endomorphosis/ipfs_cambodia_laws

IPFS Cambodia Laws

Structured corpus of Cambodia national laws and legal instruments.

This dataset follows the same schema used by the IPFS Datasets legal corpora:

• data/laws.parquet: one row per law / legal instrument with full text.
• data/articles.parquet: one row per article / section (law_id joins laws.id); this is the default config.

Wave 10 build (2026-10-13): additive merge by instrument ID (longest-text-wins) of the previously published corpus with newly collected instruments (FAOLEX via Wayback CDX enumeration + live FAO PDF hosts, and WIPO Lex), with a normalized-title duplicate guard.

Counts

• laws_rows: 274
• articles_rows: 12027

Schema (data/laws.parquet)

Columns: id, title, text, source_url, source_type, jurisdiction, country, language, eli, date, date_issued, retrieved_at, license, law_status, identifier, official_identifier, article_count, json_path, metadata_json

Types: all text columns large_string; article_count int64 (= number of rows in data/articles.parquet for that law).

Schema (data/articles.parquet)

Columns: law_id, id, title, text, source_url, document_number, article_number, record_type, metadata_json (all string).

Fix 2026-10-16

Batch fix-20261016-kh-v5 on parent 91a0f98caecd00c45ed421c5ad513d67aef31ca2. Every id is listed in manifests/removed_ids_20261016_v5.json and manifests/changed_ids_20261016_v5.json.

• Laws 155 -> 50: 105 rows that are not Cambodian legislation removed, each checked by source path and content:
  • justice library sample judgement or civil form: 45
  • justice library sample divorce: 11
  • un sanctions list notice: 1
  • un security council resolution: 11
  • misc form presentation letter textbook: 12
  • non law personal material form: 1
  • faolex policy strategy plan: 24
• Articles 8,301 -> 8,002:
  • 26 rows of removed laws;
  • 218 table-of-contents entries parsed as articles (202 heading + page-number rows, 14 contents rows sharing an id with the real article body, which now has the plain id, and 2 standalone contents rows);
  • 51 heading-only rows whose body could not be found in the law text (0 heading-only rows got their body back from the law text);
  • 4 exact duplicates.
• Contents text cut from 6 article rows that also hold real body text (rows kept):
  • kh-e1-9e-9f-e1-9f-81-e1-9e-85-e1-9e-80-e1-9f-92-e1-9e-8f-e1-9e-b8-e1-9e-80-e1-9f-86-e1-9e-8e-e1-9e-8f-e1-9f-8b-e1-9e-85-e1-9f-86-e1-9e-96-e1-9f-84-e1-9f-87-e1-9e-article-84: contents head cut (body kept) (170,488 -> 170,419 chars; every cut line classified as a contents entry, heading, page number or OCR fragment)
  • kh-law_007_0511_en-article-84: appended contents block cut (2,130 -> 1,539 chars; every cut line classified as a contents entry, heading, page number or OCR fragment)
  • kh-law_007_0511_kh-article-78: appended contents block cut (20,452 -> 12,714 chars; every cut line classified as a contents entry, heading, page number or OCR fragment)
  • kh-law_030_1207_kh-article-1303: appended contents block cut (18,325 -> 1,043 chars; every cut line classified as a contents entry, heading, page number or OCR fragment)
  • kh-e1-9e-94-e1-9f-92-e1-9e-9a-e1-9e-80-e1-9e-b6-e1-9e-9f-e1-9f-a9-e1-9f-a8-20-e1-9e-9f-e1-9f-92-e1-9e-8f-e1-9e-b8-e1-9e-96-e1-9e-b8-e1-9e-93-e1-9e-b8-e1-9e-8f-e-c1e2f5ed-article-22: mid-row contents block cut (body, signature and annex kept) (26,401 -> 23,426 chars; every cut line classified as a contents entry, heading, page number or OCR fragment)
  • kh-prakas_098_1209_moj_jp-article-22: mid-row contents block cut (body, signature and annex kept) (26,157 -> 23,166 chars; every cut line classified as a contents entry, heading, page number or OCR fragment)
• The table-of-contents gate decides by content only (dot-leader lines ending in page numbers, incl. full-width and Khmer digits); it scans every whole row (start, middle and end), and every block it cuts is checked line by line before it is cut.
• 485 article rows that share an id with a different body keep their text and get <id>--<sha8>, where sha8 = sha256((law_id + "|" + title + "|" + text).encode("utf-8")).hexdigest()[:8] computed over that row's published law_id, title (empty string if null) and text (not sha256(text)). The first body of each id, in source order, keeps the plain id.
• article_count recomputed for 18 laws; it now equals the article rows of each law.
• HTML entities: 0 rows changed. Email redactions: 0. Institutional addresses printed in official documents are kept.
• Parquet column types unchanged from the previous release.

