Repository path: workshop/translations/malavoglia-i/R23-v1/translation.md · rendered 2026-09-09

T-malavoglia-i-R23-v1 — Verga, «I Malavoglia», opening of Chapter I: the low pole crossed by device

Source. As T-malavoglia-i-R06-v1, -R21-v1 and -R22-v1: I Malavoglia (1881), Capitolo Primo, opening through «chi comanda ha da dar conto», same copy-text file (workshop/experiments/E-20260808e-low-pole/arms/source_it.txt), 595 Italian tokens.

Regime. R23 — device-crossed low, frozen this session. Two switches, four cells.

Why this is one artifact and not three. The four cells are a single manipulation, not four translations: ∅ is a rendering and A, B, AB are minimal revisions of it under R23 §Procedure. Filing them apart would invite a reader to treat them as independent hands, which they are not and must never be read as. The cell texts sit beside this file as cell-A.txt, cell-B.txt, cell-AB.txt; cell ∅ is T-malavoglia-i-R22-v1, unaltered, and is reproduced here only as cell-null.txt for diffing.

That the two +I cells are the two long ones is the first thing the quadruple measured, and it was not looked for: located English is longer than its placeless equivalent at these sites, by 2.9% over the span. Was taken for a man with a good head → had the name of a man with his head screwed on; was nobody's fool → knew how many beans make five; a small thing → a little bit of a thing. The device that reaches the register also lengthens, which is Berman's allongement turning up as a property of the means rather than of the translator. Recorded, not chased.

Contamination: high, declared, and one of the three bases is new

Three bases, in descending size:

1. ∅ is the lead's own prior artifact and the other three cells are edits of it. This is R23's procedure, and by design: byte-level identity between cells is the instrument. The quadruple is one translator's four passes and no cell is an independent opinion of anything. Note (bhb)'s standing figure for the lead against itself — 37 contiguous tokens, revised to 43 by R22~R21 — is not a comparator here, because these cells are meant to be identical outside the touched sites.
2. R22's own declared exposure carries in. T-malavoglia-i-R22-v1 records, as a violation of continue-prompt.md §5.2, that Craig 1890's English at all seventeen census sites had been read before it was written. Everything derived from ∅ inherits that exposure.
3. New this session, and small: while locating the source file the lead read the first ~800 characters of craig1890_span.txt — Craig's English for the first paragraph, covering the region of sites IT-01 to IT-03. Reading stopped there and no further Craig was opened before or during writing. Declared because it happened, and measured below.

Consequence, fixed before any measurement and not revisited after one: the lead's quadruple is a secondary human replication and never a primary arm. The primaries of E-20260809g are written by two panel models from the source and the rule sets alone, blind to the hypothesis, to this file, and to each other. What this quadruple is for is the log — what each device can and cannot reach — and a log is not a score.

The measurement, run immediately after the draft was fixed and before the design existed

tools/dependence_check.py, per CLAUDE.md's measure before choosing, run on the frozen cells before E-20260809g/design.md existed. Full output in dependence.json beside this file.

Basis 3 leaves no measurable trace. Craig's opening paragraph was read this session and the three cells share no twelve-gram with him and no run above 8 tokens — below every threshold the project uses. The declaration in basis 3 stands as declared, because the reading happened; the quantity it might have hidden is measured and is nil. That is the fourth time a contamination: label and its measurement have pointed different ways, and the standing rule holds: report the number, not the fear.

The 51-token within-quadruple run is not a note (bhb) datum and must not be filed as one. Note (bhb) counts the lead matching itself across independent renderings; these cells are edits of one text and are supposed to be identical outside the touched sites. Any future high-water mark for lead-against-itself independence is still 43 (R22~R21).

The four cells

cell-null.txt · cell-A.txt · cell-B.txt · cell-AB.txt, in full, beside this file. The opening paragraph of each, so the manipulation is visible without opening four files:

∅ — Once there were Malavoglia thick as stones on the old Trezza road; … The parish book had them down as Toscano, but that meant nothing, because since forever, at Ognina and Trezza and Aci Castello, they'd been the Malavoglia … Now the only Malavoglia left in Trezza were skipper 'Ntoni's, the medlar-tree house ones …

A — … The parish book had them down as Toscano, but that meant nowt, because since Adam was a lad, at Ognina and Trezza and Aci Castello … the medlar-tree house lot …

B — … The parish book had 'em down as Toscano, but that meant nothin', because since forever, at Ognina an' Trezza an' Aci Castello … all o' 'em good sea folk …

AB — … The parish book had 'em down as Toscano, but that meant nowt, because since Adam was a lad … all o' 'em good sea folk … the medlar-tree house lot …

Translator's log

Frozen with the cells, before E-20260809g/design.md existed. Coverage codes per R23 §Procedure: P predicted by the rule set · D deviation with reason · INT decided by interpretation · L a loss the rule set caused.

1. The thing the crossing made visible before anything was measured

R22's log said the placeless shelf is "a register of worn idiom and short words — it can be plain but it cannot be fresh". R23 was built to ask which half of the banned clause bought the freshness. Writing the cells answered a different and blunter question first, and the answer is arithmetic rather than judgement:

The respelling device cannot be aimed. At the eight independently voted Italian narration sites, cell B differs from cell ∅ at three, and at two of those three the only difference is the coordinator and → an'. Cell A differs from ∅ at eight of eight.

