Repository path: workshop/experiments/E-20260902-arabic-function/design-frozen.md · rendered 2026-09-09

Is the marking about the language, or about the scripture?

Frozen before any English of these six loci was opened, and after the span D translation was frozen and committed at aa0c9983. No API call is planned; expected cost $0.

The question

RS-20260831-arabic-in-persian counted what four published English hands do where «گلستان» changes language into Arabic and English cannot. Across fifteen loci in حکایات ۱–۳۱, unframed Arabic is marked in 0.543 of present cells — Eastwick at 11 of 12, Ross at 8 of 10, Gladwin at 0 of 12, Arnold dropping 10 of 12. The page's conclusion, and the correction it forced on register rule V5, is that the language change can be signalled and half the tradition signals it, so an arm that declines to mark is declining an available repair rather than recording an impossibility.

That count classified the loci by one thing only: whether Sa'di frames the quotation. It did not ask what the Arabic is for. Every one of the fifteen is scripture, hadith, or an Arabic sentence quoted as a maxim — Arabic in its authority-bearing role, which is the role Persian prose usually quotes it in.

Span D's translating met the other kind. حکایت ۳۳ is the longest tale in the chapter, and it describes the king's garden and the boy who serves in it with two whole Arabic couplets that assert nothing — pomegranate blossom as fire hung on a green tree (b255), and a cupbearer who sees the thirsty and does not pour (b260). No authority is invoked. Sa'di is quoting Arabic verse for its beauty, the way an English writer might drop in a line of Latin poetry to decorate a description.

So the 0.543 may be a mixture of two very different practices, and the question this unit asks is:

Is the typographic mark those two hands supply a signal that says this is another language, or a signal that says this is holy writ?

The answer matters to the translating, not only to the count. If the mark is about foreignness, then V5 declines a repair that is available everywhere Sa'di changes language. If the mark is about scripture, then at b255 and b260 — the places span D actually had to decide — there was no available repair to decline, and V5's corrected sentence is still too strong.

Materials

The six loci RS-20260831 enumerated and then deferred, numbered 16–21 in E-20260831/loci-frozen.md. Its primary was narrowed to حکایات ۱–۳۱ after freezing because the alignment work for these was not finished; their partial codings are in that study's coding.md §Supplement and are not consulted before this study's own coding is complete. Five of the six are in the text span D has just translated, which is why they are reachable now.

The four hands are A-gulistan-hands's, already stored: E-20260831/materials/{GLA,ROS,EAS,ARN}-*-bab2.txt and the ABBYY layer under materials/raw. Italic is one of the four marking signals and _djvu.txt throws font information away, so abbyy_italics.py is the instrument, exactly as in the earlier study.

Cells: 6 loci × 4 hands = 24.

The classification, frozen from the Persian

Each locus is placed in one of three classes by what the Arabic is doing in Sa'di's argument, read from the Persian alone:

The fifteen loci of the earlier study will also be classified this way, from the Persian — but that classification is declared post hoc, because their outcomes are already printed in RS-20260831. It is a secondary partition and will be labelled as one. The primary test is the 24 cells above, whose outcomes nobody has seen.

The confound, named before the run

ORNAMENT is entirely verse and two of the three SCRIPTURE loci are prose, so ornament against scripture is partly verse against prose. Two things bear on it, and both are stated now:

1. RS-20260831 established that Eastwick's own verse is roman — his italic is a dedicated signal and not a verse convention.
2. b315 is the discriminating locus, and it is named in advance. It is verse and authority-bearing. If b315 is marked while b255 and b260 are not, function decides. If all three verse loci behave alike against the two prose ones, form decides and this study reports that instead.

Predictions, registered

1. Across Eastwick and Ross together, at most 1 of the 4 ORNAMENT cells is marked, against at least 3 of their 4 SCRIPTURE cells. (If marking were about foreignness, their base rates — 11/12 and 8/10 — predict roughly 3 or 4 of the 4 ornament cells marked.)
2. Gladwin marks none of his six, as he marked none of twelve unframed.
3. Arnold drops at least 4 of his 6, as he dropped 10 of 12.
4. No hand marks an ORNAMENT locus while leaving a SCRIPTURE locus plain in the same book. This is the within-hand ordering, and it can fail even if prediction 1 passes.

Failure criteria

• If more than 6 of the 24 cells cannot be aligned (25%, the bound the earlier study used), the study reports the alignment failure and reports no rates.
• If Eastwick and Ross between them have fewer than 3 present ORNAMENT cells, prediction 1 is not testable and the page says so instead of reporting a rate on one or two cells.
• A locus a hand drops is coded DROPPED and is neither marked nor plain; drops are reported separately and never counted as "not marked".

Coding

Per locus per hand: MARKED (italic, quotation marks, Arabic type, a footnote naming it as Arabic, or a doubled rendering) / PLAIN / DROPPED / UNALIGNED, with the English quoted verbatim in coding.md so every call is checkable. Same four signals as RS-20260831, so the two counts combine.

internal-judgment-only on every coding: it is the lead's reading of uncorrected OCR, and no second reader will have seen it.