Repository path: workshop/translations/dakghar/collation.md · rendered 2026-09-09

«ডাকঘর» — the copy-text, and the page-image gate

What is new here. ARM-legend's craft report closed with "No page image was consulted. The Canth precedent puts the false-positive rate of an unchecked diff at roughly one in seven, and every collation figure here inherits that" (ES-20260810-craft-report-legend §5). This work's copy-text is a page-image edition: the transcription is a Wikisource proofread transclusion of a scanned PDF whose every page is on Commons. The gate is therefore transcription-against-image, not e-text-against-e-text, and it is the first collation in this project that can be wrong in only one direction.

1. The edition, and what cannot be established about it

No year is printed anywhere in the front matter. The 1958 date is a library accession, not an imprint. Two facts pull in opposite directions and neither settles it: Indian Publishing House is the publisher of the 1912 first edition, which is evidence for an early printing; a bound-in photograph of a staged production is evidence against its being the first impression, since the play was first written in 1912 and the photograph postdates a performance. The year is recorded as not established, and no page in this repository may cite this copy-text with a date.

One witness only. No second free Bengali witness of «ডাকঘর» has been located. The three predecessor long works each had two witnesses and each found the copy-text question live at some span; this arm does not have that instrument and says so. What it has instead is the page image, which answers a different and in one respect better question: not which printing is right but whether the electronic text is what the printing says.

2. The gate on span A — complete, printed pp. 1–11 (PDF 5–15)

All eleven pages of span A were read as images against the transcription. Four divergences, all four transcription errors, none a variant reading; the copy-text file source-ipublishinghouse.txt carries the corrections and this table is the record.

Rate: 4 errors in 1,147 words, 3.5 per thousand, in a text marked Validated. Three of the four are invisible to a reader and the fourth ([34]'s ঠাকুদ্দা) would have been read as a source-side name variant by anyone collating two e-texts — which is exactly the false positive ES-20260810-craft-report-legend §5 warned the record was carrying. It is a transcription slip, and only the image could say so.

One reading not resolved. At [15] the transcription has চুপ করে সহ্য করে; the image at this resolution does not distinguish সহ্য from সহ in this face. Left as transcribed, flagged, and to be settled at a higher resolution when span B runs its gate.

3. What the gate found in the source rather than in the transcription

Three naming variations are in the print, confirmed at the image, and are carried into the English under V4 (translation log D7, D10):

1. মাধবদত্ত → মাধব. The speech-head is the full name for the first eleven heads (printed pp. 1–3) and the bare given name from p. 3 to the end of the play. The switch falls at [15].
2. অমল → অমলবাবু, once, at [55].
3. পিসে মশায় (4×) / পিসেমশায় (12×) — printed both ways. Word-division, not lexis; recorded, not rendered.

And one probable misprint, carried verbatim and not emended: the stage direction at [38] reads ( অমলগুপ্তের প্রবেশ ), giving the child a surname that appears nowhere else in the play. A one-witness undated copy-text is not a text this project may silently correct.

4. Standing rule for this arm

The copy-text gate is run on every span, against the page images, before that span is translated, and its divergences are recorded here before the freeze. This is the predecessor arm's practice (ARM-legend ran a copy-text gate four times and it found something wrong with the record all four times) carried over to the better instrument.

5. The gate on span B, stage one: what the file was missing (2026-08-12)

The first thing the span-B gate found was not a transcription error but a hole in this repository's own copy of the text. Ten of the span's 174 speeches were empty in source-ipublishinghouse.txt — a speech-head with nothing after it — and every one of the ten is the last speech printed on a page.

They are not missing from Wikisource. Fetched page by page from the পাতা: namespace, all ten are present and complete. They were lost by this project's extraction at S160, and the loss is silent: nothing in the file marks it except the empty head, and build_index.py has no view of this file at all.

101 Bengali words, 4.7% of the span. The span's true extent is 2,140 words — the ten speeches add 101 and four of §6's emendations join words the transcription had split, so the arm page's declared 2,043 is 97 short, not 101. Span A has no empty speeches; section ৩ has sixteen, and they must be restored and gated before span C is translated.

Two of the ten are load-bearing for the scene rather than incidental: [127] is the watchman's what if I seized you and carried you off?, without which Amal's next line has no antecedent; and [242] is the boys' whole reaction to the toys, the emotional centre of that exchange.

6. The gate on span B, stage two: transcription against the page image

All 22 pages (PDF 16–37, printed 12–33) read as images against the restored transcription. Twenty-six emendations in twenty-one speeches. Three matter to the sense:

Carried verbatim, because the print has them: [173] শক্রতা for শত্রুতা; [250] কন for কেন; [253] তোমারা for তোমরা. A one-witness undated copy-text is not silently mended (§1).