Known gaps

9 laws have no article rows (full text in laws.text); each was checked by batch fix-20261017-kh-live51-split and has no article headings to split:

• kh-021_hang-20jun-20lun-20--20khmer-20translation: English 'Unofficial Translation' of a Ministry of Justice forwarding order and a police report on a property freeze (letters and a table); no article headings.
• kh-4.guideline-20for-20implementing-20mla-20request: Ministry of Justice guideline (letter with lettered points A-D); no article headings.
• kh-e1-9e-98-e1-9f-81-e1-9e-9a-e1-9f-80-e1-9e-93-e1-9e-80-e1-9f-92-e1-9e-9a-e1-9e-98-e1-9e-9a-e1-9e-8a-e1-9f-92-e1-9e-8b-e1-9e-94-e1-9f-92-e1-9e-94-e1-9e-9c-e1-9f: Textbook (lessons with questions and answers) on the Civil Code and the Code of Civil Procedure; not article-structured. Its heading-like lines are quotations of code and constitution articles inside the commentary (e.g. civil code articles 1-5, constitution articles 14-17), not articles of this document.
• kh-e1-9e-9f-e1-9e-b6-e1-9e-81-e1-9e-b6-e1-9e-89-e1-9e-b6-e1-9e-8f-e1-9e-b7: Kinship chart (table of relatives by degree); no article headings.
• kh-e_book_guideline: Ministry of Justice manual for the electronic document-management system (numbered sections of a user guide); no article headings.
• kh-faolex-cam1073: OCR text layer of a scanned two-column English translation (Proclamation on aquaculture licences, 1989) that interleaves the columns: the article 4 and 5 headings sit on adjacent lines, the body of article 4 follows the article 5 heading and the article 1, 3, 6 and 8 headings are illegible; splitting would put text under the wrong article. Left as whole-law text.
• kh-e1-9e-9f-e1-9f-81-e1-9e-85-e1-9e-80-e1-9f-92-e1-9e-8f-e1-9e-b8-e1-9e-96-e1-9e-93-e1-9f-92-e1-9e-99-e1-9e-9b-e1-9f-8b-e1-9e-80-e1-9f-92-e1-9e-9a-e1-9e-98-e1-9e-9ee5c0bd (Japanese commentary on the Code of Civil Procedure, 224 rows): two-column pages whose text layer interleaves the columns line by line; 15 article headings stay inside other rows. A column-wise re-extraction needs the official PDF, which is not available (the official site answers with a block page, treated as a robots.txt disallow).
• Khmer structural heading lines at row ends in legacy Limon encoding: fixed in the bilingual commentary (batch fix-20261017-kh-headings) and in the 10 other Limon laws (batch fix-20261017-kh-limon-tails, 438 rows); no row of a Limon-encoded law ends with one now; 5 rows end with a heading at the foot of a page followed by the page number, not cut (page_foot_headings_not_cut).
• kh-faolex-cam161033: Prakas No. 099 MAFF on Good Agricultural Practice: the text layer is a Limon-encoded title page followed by unreadable OCR (Latin glyph noise); no legible article headings.
• kh-instruction_002_1208_moj_kh: OCR text of a one-page Ministry of Justice instruction letter (numbered points ១., ២.); no article headings.
• kh-sub_240_0814_ank_bk_kh: The text layer holds only repeated page headers ('05/04/20', 'www.moj.gov.kh Docs supported by pheakdeylaw.com'); there is no body text to split.

The other 15 laws of the previous list now have article rows (see the section 'Fix 2026-10-17 (live-51 article split)').

Borderline documents kept (official Ministry of Justice publications, but administrative guidance rather than legislation):

• kh-4.guideline-20for-20implementing-20mla-20request: Ministry of Justice guideline on handling mutual-legal-assistance requests (administrative guidance, not legislation)
• kh-e_book_guideline: MoJ guideline for the electronic document-management system (administrative manual, not legislation)

Article rows that held more than one article are split (batch fix-20261017-kh-live51-split): kh-e1-9e-9f-e1-9f-81-e1-9e-85-e1-9e-80-e1-9f-92-e1-9e-8f-e1-9e-b8-e1-9e-80-e1-9f-86-e1-9e-8e-e1-9e-8f-e1-9f-8b-e1-9e-85-e1-9f-86-e1-9e-96-e1-9f-84-e1-9f-87-e1-9e (82 rows for articles 1-84, -article-1 ... -article-84; articles 20-21 are not in the text; the false split -article-85 is removed), kh-law_007_0511_kh (articles 2-84) and kh-law_030_1207_kh-article-1303 (article 1304 now has its own row). The 9 other laws in legacy Limon-encoded Khmer (including kh-law_030_1207_kh) now have one row per 'maRta N>-' heading (batch fix-20261017-kh-merged-split, see the section 'Fix 2026-10-17 (merged headings and Limon split)').

OCR batch ocr-20261017-kh-pressocm (pressocm.gov.kh scans): the 226 instrument posts held for OCR were OCR'd (tesseract 5.5.0, Khmer model); 223 were added (see the OCR section below), 2 have too little usable text and 1 duplicates another post; per-post status in manifests/ocr_status_pressocm.json (the deepen batch's manifests/held_for_ocr_pressocm.json is unchanged). OCR text is machine-read: expect character errors (see the OCR section); the official scans prevail.

