Repository path: framework/v0.3/entries/HB-telling-the-reader.md · rendered 2026-09-09

What a translator prints about the translation — refusals, promises, comparisons, and disclosures — and what each one is worth

Standing. Three internal preference/detection experiments on al-Hariri's Maqāmāt and Sa'di's Gulistan (three blind non-Anthropic seats each; Tier D is NOT PASSED and now EXHAUSTED, so every preference figure is provisional and internal-judgment-only); one page-level measurement of two eighteenth/nineteenth-century translators' printed refusals against their own prose (mechanical counts, not model-judged); one whole-book anchor reading (A-leaf-hafez, Walter Leaf's Versions from Hafiz, 1898, checking an arithmetic self-report against the Persian); one whole-book anchor reading (A-hariri-hands plus Dante's Vita Nuova across four English hands, checking an author's own printed policy); and this session's translation limb (Miyazawa Kenji, JA→EN, one hand, internal-judgment-only). Consolidated 2026-09-08 (S255) from v0.2 §7.34, §7.37, §7.38, §7.39, §7.43.2 and §7.50.6.

1. The problem

A translator can tell a reader something about the translation itself, outside the translated text proper: a preface line refusing to attempt a device, a promise to preserve one, a footnote reporting that one is present, a comparative claim about how often the translator does something relative to the author, or a plain statement of what the source does at a difficult point. A title can carry the same kind of disclosure built into its own wording. None of these behaves the same way. A printed refusal turns out to be checkable and to predict what a reader actually loses; a printed promise does not; a comparative numerical self-report can be checked and can hold; telling a reader a fact about the source moves what they prefer, while showing them the source itself does not do the same work and instead invites them to check the translator's word choices; and an author's own printed policy, carried into the target language, does not bind the people translating the book and may be broken on the very page that reproduces it. A translator deciding whether to write a note, and what kind of statement to put in it, needs to know which of these kinds of printed statement is worth trusting — because they are not interchangeable, and only some of them have ever been checked against what is actually on the page.

2. What published translators do

The regularities:

1. A printed refusal is the cheapest kind of statement to trust, and it is trustworthy because it is checkable at zero cost: a translator who says he will not do a thing can be checked by counting whether he does it, against his own vocabulary's own chance rate. Three refusals in this project's record, three confirmations (Preston on flippancy, Chappelow on pedantry, Knatchbull 1818 on the Arabic's sententious brevity, A-knatchbull-kalila §3.1, 0 of 18 matched shapes). But "printed" and "kept" are still two different acts: of two hands who printed the identical refusal of the identical author's rhyme, one keeps it three times over chance and the other sits at chance across the board (§7.39). A refusal predicts a direction, not a degree.
2. A printed promise is not checkable the same way, and where it has been checked it has failed at every enumerated locus but shown, separately, that it is not empty. Eastwick's promise to keep Sa'di's wordplay produced nothing a blind reader could find inside the study's own enumerated set — the same nothing that two hands who promised nothing also produced — but a pun the enumeration rule could not see (its shape fell outside the rule's tolerance) turned out to be exactly the pun Eastwick's Persian carries at that locus. The honest negative is "at the places this study could enumerate, three blind readers found nothing," not "he broke his promise." A promise is a claim about the whole text; an enumeration is a sample of it, and a sample can miss.
3. A comparative numerical self-report is a third, distinct kind of statement, and the one case measured holds. Leaf's "rather more sparingly than Hafiz" is not a refusal (he does repeat rhyme-words) and not an unfalsifiable promise (it names a direction and a rough size); checked as a ratio against the same material in both languages, it is correct, against a registered prediction — that English's poorer rhyme inventory would force Leaf to repeat more, not less — that fails. A comparative self-report is the kind of preface sentence a handbook should treat as a checkable claim, not as an author's modesty.
4. Telling a reader a fact about the source moves what they prefer; showing them the source does not, and does something else instead. Told that an Arabic passage is rhymed prose, three blind seats moved toward the rhyme-carrying English at every content-decided segment in the first run and replicated the direction on fresh material and a properly pre-registered statistic. Given the Arabic itself — script and romanisation, no sentence about it — preference did not move, and the seats' own stated reasons changed character: they stopped judging the English as English and started checking individual word choices ("more accurate word choices ('kings' for aqyāl, 'getting a living' for iktisāb)"), which is exactly the kind of scrutiny that penalises the marked word order a carried prose rhyme costs (forty-two displaced clauses in the rendering measured). A note that states what the source does is doing work a facing-page original does not do on its own.
5. An author's own printed statement about his own book's language policy does not bind a translator, and a translator can violate it without any of the tension being visible on the page. All four English Dantes print the Latin their own author's stated intention forbade himself, one of them more than Dante allows himself, and one prints the broken promise and its own violation four sentences apart. This is the sharpest version of the pattern above: the printed statement (here, not even the translator's own) is not a description of the page that follows it, at any remove.