The reason is structural and it is a fact about English rather than about this passage. Verga marks his narration lexically and idiomatically — bighellone, testa quadra, codino marcio, la sapeva lunga, nè carne nè pesce. Respelling attaches to function words and to the -ing ending, whose distribution in an English sentence has nothing to do with where its source marked register. So the two devices are not two ways of doing one thing. The located idiom lands on the marked site; the respelling lands wherever an -ing or an and happens to be, which at these sites is mostly nowhere.

If that is right, it is worth more than the register measurement the run was built for, because it says the cheap device is not a substitute for the expensive one at all — it is a different device that reaches different places. internal-judgment-only; the model arms in E-20260809g are where a hand that has not read this paragraph gets asked to find the reach.

2. Cell A, site by site — what the located shelf actually offered

The eight voted narration sites, plus the four non-site changes +I reached. The ∅ reading is R22's and is not re-litigated.

Two of these are INT and both are the same worry. Neither nowt nor summat and a rotten old Tory do not merely sound low; they sound like a specific country's specific class and a specific country's politics, and W8 forbids the parodic. A Sicilian fishing village whose patriarch is nearly made a councillor until the clerk calls him a Tory is, to an English ear, a joke the Italian does not make. I kept them, because the cell exists to show what +I costs and softening it would hide the cost — but the cost is here, in writing, before anyone rated anything. INT.

A reached every one of the eight sites, and the shelf was never empty. That is the asymmetry with B in §3 and it was not engineered: I did not choose the sites, and the three annotators who voted them (RS-20260808e §4) were not asked about devices.

3. Cell B, and the reach of respelling

+S was applied by a stated policy, not ad hoc, so that the census is checkable:

• -ing → *-in' on verbal and nominal -ing forms of two syllables or more: doin', rowin', weavin', saltin', havin', gettin', sayin', sayin's, goin', spittin', plottin'. Not on thing, things, bring*.
• and → *an' throughout, of → *o' throughout.
• object them → 'em (two instances).
• nothing → *nothin'*.

What was available and declined, and this is the decision the cell turns on. English's other standard respellings are h-dropping ('is, 'im, 'ave) and the t' article. I declined both. The reason is R23's own boundary: h-dropping and t' are not neutral respellings of ordinary speech, they are the orthographic signature of particular English speech — London and the North respectively — and admitting them would have made B a +I cell wearing an apostrophe. So B is the best case for the framework's assumption, deliberately: it is the most placeless respelling a translator could manage. If B still reads as located, the assumption fails at its own best case. INT, and it is the INT that matters most in this file.

The reach, counted rather than estimated. Across the whole span B differs from ∅ at 47 tokens, of which 45 are respellings (an' 25, o' 5, -in' forms 12, 'em 2, nothin' 1 — and more than half of the whole device is one word, the coordinator and). Across the eight voted narration sites it differs at three:

Five of eight sites are untouched and one more is touched only at a conjunction. L — and it is a loss the device causes, not the rule set.

4. Cell AB, and the one place the two devices interact

AB is A with §3's policy applied. It is not simply the union: three of A's located choices create respelling opportunities that ∅ did not have — had the name of a man → o', dead spit of his grandad → o', a little bit of a thing → o' — so +I increases the reach of +S. Counted: o' rises 5 → 7 from B to AB (three new, one lost with the substituted phrase), and at IT-10 the respelling exists in AB and cannot exist in B at all.

The interaction runs one way only. Nothing +S does opens a site to +I. If a later design wants an additive model it should know that the design is not symmetric, and this is the reason. P, and it is the cleanest thing the quadruple found about the two devices together.

5. W6, the addition rule

Nothing was supplied in any cell. The three cuts R22's log records (help each other out, plenty more, rotten old nothing but a) stand cut. +I invited two supplies that were refused: a right good housewife (right is an intensifier the Italian's buona does not carry) and knocked all over the shop, poor devils. Both cut before freezing. P.

6. W3, and the deviation

A and AB come in at 612 against 595. Under R23's minimal-pair clause the excess is reported rather than cut, because cutting it would mean editing wording the switch never touched. The whole excess is attributable to eight located substitutions being longer than what they replaced; the largest single contributor is IT-12 (was nobody's fool, 4 → knew how many beans make five, 6). D, with the reason.

7. Losses

• L1 — B cannot reach five of eight marked sites (§3). This is the finding, not a defect of execution.
• L2 — A puts the book in England at two sites and possibly at all eight (§2's INT).
• L3 — every loss R22's log records at ∅ (L2–L6 there: la paranza, the doubled reazionario, the mullet proverb's rhyme, Sicily, the proverbs not being English proverbs) carries into all four cells unchanged, because no switch reaches them.
• L4 — the cells cannot show what a translator would have written from scratch under +I or +S. They show what a placeless rendering becomes when a device is permitted. That is the right instrument for pricing a device and the wrong one for describing a programme, and the E-20260809g model arms are built the same way for the same reason.