Repository path: framework/v0.3/entries/HB-footing-and-address.md · rendered 2026-09-09

When a source grades speaker against addressee, and English has no matching slot

Standing. Evidence classes X1a (published translations read against source, machine- or lead-coded) and X2 (machine-measured, recomputed); no X3 — Tier D is NOT PASSED, so nothing here is a jury's word that one rendering carries the relation better than another. Most coding is the lead's own, unanchored (internal-judgment-only); two studies bought an independent second coder (RS-20260814b, RS-20260816d), reported rather than assumed. Consolidated 2026-09-04 (S245) from framework/v0.1 whole and framework/v0.2 §2–§6, §10.

1. The problem

A source grammar can grade the speaker against the person addressed — a pronoun distinguishing formal from familiar (Russian ты/вы), a verb inflecting for the addressee's or referent's standing (Japanese honorific morphology, Bengali verb-final politeness), a clitic glued onto whatever word is handy (Russian «-с»), a plural verb form applied out of respect to someone the speaker names but is not addressing. English marks almost none of this grammatically: one second-person pronoun, no addressee-agreement on the verb, no honorific affix. Two questions follow. First: when English cannot carry the device, does it carry the relation — does a reader still learn who defers to whom — or is something actually lost? Second: if a translator tries to compensate, where does the marking go, what does it cost, and does it end up saying what the source said?

2. What published translators do

The regularities, each with what it was measured on:

1. Where the source grades a relation in a slot English fills compulsorily with one form — above all the personal pronoun — the marking does not reach English, and is not compensated elsewhere either. RU ты/вы 0/11; JA honorific pronouns 0.022/13; BN total loss; DE 0/9; two further Japanese strata at 0.000. Six measurements, four pairs, all at the floor. (RS-20260802e, RS-20260812i, RS-20260813e, RS-20260812h, RS-20260806e, RS-20260814b.)
2. S1 — the loss this line of work exists to repair turns out to be uncommon. Five attempts across five pairs never produced a population of sites where a competent English rendering actually loses a grammatically-marked relation. Dropped devices usually transfer anyway: King's device-free German scores 8/9; Garnett's deleted clitic transfers at error 0.048; a close Japanese rendering carries no device at 33/51 sites and the phenomenon that would matter occurs at only ~8% of them. Expect most sites to need nothing. (v0.1 §2–§3, v0.2 §2.)
3. The one exception is a site where the marking is all the source has. Russian «Что-с?» — one pronoun, one clitic, no other content — is mis-stated by every rendering dropping the clitic and correctly stated by the one keeping it (What, sir?). n = 1, but the same shape recurs in Kohlhaas's A2, where the speaker's content works against the deference his address claims. Where content and relation pull apart is where the device earns its keep; elsewhere it is close to redundant. (v0.2 §2, §5.)
4. No predictor of where a lost marking lands has survived a test, in four attempts. A positional account (English's free utterance-final slot is why «-с» reaches English at 42/43) failed on Japanese honorific endings in the same slot: 0.200 against a registered 0.50 — and 1,762 words of published dialogue use no sir at all. A part-of-speech account (御 should cross because it attaches to a noun) crossed 0/14. Two further attempts inside ARM-carrier failed at their own instrument gates. A framework recommending "look for the free slot" would be recommending something no translator here does. (v0.2 §2, §10.2.)
5. Where English does carry a grammar-only site in a published hand, it is a verb, not a noun. Across 5,361 words, five hands, one site: Waley's deferential auxiliary — "but even if you will not deign to have any dealings with us…" An earlier reading proposed a rank-noun device instead; given a second source it was withdrawn — the machine hands use it zero times on the new material, and the earlier case turns out to depend on a daimyō being present to address, not on a resource this scene lacks. (v0.2 §10.3.)
6. Device categories are alternatives, not ingredients — three separate contrasts say so. A pronoun and an adjective change different things (a suffix modulates a relation; an adjective states a fact the source did not). Carrying Kohlhaas's asymmetry on the pronoun made the address-noun redundant — seven vocatives collapsed to zero. Inside one Chekhov sentence, where a noun phrase and a clitic mark the same relation at once, the noun phrase carries the whole load and the clitic is redundant on top of it — until the noun phrase is absent, and then the clitic is the only thing there (item 3's «Что-с?»). (v0.1 §4, v0.2 §3.)
7. The pronoun category cannot reach an up-address relation in English at all — no deferential second-person pronoun exists — and the one Kohlhaas site that mattered (A2, the up-address) is exactly where a pronoun-based compensation fails while it succeeds elsewhere. A device can also assert a register the source does not: the archaic pronoun dragged fifteen archaic verb forms in with it, ageing speech the German does not age. (v0.1 §4.)
8. Nothing is actually unreachable, when a translator is asked to try. Forcing every device at 52 Genji sites left 51 carried, one abandoned — zero sites with no device to reach for, contrary to this project's own earlier notes. (v0.2 §10.5 item 1.)
9. Carrying everything is mostly bulk, and it manufactures a grading the source did not make. The maximal rendering ran +21.1% words, one marker every 13 words against the source's one every 23.6 characters — a comparable rate, a very different visibility (the source's marks are inflections already there; the English ones are extra words). Against a bulk-matched decoy, only 30% of the naturalness cost (0.417 of 1.389 points) is attributable to the deference devices; 70% is what adding a fifth more words costs regardless of content. Worse, the direction of the perceived grading follows local device density, not the source — whichever character a paragraph happens to be about reads as higher-status, confirmed on a matched pair holding events fixed. A translator carrying every mark is not carrying the source's grading; they are manufacturing one that tracks which paragraph is about whom. (v0.2 §10.5 items 1, 2, 2a, 6, 6a.)
10. A rule that marks whoever a paragraph's own devices put highest delivers a smoother wrong answer. Tried at all twenty graded paragraphs of a censused span, it solved the source's three direction-reversals in writing, and none of the three solutions reached a blind judge. The open question is not a finer rule of the same shape but whether anything below the scene is the right grain, since English marks with lexis that persists across paragraph boundaries a source inflection does not cross. (v0.2 §10.6.)
11. A title already in the prose outranks any amount of grammatical marking on people actually present. Where the source marks two present characters eleven and twelve times, three blind readers unanimously name a character who is not in the room and never speaks, because his rank noun is the only one on the page; in three of six segments, marking scores are driven by nouns present in every arm, including the unmarked baseline — a lexical channel no honorific-morphology census was counting. (v0.2 §10.5 items 6b, 7.)
12. Independent readers cannot agree on which sites are "silent." Two readers of the same 70 utterances agree on how much footing a determinate line carries (within one point, 9/9), but on whether a number is even possible they agree at 0.485 — worse than chance — one calling 87% of sites determinate, the other 35%. "Watch the places where English says nothing" has no denominator. And a choice made at a genuinely silent site is taken for a reading of the source anyway: identical English lines read as polite-wife/plain-husband or the reverse purely by which name labels which line, flipping sign in 3/3 hands when the labels are swapped — the deference lived in the label, not the sentence. A separate case shows the lexical channel can be invisible to a grammar-built instrument: an English "sir" rendered as an address term rather than a polite predicate scores 0 on an honorific coder though the mark plainly crossed by a different route. (v0.2 §10.7, Q5/SWAP.)
13. Where a source marks a relation by a change across a passage, English carries it only if the device has any English exponent at all. A pronoun switch (ты→вы mid-scene) is not recovered by independent readers (+0.134 against a null-favouring bar of +0.40); a vocative-borne change on the same design is recovered at +0.786. A per-utterance instrument cannot see this class at all: discordance is a property of a text's trajectory, not of any one line, which is why no per-utterance admission condition for it ever reproduced. (v0.2 §6, RS-20260808b §4.2, RS-20260809a.)

3. What this project's own practice found

The lead's own forced renderings (internal-judgment-only, hypothesis-aware, excluded from every registered figure above) supply the process evidence the published censuses cannot:

• T-kohlhaas-R20-v2 (DE→EN, whole Luther scene, contamination: high): rendered under the pronoun device the census had excluded as one-sided. Carrying the marking on the pronoun made the address nouns redundant (seven vocatives to zero) and dragged fifteen archaic verb forms in with it — where items 6–7 above were first seen, not read off a published hand.
• T-genji-yomogiu-R06/R27/R28/R33 (JA→EN, one span, one hand, four regimes in one session, the source of items 8–10): the only design here where the translator's intent is the manipulated variable and the source text is held fixed. untested in the D-20260724-04 sense — one hypothesis-aware hand self-reporting on its own sentences — citable for what was reachable at what price, never for what English "does."
• RS-20260812d's floor measurement: the same translator re-rendering the identical Japanese passage twice at temperature 0 produced texts only 0.8333 similar on a 0–3 scale. Before a future design proposes to detect a shift this size, it must measure what one translator's own repetition moves and set its bar above that — several of this project's earlier statistics never did.
• The lead's own frozen translator's log (v0.2 §10.7, RS-20260816d Finding B): classifying 23 Dickens-into-Japanese sites, calls 4 "the English words fixed the level" and 19 "nothing did, and I supplied it" — echoing item 12's finding that independent readers cannot agree where the silence is.

