Repository path: workshop/experiments/E-20260823b-embedding-carriage/design.md · rendered 2026-09-09

E-20260823b — a tale three deep with nothing to mark the depth: what carries the level structure into English?

Track T1, ARM-alf-layla step 8. Study limb of span H (the copy-text's fifth night, pp. 37–42). This design and its locus table are frozen from the Arabic before the span is translated and before any English rendering of it is opened. The locus enumeration below was made from ../../translations/alf-layla/source-hindawi-2022-spanH.txt alone.

1. Where the question comes from

The fifth night is the first place this work stacks a tale inside a tale inside a tale. Its architecture, read off the Arabic:

Three tellings are nested inside the book's own telling, and one is named and withheld.

The copy-text marks none of it typographically. There is one heading in the span (فلما كانت الليلة ٥, a night heading, not a tale heading); there is no rule, no break, no indentation and no quotation mark at any of the five seams above. V4 of the binding register already refuses quotation marks for speech, and V10 refuses a supplied heading at an embedded tale — but V4 and V10 were written at spans A and C, where the nesting was two deep. This is the first span in which obeying them means asking an English reader to hold three levels at once with nothing on the page to help.

So the translator's problem is real and immediate, and the question the study limb asks is the question the translating raises: when the source marks narrative depth only from inside the sentences, what reaches English — the source's own in-text machinery, or apparatus the translator supplies?

The Arabic's in-text machinery is of three kinds, all enumerable before any English is opened:

• Entry and exit formulae — وما حكاية …؟ / وكيف كان ذلك؟ to request a tale, هذا ما كان من حديث … to close one. The copy-text uses them unevenly, which is itself the span's most interesting formal fact (see L4 and L5 below).
• Vocative intrusions — the level-2 teller breaking into his own level-3 tale to address his level-2 hearer: مثل ما قلت لك أيها العفريت, and, wedged between subject and clause, فلما تحقَّقَ الحكيم — أيها العفريت — أن الملك قاتِلَه.
• Verbal identity across levels — the plea أَبْقِني يُبْقِيك الله، ولا تقتلني يقتلك الله, spoken by the fisherman to the ifrit at level 2 (span G, p. 33) and by Ruyan to Yunan at level 3 four times in this span, and turned in the closing sentence back onto the ifrit. The embedding exists because of this sentence: the fisherman tells Yunan's story precisely to prove what happens to a king who hears that plea and refuses it. If the sentence is not the same sentence in English at both levels, the reason for the nesting is invisible on the page.

2. The registered hypothesis

H1. The programme difference measured on the night boundary at RS-20260815f-night-formula reproduces at the tale boundary. There it was found that Burton marks all 33 of his nights and writes the closing formula in two near-identical forms (similarity 0.9428), while Lane has no night formula at all and replaces the nights with eight chapter headings that name them from outside the story. H1 says this is one habit and not two: Burton keeps the source's in-text machinery and supplies little; Lane deletes the in-text machinery and supplies apparatus in its place.

H2. A hand that reaches the work through a second language rewrites the architecture more than either hand that reaches it through Arabic. Forster 1802 renders Galland's French, not an Arabic recension; the transmission has one more editor in it and Galland's editorial hand on the frame is the best-known fact about his version.

Both hypotheses can fail, and the direction of failure is informative in each case.

3. Materials

Four hands, three transmission lines:

LANE and BURT are not independent witnesses — Burton read Lane, and the project has already measured two hands on one author reading each other (RS-20260822c-persian-hands). FORS is independent of both in transmission and dependent on neither in wording, which is why it is here. Payne 1882 is not freely reachable and was searched for at span A.

Conditional secondary check, declared in advance. If FORS diverges from both Arabic-line hands at a locus, the divergence is attributed to Galland only if Galland's French (freely readable, fr.wikisource) is consulted at that locus. Otherwise it is reported as Galland-line, not as Galland's, and the ambiguity is stated.

Second Arabic witness. The same loci are checked against witness 2 (the older, unrelated Arabic Wikisource transcription) as part of this span's ordinary collation. A device present in one digitisation and absent from the other is an edition's property; a device in both is the tradition's.

4. The locus table — frozen from the Arabic, before any English for it exists

Ten loci. Every one was located by reading the Arabic source file for span H (and, for F1, span G, already rendered). The classification column says what kind of level-machinery the locus is, and is fixed here.

The cross-level formula, F1–F6, enumerated separately because it is scored differently (§5.3). All six are one Arabic sentence, or a transformation of it:

5. Procedure

Order, and it is a gate. §4's table is frozen at a commit before the span is translated. The span is then translated and its log frozen at a second commit. Only then is any English rendering of this span opened. The lead has read Lane and Burton on spans A–G and is primed on their manner; it has never opened Forster. The span-G contamination event (five lines of Burton printed by a grep) is on the record and its refusals stand; no comparable event may occur here, and the extraction in §5.1 is written to write to files without printing.

5.1 Locating each hand's text

Each hand is located by searching its volume for the tale, not for the locus, so that the search string cannot select the answer: LANE/BURT/FORS are cut at the start of the Tale of the Fisherman and the Ifrit and at the end of the Tale of King Yunan (Douban / Doobán / the corresponding name in Forster), and only the cut span is read. Where a hand's night-division does not correspond to the copy-text's (it will not, for LANE and FORS), the tale, not the night, is the unit, and that is recorded.

5.2 Scoring L1–L10

Each locus × each hand gets two independent codes, both applied by the lead reading the hand's text, and both defined here:

(a) SOURCE-MACHINERY — is the source's own in-text device present in this hand's English?

