Repository path: framework/v0.3/entries/HB-long-works.md · rendered 2026-09-09

A work too long for one sitting: the register that binds the next span, the text that shifts under it, the method that tires

Standing. Written at the project's close (2026-09-09, S257) as a consolidation of the record without the translation limb the procedure's step 6 requires — no fresh passage was translated under this entry, so it stays status: draft and its §5 Application reads "not applied". This family's evidence is craft reports more than measured results: five closing essays, all the lead's own coding of its own six long works (internal-judgment-only), with four measured limbs beside them (X2 counts on stored texts; X1b readings of published hands). Tier D is NOT PASSED and EXHAUSTED (config/models.md, RS-20260906-tierD-verdict-v3): nothing here says one rendering is better than another on a jury's word; X3 is cited only to be excluded. framework/v0.2 has no numbered section for this family (§7).

1. The problem

A short piece is frozen in one sitting with every policy still in view. A long work is frozen in pieces, weeks apart, with other work between (R05): a term settled in span 1 binds span 5; a rule learned in span 4 cannot reach the frozen span 1; the source can prove to be a corrupt reprint or one recension among several; the translator who finishes is not the one who began; and the method tires in ways a single sitting never shows. Charter §3 wanted a long work for sustained voice, cumulative characterisation, motif and terminology consistency, fatigue of method. Six were translated whole under R05 (Verga; Canth; Mikszáth Part I; Tagore; the Nights' frame and seven nights; Sa'di's chapter II); a seventh (Lazarević) was chosen and not begun.

2. What published translators do

1. Two published hands of one long work are two books, not two poles. The hand that deletes the outermost frame keeps every formula inside it exactly; the hand that keeps every frame varies every formula (RS-20260823b §10). What each refuses to lose survives as a comparison; which is more faithful does not (ES-20260826 §4).
2. Where recensions diverge, a hand's absence is permanently ambiguous between the translator cut it and his Arabic lacked it; attribution was possible at two loci only (RS-20260823b §5–§6).
3. Published hands were used as contamination baselines, not read as translations (ES-20260810 §5; ES-20260813 §6).

3. What this project's own practice found

(The arms — ARM-longwork, ARM-legend, ARM-dakghar, ARM-alf-layla, ARM-gulistan — and the craft reports' §1.)

3.1 Choosing, and the gate that runs on span 1

The binding constraint was free source and free comparator together: two of five candidates died for want of a reachable English and two on scale, since a contamination: line with no comparator is a placeholder (RS-20260727-longwork-selection §1). Contamination cannot be measured before something is translated: measure on span 1 and be prepared to throw span 1 away (§5). The gate's one known-ground-truth case is accidental (232 of 2,566 words primed by reading Dole, declared before the tool ran; the longest run fell inside the primed 9%; p ≈ 0.090 — consistent with detection, not validated, §4). Measure-once is wrong: Verga's runs went 15 → 14 → 11 → 14, so S042's "monotone" reading described three spans, not the work (ARM-longwork S045); ARM-gulistan re-measured after every freeze. Where no English exists, declare none, not measured and record the routes searched (verter/plan.md §2b). W3 step 5's criteria (wiki/plan.md) applied S256.

3.2 The binding register, and what it caught forward

R05 rule 4: register.md carries the terminology, naming and voice decisions (the Gulistan register adds figure policy) that bind every later span, each citing the log decision that created it, plus an unresolved section the next span must settle; revisions need a numbered erratum. The Gulistan register opened at fourteen rules, grew to nineteen, closed unextended at span F (gulistan-bab2/register.md). Caught forward: Verga N15 mandra → herd, set when every referent was equine, broke on sheep in span 4 (E5); the Nights' question 12 wrote down in advance what would falsify T2 slave-girl — span J met the bar and the erratum was still not needed (ES-20260826 §5). A play's register is a register of address: five of Tagore's nine rules govern what people are called; V3 paid for itself 16 times in span A (ES-20260813 §2).

3.3 The register is blind backwards — measured on four works

Named on Verga at S047 (two rows wrong about the frozen prose they described; E6), then measured by re-rendering the first span under the closed register, blind to the first rendering:

• Canth (RS-20260806d-first-span-again, 1,795 Finnish words): 13 of 234 divergence sites attributable to something learned later; 11 of 13 caught by no erratum — all five errata came from the collation or a comparator. Coding is the translator's own; the numerator survives as a claim about eleven named sentences, the 234 does not (§7).
• Mikszáth (RS-20260810x, via ES-20260810 §2.4): 160 sites in 1,963 English words; 5 coded exclusively to a later rule, 9 with one governing; 141 residual unaudited. Of 22 decisions logged before the comparator opened, 12 same, 10 different; D66's argued choice was the one the register-less habit had already made.
• Tagore (RS-20260813d, via ES-20260813 §5): (B) is zero — every later rule was anticipated by span A's practice; only a lexical chain moved (§3.5).
• Sa'di (ES-20260905 §2, §4.1): 13 of 22 inherited figures survive the register, 9 do not; V5's frame/gloss test reclassifies 4 of 7 Arabic insertions; two designation collapses corrected — one of them the collapse D45 predicted three spans earlier — and one false friend caught by reading.

