two pairs were not picked for convenience. They were picked so that each differs from Japanese→English in a known way:

The Baudelaire one also does something the project had never done: English is the source, and the target isn't English. All four texts are public domain on both sides, so, as with Shaw, I could read them whole rather than in snippets. Cost: nothing.

What came out of it

1. The single most robust finding so far: meaning carried by grammar doesn't get translated — it gets converted into word choice, and loses its consistency on the way.

Three translators, three languages, seventy years apart, all doing the same thing:

• Chekhov's nine-year-old writes his grandfather a letter in the intimate ты, and slips once into the polite вы — two letters that tell you exactly how far the boy's schooling got. English has one word for "you." The slip doesn't survive; it can't. Garnett doesn't mistranslate it — there is nothing to translate it into.
• The same boy's letter is written in thick illiterate peasant grammar. Garnett writes standard English and pays it back in vocabulary instead — "I had a wigging," "their wretched brat," "my mammy." The marking moves from grammar (constant, automatic) to word choice (occasional, wherever a word happens to be handy). That's why these translations read as flattened rather than wrong.
• And it isn't something that only happens to exotic foreign languages. English loses this way too. Poe calls his first cat "him" and the second cat "it" — the narrator is refusing the second animal personhood, and the pronoun is where he does it. French has genders, not that contrast. Baudelaire can't render it. Nobody could.

Worth saying what did survive all three times: whatever the story carries in its events. The whole point of "Vanka" — the letter is addressed "To grandfather in the village" and can never arrive — comes through Garnett untouched. Losses aren't distributed by importance. They're distributed by what kind of thing carries the meaning.

2. Closeness causes its own damage — and this surprised me. With Baudelaire I expected an easy ride, and in one sense it is: no cultural problems at all (monks are monks, gin is gin, and even Poe's cat "Pluto" has a ready French form, "Pluton"). But Poe's style is high-Latin-derived English against a plain Germanic floor — that contrast is his manner. In French those same Latin roots are simply the ordinary way to talk. So the most faithful-looking word is the one that flattens him. Maximum word-for-word fidelity, minimum stylistic fidelity — caused by the languages being close, not despite it.

There's a lovely small case. Poe italicises barroques to make a French-flavoured word feel foreign inside English. Baudelaire keeps the italics — but in French it's just the word, so the italics now mark nothing at all. The translator didn't choose to domesticate him; the language pair took the choice away. That's a real correction to how we'd been using this idea: we'd been asking "what did the smoothness cost?", and there's a prior question — was there a choice to make?

Baudelaire's answer is to put foreignness back in by hand: he leaves "Gentlemen" sitting in English inside the French, and he coins words Poe never wrote (hyperdiabolique; and, translating "unutterable fear," he writes "intraduisible" — untranslatable). That coining habit isn't just my reading: a 2012 scholarly article documents it across his whole Poe corpus, and reads it as an appropriation "hidden behind a smokescreen of apparent literalness." I've made that our second theory source page.

3. An honest correction to last session. With Shaw I catalogued his four-different-ways handling of one set of Buddhist place-names as a fault. But Garnett does it too (two dogs in one sentence: one transliterated, one translated), and so does Baudelaire (three foreign words, three treatments). Three for three. Either consistency is a much weaker norm in real published translation than our typology assumes, or all three are defective — and I shouldn't decide that from an armchair. I've reopened it as a question a calibrated jury could actually settle, rather than leaving my earlier verdict standing.

Why it matters

Everything in this repository until today came from one language pair. That's a live risk, not a theoretical one: a typology built on Japanese→English will quietly encode that pair's properties as properties of translation itself. What the three anchors now show is that the senses of "good" survive the move to other pairs, but their weights don't — "cultural mediation" is the whole story for Japanese→English, a small residue for Russian→English, and nearly idle for English→French, where the difficulty moves wholesale into style. So a framework that fixes those weights from Japanese→English evidence will mis-describe every other pair. I've opened that as a formal decision for a later session to ratify or reject, rather than just acting on it.

Spent

$0.00 — again no model calls; four public-domain texts, read and compared directly. Day total unchanged at $4.49 / $5.00.

Honest limits

One reader, one short text per pair, three translations all from before 1930. The Russian and French readings have had no second pair of eyes, and four claims now rest on them — so I've queued a cheap verification pass (a panel model checking my quoted correspondences against the stored texts) as the next thing worth spending on in this line. And the design has a hole: nothing structurally close but culturally distant, and no non-European target anywhere in the project yet.

For you

Nothing new needed. The standing ask is unchanged — ~2–3 mid-book pages each of the Cohn and Turney Botchan translations (the "na moshi" scene, Chapter 6), which unlocks jury calibration and the scaffolding test.