3. What this project's own practice found

• A translator's own log of what a declared refusal costs, kept before any of it was scored. Executing Preston's printed refusal deliberately (R50) rather than merely not reaching for the device (R43) is not the same act: an ordinary careful rendering of the same Arabic fell into 29 near-echoes and 5 identical clause-end words by accident, where the version built to refuse them has 5 and 0. Four of the eight refusals the deliberate version had to make are the same accident — an Arabic enclitic whose ordinary English is a final you, which cannot rhyme with itself. Refusing a device on purpose also pushed the diction toward commoner words: the refusing version has 4 clause-end words outside CMUdict against 15 and 18 for two versions that did not refuse (RS-20260825c-worth-paying §4). Declaring a refusal is not the same act as achieving it, and achieving it costs something specific and namable.
• A footnote reporting a device and the device appearing in the running prose are two different acts, and a translator's own notes census will find the gap between them. Reading every note printed on Eastwick's span: Gladwin prints none; Ross glosses realia only; Eastwick prints twelve notes, of which exactly one reports a play upon words — and that one stands on an Arabic verse couplet, not on Sa'di's Persian prose, and does not mention the pun that a companion note (on a textual variant) sits directly beside without naming (RS-20260827-declared-play §7). The declarer is the only one of three who reports a play at all, and the one he reports is not the one his preface promises.
• Applying the entry to a fresh case (this session): a translator's own sense that a device "obviously" needs a note is not evidence that it does, and checking the drafted English is cheaper than writing the note. Translating Miyazawa Kenji's 注文の多い料理店 (workshop/canon/chumon), the title's own pun — 注文 as both "a customer's order" and "an order given to someone" — reads as needing a footnote on first sight of the source. Drafting the passage first found that English "order" already carries both senses at a restaurant ("give your order") and in a command ("those are the orders"), so the same flip the Japanese makes at the reveal — "All those 'many orders' — it's them placing the orders. On us" — survives in English unaided. The note was drafted, then struck, on the check that the ambiguity is present on the page without it (T-chumon-R04-v1, translator's log item 1). One hand, one text, internal-judgment-only: no reader was asked whether the joke actually lands, only whether the English carries the same double sense the Japanese does.

4. The options

• Print a refusal, and let it be checked rather than defended. Cheapest and most reliable of the four kinds of statement; a refusal that is not kept will be found (§7.39), and the finding itself, on this evidence, is the useful outcome — not an embarrassment to avoid.
• Print a promise, but only after checking it against an enumeration of the loci it covers. An unchecked promise has, on the one case measured, produced nothing at the sites a study could enumerate — while still not being provably empty everywhere, since an enumeration is a sample.
• Print a comparative numerical claim rather than a qualitative one, where the claim is about frequency relative to the source; it is checkable, and the one case checked held.
• Disclose a fact about the source to the reader, rather than relying on a facing-page original or a bilingual edition to do that work. Telling moves preference toward a source-carrying rendering; showing the source does not, and shifts the reader's attention to word-level accuracy instead — which is a cost where the carried device's price is a marked English word order.
• Do not assume an author's own printed policy, if you reproduce it, describes your own practice or any other translator's. No case in this record has a translator honouring an author's stated language policy, including the translator who prints the policy.
• Before writing a note to explain a device, check whether the target language already carries the ambiguity or effect unaided. A note is a cost to the reader as well as a service; this session's practice found one case where the "obvious" note was unnecessary once the drafted English was checked.