3.4 The register's price, counted — and invisible to copy-editing readers

The Nights counted it: span H, 3 of 4 sound-figure losses were the price of terms frozen earlier; span J, 2 of 3 (ES-20260826 §5). RS-20260826c-register-cost (11 loci, bound vs refused, 3 seats, 174 live calls): seats queried both versions at the same rate; the frozen term was quoted twice in 111 calls, the refused alternative never; at L1, the register's own predicted failure, 0 of 11. The instrument catches blatant seeded defects at 22/22, subtle ones at 0/10, so the null is not costs nothing. What gets flagged is the programme — calques, cognate figures, coordination — adopted once and never rethought.

3.5 Chains are visible only once every span that carries them exists

Sa'di V19 (a word repeated across tales keeps one English word): nine chains in span D against one in C, two in E; D70 found a pair crossing the D→E boundary; span F found عارف at three points (b026 → b321 → b368) and مدّعی at four (b048 → b306 → b320 → b345), invisible until the first-translated tale was re-rendered under the closed register (ES-20260905 §4.3). Mikszáth V15 (five names for an umbrella, none normalised) and V21 (a varied designation is not collapsed, a repeated one not varied; the inherited rendering broke both, 4 of the 10 differing rows) (ES-20260810 §3). Tagore: kṣep- rendered three ways in v1, one root in v2, paying at a sixth site where V4 forbids collapsing pāgal onto it — a chain has to be opened at the first occurrence or it cannot be closed at all (ES-20260813 §4.3). Verga: gnà, kept at D4 for an unrelated reason, was the only instrument for a T/V switch 8,600 words later (ARM-longwork S045).

3.6 The text under the arm can be the wrong text; only length finds it

• Canth: three free e-texts are one transcription of one 1917 reprint; against the 1886 first edition, 59 real divergences (4.4 per 1,000), five corruptions, two invisible until the sentence had to be rendered. Yet 0 of 7 first-edition readings changed an English word in span 1 (ES-20260806 §3.6, §4).
• Mikszáth: 164 aligned divergences, 18 substantive; of four readings flagged will change an English word, one did; a paragraph division and an italic were found only by translating. A gate run once is a gate that has been run once (ES-20260810 §2.3, §3.4).
• Tagore: 24 speeches, 335 words, 5.8% of the play, absent from the project's own copy — each the last speech on its page; only the third span showed a rate and a cause (ES-20260813 §4.3a).
• Sa'di: single witness declared seven times, not re-derived; the S227 fetch byte-identical over 68 blocks; one published block count wrong (gulistan-bab2/collation-chapter.md, collation.md).

3.7 Recensions: the source is a variable, and regularising it destroys evidence

Ten Nights spans collated against a second Arabic witness: 18 to 83 non-identical stretches per span at similarities 0.928–0.952 — a verb censored twice, 190 words missing, and at span A a copy-text shorter than the Arabic behind both hands, leaving one locus permanently ambiguous (ES-20260826 §1). T1 made the sage Ruyan where every English Nights has Douban — page against tradition, chosen in public (D85). V15 refused to regularise three ways of marking a night, which made a measurement possible: 56 of 56 boundaries found at ±1, 0 false alarms, 30 bodies — the dawn formula does not make a seam findable; it places it (7/7 with it, 4/8 with a neutral substitute, 1/7 with nothing; P = 0.00233); a heading-only boundary was invisible to all 22 bodies (RS-20260816f §2–§5).

3.8 The self-match: the lead is not an independent sample of itself

Two lead renderings made in different sessions, unseen by each other, matched at up to 37 tokens (note (bhb) as first recorded, since revised to 51); Tagore's two passes at one page shared 60 twelve-word runs, longest 25, while neither shared one with Mukherjea (RS-20260813d §2). Gulistan span F, whose brief required reading R52-v1 in full before drafting, matched it at 536 shared 7-grams, 153 15-grams, a 51-token run, against 0 shared 7-grams and 4–5-token runs for the lead's seven other Gulistan renderings — the highest self-match recorded, produced by the comparison task itself (ES-20260905 §3; note (bta)). The nine checked loci stand; the measured rather than recalled framing does not.