Three readings the gate could not settle, all in a margin the scan loses — the gutter on printed pp. 14, 16, 18, 20 and the outer margin on p. 33: [136] পিসিমা রা …, [140] যে দেশের জানে না, [259] কে চিনিয়ে দিয়ো না. Each is flagged on the translation at D34 and none is emended.

7A. The gate on span C, stage one: fourteen lost speeches, not sixteen (2026-08-12)

Span B's forecast was that "section ৩ has sixteen" empty speeches (§5). Sixteen speech-heads in the file are indeed empty; only fourteen of them are losses. The whole of span C — PDF 38–60, printed 34–56 — was refetched from the পাতা: namespace (Bengali-numeral page titles) and aligned against the file speech by speech: 166 upstream blocks against 170 file units, the four extra being (cont.) units where the file splits a trailing stage direction off its speech, which the print does too.

[344] and [360] are empty in the print as well. Both are a speech-head (ঠাকুর্দ্দা) followed by a centred stage direction on the next line — (অমলকে বাতাস করিতে করিতে), (তাড়াতাড়ি উঠিয়া দাঁড়াইয়া) — which the file carries as its own unit with the speech in the (cont.) after it. The file is right and the forecast counted layout as loss. A count of empty heads is not a count of missing text, and the only thing that could tell the two apart was going back to the page.

The fourteen real losses, all restored:

234 Bengali words, 9.4% of the span — twice span B's rate (4.7%), and the same signature: every one is the last speech printed on its page. The span's true extent is 2,501 Bengali words against the arm page's declared 2,268.

Three are load-bearing. [343] is Amal's the dark keeps coming down over my eyes… everything feels like a dream — the turn where the play stops being about waiting and starts being about dying; without it Grandad fans a boy who has said nothing. [403] is I can see all the stars — all the stars on the far side of the dark, the line the whole third section builds to. [409] is Madhab whispering to the dying child to ask the King for money.

7B. The gate on span C, stage two: transcription against the page image

All 23 pages read as images against the restored transcription. Forty-five emendations in thirty-eight speeches. Seven change the sense:

Nine sites where the transcription drops the honorific তাঁ-: [306], [314], [315], [372], [373], [380], [400], [404], [409] — every one of them a pronoun for the King or for the fakir, and [400]/[409] are inside speeches restored at stage one. This is the span's largest single class and it is not orthographic: তার/তাঁর is the play's only surviving grammatical mark of deference once আপনি has gone (register.md V5), so a transcription that flattens it erases the one signal the register instrument is watching.

The rest: [354] সুর্যাস্তের for সূর্য্যাস্তের; [356] সেইটােতেই, a malformed vowel cluster that is not a Bengali word; [404] অৰ্দ্ধরাত্রে typed with Assamese ৰ (U+09F0) for Bengali র; [340] ষে for যে; [371] হাহাহাহ for হাহাহাহা; [372] ঠাট্ট for ঠাট্টা; [319] ভিক্ষ for ভিক্ষা; [265] ছেলেবুড়ে for ছেলেবুড়ো; [379] রাজকবিরাজকে ও for রাজকবিরাজকেও; six em dashes flattened to - or – ([265], [283], [291], [311], [355], [375], [417]); and six places where the print's ।/! were exchanged ([283], [293], [298], [353], [363], [373], [405]) — including [298], where the transcription is the more emphatic of the two.

Restored to the print, against the transcription's modernisation: [307] and [406] পরিক্ষার, which the transcription normalises to পরিষ্কার at both sites; [331] ফুরচ্চে and [350] হচ্চে, normalised to ফুরচ্ছে/হচ্ছে. The play's চলিত is -চ্চ- throughout and the transcription is inconsistent about it.

One class recorded and deliberately not rendered: the hasanta. The print marks the চলিত verb stem with ্‌ — আস্‌বে, দেখ্‌তে, শুন্‌চি, থাক্‌লেই, বল্‌ব — and the transcription drops it at roughly forty sites in this span while keeping it at others. No reading in the play turns on it, and hand-applying forty diacritics to the file the translation is made from would risk introducing exactly the errors this gate exists to remove. It is counted here and left as transcribed. Span A's and span B's tables did not name this class; it is present in both.

Two readings the gate could not settle, both where the scan loses the gutter or the ink is faint: [400] তাঁর কালের চেয়ে (printed p. 52, inner margin) and [430] বলো/বোলো (printed p. 56, the faintest page in the book). Neither is emended; both are flagged at D48.

7. A correction to §2's own record

§2 described this copy-text as "a text marked Validated". That is an Index-level claim and it does not hold page by page. Read from each page's pagequality level:

Span A's rate of 4 errors in 1,147 words, 3.5 per thousand was therefore measured on a mostly validated stretch. Spans B and C are mostly proofread only, their error rates are higher, and the rates are not comparable and are not compared here. What the figures do jointly support is the standing rule: the gate runs on every span, and the instrument that catches the worst errors is the page image, not a second e-text.