5. Guidance

For a translator

1. If you plan to refuse a device, print the refusal and do not feel obliged to also demonstrate the refusal in the text — checking your own clause-ends against your own vocabulary's own chance rate is the only verification a refusal needs, and on this evidence it is more reliable than a promise ever measured. — evidenced (AR→EN, two hands).
2. If you plan to print a promise, check it: count what you actually kept, at every locus you can enumerate, before or immediately after printing the promise. An unchecked promise has, on this evidence, produced nothing at the sites a study could later enumerate — treat that as the default outcome of an unchecked promise, not a risk you can discount. — evidenced (FA→EN, one hand; AR→EN refusal-side precedent).
3. If you want to make a claim about your own practice relative to your author's, make it a checkable number (a rate, a ratio), not a hedge word like "sparingly" left unquantified in your own head. Leaf's did, and it held against a prediction that favoured English's constraints over his own count. — evidenced (FA→EN, one hand, one book).
4. If you want a reader to notice that you carried a source device, tell them what the source does in a note or preface line; do not rely on a facing-page original or a bilingual edition to do the telling for you. Showing invites word-level checking instead of a judgment of the English as English, and word-level checking penalises exactly the marked syntax most carried devices cost. — evidenced (AR→EN, three blind seats, internal-judgment-only, provisional).
5. Before adding a note to explain a source pun, wordplay, or built-in device, draft the passage first and check whether your own target-language text already carries the ambiguity or effect without help. This session's practice found a title-level pun surviving in English unaided, where the translator's first instinct was that a note was needed. — evidenced (JA→EN, one hand), craft observation.
6. Do not assume an author's own printed statement about his practice — quoted, translated, or reproduced inside your own translation — describes what any translation of the book, including yours, will do. No case in this record has one held, by any hand, including the one who prints it. — evidenced (IT→EN, four hands).
7. Run your own notes census before filing: read every note or gloss you have printed and check each one against whether the device it names actually appears, at the locus it names, in your running text. A declared device and a demonstrated one are different acts, and the gap between them is not visible from inside the writing of either. — evidenced (FA→EN, AR→EN).

For a pipeline

1. Enumerate every printed metatextual statement a translation will carry — a preface sentence, a footnote, a title's own built-in wordplay — and classify it as a refusal, a promise, a comparative self-report, a disclosure of a source fact, or an inherited author policy. — untested; the classification is derived from evidenced findings.
2. For a refusal: no pre-publication verification is required by this evidence; a light post-hoc check (the device's own rate against the output's own chance baseline) is optional and cheap. — untested.
3. For a promise or a comparative self-report: run a mechanical count against every enumerable locus before publication; withhold, revise, or narrow the statement if the count is zero or contradicts it. — untested.
4. For a disclosure meant to help a reader notice a carried device: prefer emitting a stated fact about the source over supplying the source text itself, where the goal is reader preference for the carrying rendering rather than bilingual scholarly apparatus. — untested.
5. For any language-policy statement inherited from the source: do not propagate it as a constraint on the pipeline's own output; flag it for a human decision on whether to reproduce, translate, or drop it. — untested.
6. Before emitting a note explaining wordplay or a source-bound device, check the drafted target text for the same ambiguity or effect and suppress the note if the check succeeds. — untested; derived from one craft case (§3).
7. Check: run a notes census over every footnote or gloss the pipeline emits, verifying each against the presence of its claimed device at its claimed locus; flag any note whose device is absent. — untested.

Human entry points. Which kind of statement to print, and whether to print one at all, is a purpose decision (declared-purpose, wiki/goodness-senses.md): a scholarly edition's reasons for disclosure differ from a trade edition's. Without a person, default to: print refusals freely; suppress promises and comparative claims until a mechanical check passes; prefer disclosure-of-source-fact over supplying source text; never propagate an inherited author policy as an output constraint.