3.9 Cadence, and fatigue of method

A three-session cadence over a five-track rotation was unmeetable the day it was declared: Canth's seven visits all fell every fifth session (ES-20260806 §2; note (bjp)). Declared in units of the rotation, a five-session cadence held on three of the four arms that declared it (§3 table). The method also tires the other way: a translator holding the whole book is not uniformly better at its first page — Canth's reviser collapsed sour tears into bitter; Tagore's re-render enlarged one realia loss (ES-20260806 §5; ES-20260813 §4.2). Decisions adopted by drift (Verga D19, a contraction policy binding 11,000 words) are visible only in serial work (R05 §6).

3.10 What the length pressured, per sense

Verga moved consistency and voice and returned nulls on naturalness and affect — the latter structural: a log records choices, affect is reception (ARM-longwork). Canth's address grid closed only at the end (V28); its fifth, speaker-determined cell is what English loses (ES-20260806 §3.1). Tagore's grammaticalised deference is plotted, not pervasive: ten sites in 430 units (ES-20260813 §3).

4. The options

• Which work: both sides free or the work the criteria want with none, not measured declared — buying voice, chain and fatigue evidence, forfeiting every two-hands comparison (verter/plan.md §3).
• Gate once or every freeze: once would have carried a wrong description of Verga (S042's reading, corrected by re-running at S045); every freeze caught span F (§3.1, §3.8).
• Register forward-only, or re-open span 1: forward-only left 11 uncaught sites in 13% of Canth and a broken chain in Tagore; re-opening by re-rendering for comparison produces the 51-token self-match unless drafted blind (§3.3, §3.8).
• Regularise the source's inconsistency or carry it: carrying it cost the reader nothing at a boundary with prose and the whole boundary where there was only a heading — what the copy-text costs its own reader (RS-20260816f §6).
• Page or tradition where recensions diverge (Ruyan/Douban): either states what the translation is for (§3.7).
• Cadence in sessions or in rotations: only the second is meetable by construction (§3.9).

5. Guidance

For a translator

1. Confirm the comparator exists and is reachable before choosing; if none does, declare none, not measured with the routes searched and forgo every independence claim. — evidenced (IT→EN, FI→EN); the SR→EN selection record is a note (§6); note (bmw).
2. Translate span 1 before trusting any contamination figure, be prepared to discard it, and re-measure after every freeze. — evidenced (IT→EN, FA→EN).
3. Store the whole cleaned source once, index spans into it, break at seams, and gate the copy-text on every span including inherited ones. — evidenced (FI→EN, HU→EN, BN→EN, FA→EN).
4. Keep a binding register from span 1: rows cite their log decision; an unresolved section the next span must answer; revision only by numbered erratum. — evidenced (all six).
5. Write into the register, in advance, what would falsify a term. — evidenced (AR→EN).
6. Open every lexical chain at its first occurrence; check chains against the whole work, not within a span. — evidenced (FA→EN, BN→EN, HU→EN).
7. Do not normalise a designation the source varies; do not vary one it repeats. — evidenced (HU→EN, FA→EN), descriptive.
8. Expect the register to be blind backwards; budget a return to span 1 under the closed register, drafted blind and diffed only afterward. — blindness evidenced (IT→EN, FI→EN, FA→EN); HU→EN confounded, BN→EN null; drafting blind untested.
9. Where the source has recensions, collate before translating and publish it as the record of what the English translates; carry the irregularity; choose page or tradition explicitly. — evidenced (AR→EN).
10. Declare the cadence in units of the rotation. — evidenced (FI→EN, HU→EN, BN→EN, AR→EN, FA→EN).
11. On the last visit, review the programme, not the word: the frozen term is below what a copy-editing reader remarks on. — evidenced (AR→EN, three model seats), one run.
12. Treat a re-rendering made after reading the whole book as a reviser's pass, not an improvement. — evidenced (FI→EN, BN→EN), craft observation.