4. The options

At a site where the source grades speaker against addressee, the moves are: carry the device in the same English category; displace it onto a different category (noun, verb, epithet) where English's own category has no exponent; rely on content, trusting the scene's own words; or collapse, which is what dropping a compulsory pronoun distinction always is. The evidence says:

• The compulsory-slot case (the pronoun, above all) has no same-category fix, in four pairs. The only live choice there is displace or rely on content — never carry.
• Relying on content is usually sufficient. S1 (item 2) says the loss a compensation rule exists to repair rarely shows up; default to doing nothing and checking the scene.
• Displacement is not free even when available. It buys visibility at the cost of bulk (item 9) and can manufacture a grading the source never made (items 9–10), whichever category absorbs it.
• Exhaustive and selective displacement are not equally measured. Exhaustive carriage has a full price list (item 9); whether a light, selective purchase is audible at all is unresolved, not a pass (§6).

5. Guidance

For a translator

1. Enumerate every relational device the source grades — grammatical and lexical — before opening any other translation: does English fill the category with one compulsory form (chiefly the pronoun), or is a lexical/verbal option available? — evidenced (RU, JA, DE, BN, FI, IT, LZH, ES).
2. Where the marking rides a compulsory single-form English slot, do not hunt for a same-category fix — none exists — and do not invent an archaic pronoun as a substitute; thou asserts a register and period the source does not have. — evidenced (RU, JA, DE, BN).
3. Before compensating anywhere, check what the scene's own content already conveys. The relation usually transfers even when the device is dropped outright; expect most sites to need nothing. — evidenced (RU, JA, DE, PL).
4. Check deliberately only the site where the marking is all the source gives you — nothing in the content states the relation independently. There alone, dropping the device has been shown to cost. — evidenced (RU, DE), n = 1 each.
5. If you do compensate, choose one device category, not several: a predicating device (a noun phrase, a verb) makes an attaching device (a clitic, a pronoun) redundant on top of it; reach for the attaching category only where nothing else is carrying the relation. — evidenced (RU, DE).
6. Do not reach for the pronoun to carry an up-address relation — English has no deferential second-person form, and this is the one direction a pronoun-based compensation has failed. — evidenced (DE), n = 1.
7. At a grammar-only site, a deferential verb (deign, be pleased to, condescend) is the one category a published hand has been observed reaching for — but check the source's predicate is something its subject chooses: these devices assert volition, and one added to an involuntary state (a pang, an illness) makes a different event. — evidenced (JA), one hand.
8. Do not look for a rule about where a mark "naturally" lands — a free slot, a compatible part of speech. Four attempts at this kind of predictor have all failed. — evidenced (RU, JA).
9. Price exhaustive carriage first: roughly a fifth more words, most of the naturalness cost from the bulk rather than the devices, and a real risk of grading people by which paragraph is about whom rather than by the source. Decide whose footing a scene may mark before deciding how much. — evidenced (JA), one passage.
10. Do not trust your own or a reader's sense of which sites are "silent" — independent readers agree on this at worse than chance. Assume you supply the relation at every site of a relationship and decide it once, deliberately, for the whole dyad. — evidenced (JA→EN, RU).
11. Check what your nouns already say before adding anything grammatical: a title already in the prose, even for someone absent, outranks grammatical marking on people present, and a reader takes an undeliberate choice for a reading of the source. — evidenced (JA), one hand.
12. Where a source marks a relation by change across a passage, plan the change as one decision spanning the passage, not per utterance — and expect it to carry only if the device has some English exponent (a vocative does; a pronoun switch does not). — evidenced (RU), qualified.

For a pipeline

1. Enumerate: detect grammaticalized relational devices in the source and classify by target-slot type — compulsory single-form, or open — before any target text exists. — untested.
2. Classify English's corresponding category as compulsory-single-form or open, from a declared device inventory for the pair. — untested.
3. Set the policy parameter: exhaustive / selective / none, from declared purpose. Default selective: exhaustive is priced at +20% words and distorts direction (item 9); none ignores the residual case of item 3. — untested as a switch; the priced comparison exists for one passage.
4. If selective, freeze one footing decision per dyad for the whole passage, from its trajectory, not per utterance. — untested.
5. Render: prefer a verb/predicate device at grammar-only sites; never a pronoun-category device; flag for human review any site where the device would assert volition on a non-volitional predicate. — untested.
6. Check: scan the finished English for rank nouns/titles independent of the added devices; report per-class counts against policy and a naturalness delta against a bulk-matched decoy, so the cost of marking is not confounded with the cost of length. — untested.