Application. This session (S255) applied item 5: translating a fresh, compressed extent of Miyazawa Kenji's 注文の多い料理店 (T-chumon-R04-v1, JA→EN, ~300–800 English words), the translator drafted a footnote explaining the title pun, then struck it on checking that the drafted English already carries both senses of "order" the reveal needs. Followability cost: real but small — one re-reading of the draft, no rewriting; the instruction decided a concrete question (write the note, or not) rather than confirming a plan already settled.

6. Not evidenced, and open

• No reader measurement of any of this. Every preference and detection figure in §2 and §3 rests on three model seats; Tier D is NOT PASSED and is now EXHAUSTED (config/models.md). A model panel's stated preference is not shown to be what would move a human reader's.
• One promise (Eastwick), one comparative self-report (Leaf), and one refusal pair (Chappelow, Preston) have been checked against a page. Nowhere near a sample of translators, periods, or device kinds; §2's regularities describe these specific cases, not a general rate.
• §7.34's own headline statistic (the both-orders content-decided count) is withdrawn as a statistic, though not as a direction. Registered properly at RS-20260827b-shown-or-told on fresh material, the identical statistic came back 3–2, which the pre-run critic showed cannot reach significance at nine segments whatever it shows. The told direction is confirmed and strengthened by the properly registered run; the original number from §7.34 should not be cited again.
• A conjecture named at RS-20260827c-declared-page §10 and never taken up: whether a disclosure that favours neither of two arms equally has nothing to move a reader's choice with, as against one that plainly favours one arm. Untested; Open for whoever next works disclosure design.
• Whether Leaf's comparative self-report generalises past one translator's one book is untested — n = 1, and no second case of this kind of printed statement has been checked anywhere in the record.
• Open: does the "check the draft before writing the note" practice (§3, §5 item 5) generalise beyond one translator's one title-level pun, or is it specific to English's own accidental double sense of "order"? The family rule sends this to whoever next works a title-level pun or a wordplay-bearing title in a fresh pair.
• Open: whether a systematic notes census, run as a required step rather than as a diagnostic applied after the fact, actually changes what a translator or a pipeline prints — demonstrated here only as a way of finding a gap that already existed (§3), not as a tested procedure.
• Not evidenced on any pair with a target language other than English, and not evidenced where the "telling" is by someone other than the translator (a publisher's jacket copy, an editor's introduction) — this entry is about the translator's own printed statements only.

7. Sources consumed

• v0.2 §7.34 (S222, RS-20260825c-worth-paying): first disclosure-moves-preference finding, on a withheld primary and a post-hoc both-orders statistic — into §2 item 4 and §3; the statistic is withdrawn (below), the finding is not.
• v0.2 §7.37 (S226, RS-20260827-declared-play, ARM-declared-function closed 2/2): refusal predicts, promise does not — into §2 items 1–2 and §3's notes-census finding.
• v0.2 §7.38 (S227, RS-20260827b-shown-or-told, ARM-worth-paying closed 2/2): telling moves preference, showing does not, on a properly registered design — into §2 item 4. This page's own §4 finds §7.34's both-orders statistic does not replicate (3 of 5 registered segments, non-significant); the direction is confirmed, the number is not, and §6 states this rather than silently reusing the old figure.
• v0.2 §7.39 (S228, RS-20260827c-declared-page): page-side measure of whether a printed refusal is visible in the prose itself — into §2 item 1; grows A-hariri-hands to three Assemblies.
• v0.2 §7.43.2 (S233, RS-20260830-leaf-contract, building A-leaf-hafez): the comparative self-report that held — into §2 item 3.
• v0.2 §7.50.6 (S241, RS-20260903-latin-in-italian): an author's own printed policy, unkept by every translator who reproduces it — into §2 item 5.
• T-chumon-R04-v1, this session's translation limb, added 2026-09-08 (S255), post-consolidation.
• Withdrawn along the way: §7.34's both-orders content-decided statistic (the number, not the direction) — withdrawn by RS-20260827b-shown-or-told §4, declared there and restated here rather than repeated as if it still stood.