For a pipeline

1. Enumerate candidates against the W3 step 5 criteria; fetch-verify both sides; probe each candidate that has a comparator with dependence_check.py. — evidenced (IT→EN, FI→EN; one SR→EN selection probe, §6).
2. Commit the whole cleaned source, indexed spans, seam breaks. — evidenced (FA→EN, AR→EN).
3. Gate each span's copy-text against a second witness or page image; flag paragraph division and italics. — manual runs evidenced (FI→EN, HU→EN, BN→EN, AR→EN); automated untested.
4. Render under R05; append the log; update the register with cited rows and an unresolved list. — evidenced (all six).
5. Measure contamination after every freeze against every published hand and the lead's own prior renderings. — evidenced (FA→EN).
6. Census the target text for chains, designations and transliterations — a count of the register's rows is not a count of what the text does. — untested; the failure evidenced (BN→EN).
7. Re-render span 1 under the closed register, blind to v1; diff; code by cause with an external coder. — re-render evidenced (FI→EN, BN→EN, FA→EN); blind drafting and external coding untested.
8. Report per sense what the length pressured, naming the nulls. — evidenced (IT→EN) once.

Human entry points. Step 1's choice of work and the page-or-tradition / regularise-or-carry decisions are declared-purpose decisions and route to a person; without one, default to the copy-text and carry its irregularity. Step 7's causal coding is the one step the record marks as unaudited when the translator does it alone (RS-20260806d §8.1); without an external coder the count is descriptive only.

Application. Not applied: written at close-out without a translation limb.

6. Not evidenced, and open

• Six works, six pairs, one translator, no jury, no human reader. Every backward-blindness count is the translator coding its own two renderings; the external coder built at ARM-two-hands (~$0.15 for the 234 sites) was never run (ES-20260806 §8).
• The 234-site denominator is not stable: a single draw, indistinguishable by count from two independent hands (RS-20260807c, RS-20260808a).
• affect and naturalness remain unmet charter reasons for a long work (§3.10).
• Whether the register's price reaches any reader is unmeasured, with a measured bound (RS-20260826c §4).
• Terminology drift and the consistency sense: the one measurement (RS-20260809i, 717 Ukrainian words, four model seats) is X3, inadmissible here, and family M's.
• No published hand of a whole work was read as a translation at length.
• The seventh work (Lazarević «Вертер», SR→EN, 14,293 words, six spans, no comparator) was chosen and not begun; pairs: excludes it.
• Open: does drafting a re-render blind, then diffing, lower the self-match below 51 tokens ((bta); ES-20260905 §7)? Open: the variation half of V21 has no measurement; Worswick renders all five names (ES-20260810 §7). Open: a first span translated by two different hands (ES-20260806 §7). Open: a deliberate priming design to validate the gate (RS-20260727 §4). Open: Tagore's kṣep- chain is unrepaired; D78's fourth refusal currency goes to family E; Mikszáth V12 is unfounded, its deciding design at RS-20260810x §10.

7. Sources consumed

• workshop/regimes/R05-serial-long-work.md — the regime; two of its four declared limitations since measured. RS-20260727-longwork-selection, ARM-longwork, RS-20260728j and RS-20260728d via the arm — Verga.
• ES-20260806-craft-report-koyhaa-kansaa, RS-20260806d-first-span-again — Canth; ES-20260810-craft-report-legend, RS-20260810x via it — Mikszáth; ES-20260813-craft-report-dakghar, RS-20260813d via it — Tagore.
• ES-20260826-craft-report-alf-layla, RS-20260816f-night-seam, RS-20260823b-embedding-carriage, RS-20260826c-register-cost — the Nights.
• ES-20260905-craft-report-gulistan, ARM-gulistan, gulistan-bab2/register.md, collation-chapter.md, collation.md — Sa'di.
• workshop/translations/verter/plan.md, wiki/plan.md W3 steps 5–7; wiki/method-notes.md (bhb), (bjp), (bmw), (bta); RS-20260809i-terminology-drift — checked; X3, family M (§6).
• framework/v0.2/README.md — no numbered section belongs here (grep on long work, R05, cadence, binding register, recension, serial hits only §7.18/§7.22 limits lines — family E).
• Withdrawn, one line each: the 234 as a measure of what the register let move — RS-20260807c §3, RS-20260808a §5; RS-20260806d P1 (≤ 12) — failed at 13, same page; Verga D22's "no available alternative" — RS-20260728j, E7; Verga's monotone contamination trend — ARM-longwork S045; Verga criterion (iii) as a countable class — RS-20260727d; RS-20260810x's three first-draft claims — its own post-hoc critic, §12; D44a's Worswick prediction — ES-20260810 §2.2; Mikszáth P-C band and P-B — ES-20260810 §4; Canth D100 (herra as a voice device) — ES-20260806 §4; Tagore's premise of a pervasive social machinery — scoped to ten sites, ES-20260813 §3; the Nights' P1/P4 (Lane deletes in-text machinery; Burton preserves the formula) — RS-20260823b; RS-20260816f P1 — failed, P = 0.998; span F's "measured rather than recalled" — ES-20260905 §3, note (bta).