Human entry points. Step 3 is a purpose decision for a person; step 5's volitional-mismatch flags route to a person. Without one, step 3 defaults to selective.

Application. This session's translation limb (T-hirurgiya-R04-v1, Chekhov's «Хирургия», RU→EN, a fresh excerpt) applies items 2, 3, 5 and 12. Followability log:

• Item 2 held without a workaround: the deacon's вы / the feldsher's ты got no English device; an archaic pronoun at the moment of collapse was foreclosed before it was considered.
• Item 1's enumeration placed item 12's register break. The Russian marks the relation three ways — вы-forms and name-patronymic address, вы-imperatives turning to ты, and (outside R1's six categories) an honorific plural verb for an absent third party («Отец диакон велели»). Naming all three before drafting let the pivot land at Chekhov's own clause rather than being smoothed across the passage.
• Item 3 did the most work, at no cost. The honorific-plural device has no English counterpart and was rendered as an ordinary past tense; the adjacent lexical blessing formula in the same sentence («дай бог им здоровья») does have one ("God grant her health") and was carried directly — content already doing a job a device looked needed for.
• Item 12 cost a planning pass, not a wording cost. The English shifts from deferential phrasing ("Kindly open your mouth wider…") to blunt commands ("Pull it, then, pull it!" / "Fool!" / "Fool yourself!") at the clause both speakers break register in Russian. One read-through located the pivot before drafting; no line ran longer than its original once the pivot was placed.

6. Not evidenced, and open

• No jury has scored any of this. Tier D is NOT PASSED; nothing above licenses a claim that one rendering carries a relation better than another — only whether a device or a relation is present.
• Whether a light, selective purchase of devices is audible at all is UNRESOLVED, not a pass — the one design that tried it could not distinguish a twelve-device rendering from a mislabelled duplicate of its own baseline.
• Whether anything below the scene is the right grain for a footing decision is open: a dominance rule applied per paragraph produced a smoother wrong answer.
• The volitionality account (§5 item 7) is untested against a registered check — noticed while coding the material it describes, the weakest provenance this project counts, though cheaply testable against hands already censused.
• Period is confounded with hand in every pair with a published-hand census: every Japanese and German hand measured is pre-1930.
• Much of this evidence rests on model seats standing in for readers (charter §4); "reader" names a rating seat throughout, never a person, and no human reader is available to this project (Tom is never an experimental subject, charter §9).
• Nothing here separates a scene's own narrated events from its wording except one matched pair built to hold events fixed (item 9); every other perception figure was measured on passages whose incidents differ along with their translation.
• The reverse direction (EN→JA) has one measurement and a serious limit: its two "independent" Japanese hands share a 33-character run, closer to one observation than two.
• Open: whether the volitionality account replicates; whether a below-scene footing rule can be built that does not repeat the dominance rule's failure; whether the lexical-channel finding (§2 item 12) generalises past one clause of one chapter.

7. Sources consumed

• framework/v0.1/README.md whole (§2 R1, §3 predictions, §4 limits, §5 candidates, §7 pairs, §8 Q-a/Q-b/Q-c); §3 prediction 1 retired by v0.2 §2 rather than restated.
• framework/v0.2/README.md §2 (S1, ARM-carrier), §3 (Chervyakov device contrast), §4 (RU→EN strengthened), §5 (Q-b), §6 (Q-d), §10.1–10.7 whole (ARM-footing, ARM-honorific-hands, ARM-supplied-footing — the census, four failed predictors, what actually carries, the reachability/price/direction/lexical-channel findings).
• RS-20260802e, RS-20260803c (void; its non-void finding — occupied ≠ blocked — is used descriptively), RS-20260806e, RS-20260808b, RS-20260809a, RS-20260812d, RS-20260812i, RS-20260813e, RS-20260814b, RS-20260814h, RS-20260815c, RS-20260815d, RS-20260816d — every §Limits section read in full; RS-20260812h (Bengali «ডাকঘর») cited via v0.2 §10's own summary, not re-read whole.
• Withdrawn along the way: v0.1 §3 prediction 1, retired (v0.2 §2); the positional and part-of-speech predictors of §10.2, both failed; the rank-noun theory of §10.3, withdrawn on a second source; the insertion-reallocation claim of §10.4, withdrawn, failed to replicate twice; §10's pre-2026-08-14 sentence that nothing can be recovered, corrected by §10.5.