• KEPT — the device is there and does the same work (a request formula is a request formula; a vocative address to the level-2 hearer is one).
• WEAKENED — present but doing less (e.g. a vocative reduced to a bare O king with no address to the outer hearer; a closing formula that no longer names its tale).
• DROPPED — absent.
• N/A — the locus's material is not in this hand's text at all (a recension or abridgement gap), which is recorded and never scored as DROPPED.

(b) SUPPLIED-APPARATUS — what does this hand put on the page that the copy-text has not got, at or within two sentences of this locus?

• NONE
• TYPO — typographic only: quotation marks, a paragraph break, a rule, italics, indentation.
• HEADING — a supplied title, chapter head, or running head naming the embedded tale.
• NOTE — a footnote or endnote.
• TEXT — supplied words inside the narrative itself (e.g. continued the fisherman, said the vizier to the king) that the Arabic has not got.

More than one code may apply; all are recorded. TEXT and HEADING are the two that matter to H1, because they are the two that put the level structure somewhere the Arabic did not put it.

5.3 Scoring F1–F6

For each hand, the six renderings are extracted verbatim and compared to each other, giving a mean pairwise similarity over the five occurrences the hand has (F5 is scored on its own half; F6 is excluded from the similarity and reported separately because the Arabic itself transforms it). Similarity is character-level difflib.SequenceMatcher ratio after lowercasing and stripping punctuation — the same measure RS-20260815f-night-formula used on the night formula, so the two numbers are comparable. A hand that renders the plea identically at levels 2 and 3 keeps the architecture's joint; a hand that varies it does not.

The cross-level pair that matters most is F1×F2–F4 — level 2 against level 3 — and it is reported separately from the within-level-3 pairs.

5.4 Verification

Every count on the result page is recomputed by a script in verify.py that reads the extracted per-hand texts and the coded table, and re-derives: the number of loci per code per hand, the similarity matrix for F1–F6, and the N/A list. A count that the script cannot reproduce is struck from the result page.

6. Predictions, registered

Written before any English is opened. Each is scored HELD / FAILED / UNTESTABLE and reported as it stands.

1. P1 (H1, in-text machinery). BURT scores KEPT at more of L2, L3, L6, L7, L9 than LANE does. Fails if LANE ≥ BURT.
2. P2 (H1, supplied apparatus). LANE scores HEADING or TEXT at more of L2, L4, L5, L8 than BURT does. Fails if BURT ≥ LANE. Declared prior knowledge, and it weakens this prediction: the lead knows in general terms that both Lane and Burton annotate heavily and that Burton supplies tale headings. P2 is therefore registered as a genuine test only of the TEXT and HEADING codes at the four seam loci, not of annotation in general.
3. P3 (the unmarked seams). At L4 and L5 — the two places the Arabic changes level with no formula at all — at least three of the four hands supply something (TEXT, TYPO or HEADING). Fails if two or fewer do. This is the prediction the lead's own V10/V4 rendering can falsify from the inside: if LEAD supplies nothing there and the sentence still parses, the English does not need the help.
4. P4 (cross-level identity). BURT's mean pairwise similarity over F1–F5 exceeds LANE's. Fails if LANE ≥ BURT. Registered because RS-20260815f measured exactly this habit on a different formula and found Burton at 0.9428 with no drift.
5. P5 (H2, the Galland line). FORS scores DROPPED or N/A at more of L1–L10 than either Arabic-line hand. Fails if FORS ≤ min(LANE, BURT).
6. P6 (the refused tale). L8 — the Tale of the Crocodile, asked for and refused — is present in all four hands. Fails if any hand omits the exchange. Registered because a tale that is never told is the easiest thing in the span for a translator to think redundant, and because keeping it is the clearest evidence a hand is following the source's architecture rather than tidying it.

7. Failure criteria, registered

• The census is void for a hand if its text of the tale cannot be located unambiguously, or if its recension gap covers more than three of the ten loci. A void hand is reported, not silently dropped.
• P4 is void if any hand has fewer than three of F1–F5, since a mean over two pairs is not a measure of a habit.
• No result is reported as an effect of a hand's declared programme unless it holds at three or more loci. Two loci is an illustration and is labelled one.
• Nothing here licenses a claim about which rendering is better. The codes are descriptive; the senses this design names (style-correspondence, consistency) are the senses the description is about, and no evaluative claim is made without an anchor citation or an internal-judgment-only flag.

8. What this cannot establish

• Four hands are at most three witnesses, and probably fewer: Burton read Lane, and Forster renders Galland. Every number here is a description of four books, not a sample of English translation practice.
• No reader is measured. Whether an English reader can in fact hold three levels without apparatus is a different question, and it needs readers this project does not have. It is named here and not answered; nothing on the result page may be phrased as if it had been.
• The copy-text is one recension. Where a hand has no material at a locus, the honest verdict is N/A, and N/A is permanently ambiguous between the translator cut it and his Arabic did not have it — the same ambiguity span A's locus 8 carries.

9. Budget

$0.00. There is nothing here for a model to judge: the question is answerable exactly, on four published texts, by reading them at loci fixed in advance. The pre-run critic that this project buys for a judgment run buys nothing for a census, and RS-20260814-saj-carriage (span A's study limb, the same shape) is the precedent.

What the money would have bought, and why it is not being spent, so the choice is on the record: a reader-side run asking whether the level structure is recoverable from each hand's English. Note (brb), written yesterday, says a direct identification task on these seats is a ceiling and not a measurement, and who is speaking here is exactly that shape. The indirect form that would work — a content question whose right answer depends on knowing which level one is on — is a real design and is not one that can be built honestly in the tail of a session whose translation limb is 1,769 Arabic words and eight poems. It goes to the arm as a named step, not into this design as an afterthought.