OCR posts not in the data (3):

• kh-pressocm-33501 (https://pressocm.gov.kh/archives/33501): usable Khmer text below threshold (54 Khmer chars after noise filtering, 0 anchor words); not in the data
• kh-pressocm-70015 (https://pressocm.gov.kh/en/archives/70015): usable Khmer text below threshold (0 Khmer chars after noise filtering, 0 anchor words; pages dropped: 7 noise); not in the data; the post is in a Latin-script language, which the Khmer OCR model cannot read
• kh-pressocm-105737 (https://pressocm.gov.kh/archives/105737): duplicate within batch: kh-pressocm-104099 (not added again)

OCR pages left out of the texts: 227 of 1841 pages (121 noise, 57 no words, 40 low khmer, 4 contents, 2 wrong script, 2 low usable, 1 missing source; every page decision in manifests/ocr_pages_kh_ocr.json). The 40 pages with fewer than 50 Khmer characters of OCR output (blank pages, maps, seals, photos):

• 3394/f0_p0005 (29 Khmer chars) in kh-pressocm-3394
• 3394/f0_p0007 (32 Khmer chars) in kh-pressocm-3394
• 3409/f0_p0049 (43 Khmer chars) in kh-pressocm-3409
• 3409/f0_p0050 (33 Khmer chars) in kh-pressocm-3409
• 3409/f0_p0051 (29 Khmer chars) in kh-pressocm-3409
• 3409/f0_p0052 (38 Khmer chars) in kh-pressocm-3409
• 3409/f0_p0053 (26 Khmer chars) in kh-pressocm-3409
• 3409/f0_p0054 (18 Khmer chars) in kh-pressocm-3409
• 3409/f0_p0055 (16 Khmer chars) in kh-pressocm-3409
• 3409/f0_p0056 (3 Khmer chars) in kh-pressocm-3409
• 3409/f0_p0057 (1 Khmer chars) in kh-pressocm-3409
• 3409/f0_p0059 (20 Khmer chars) in kh-pressocm-3409
• 55399/f0_p0007 (0 Khmer chars) in kh-pressocm-55399
• 55399/f0_p0009 (0 Khmer chars) in kh-pressocm-55399
• 55399/f0_p0010 (28 Khmer chars) in kh-pressocm-55399
• 55399/f0_p0011 (0 Khmer chars) in kh-pressocm-55399
• 55399/f0_p0063 (0 Khmer chars) in kh-pressocm-55399
• 55399/f0_p0065 (0 Khmer chars) in kh-pressocm-55399
• 55399/f0_p0067 (0 Khmer chars) in kh-pressocm-55399
• 55399/f0_p0068 (5 Khmer chars) in kh-pressocm-55399
• 64708/f0_p0006 (33 Khmer chars) in kh-pressocm-64708
• 64708/f0_p0008 (21 Khmer chars) in kh-pressocm-64708
• 64708/f0_p0009 (26 Khmer chars) in kh-pressocm-64708
• 64708/f0_p0010 (31 Khmer chars) in kh-pressocm-64708
• 64708/f0_p0016 (30 Khmer chars) in kh-pressocm-64708
• 65972/f0_p0002 (42 Khmer chars) in kh-pressocm-65972
• 83215/img002 (46 Khmer chars) in kh-pressocm-83215
• 84215/img001 (12 Khmer chars) in kh-pressocm-84215
• 84219/img001 (5 Khmer chars) in kh-pressocm-84219
• 93742/img000 (11 Khmer chars) in kh-pressocm-93742
• 93769/f0_p0011 (49 Khmer chars) in kh-pressocm-93769
• 101719/img000 (21 Khmer chars) in kh-pressocm-101719
• 106739/f0_p0017 (13 Khmer chars) in kh-pressocm-106739
• 106739/f0_p0020 (6 Khmer chars) in kh-pressocm-106739
• 107966/img001 (49 Khmer chars) in kh-pressocm-107966
• 107981/img004 (28 Khmer chars) in kh-pressocm-107981
• 107981/img005 (13 Khmer chars) in kh-pressocm-107981
• 107981/img008 (35 Khmer chars) in kh-pressocm-107981
• 115706/img001 (42 Khmer chars) in kh-pressocm-115706
• 124104/img001 (45 Khmer chars) in kh-pressocm-124104

Missing source image (1): 86699/img002 of kh-pressocm-86699 (https://pressocm.gov.kh/wp-content/uploads/2023/11/photo_2023-11-02_18-07-05.jpg, not retrievable from the site; the law row is marked incomplete in its metadata).

Mostly non-Khmer pages dropped as noise (121) are concentrated in kh-pressocm-55399 (50), kh-pressocm-3409 (33), kh-pressocm-3415 (10), kh-pressocm-70015 (7), kh-pressocm-3336 (2), kh-pressocm-93769 (2) (coordinate tables, maps and the English pages of bilingual instruments).

Kept pages flagged noisy (109; noisy: true in manifests/ocr_pages_kh_ocr.json and in each row's metadata_json.pages): annex maps, GPS-coordinate and tariff-code tables and uniform/insignia plates. On these pages a second line filter removes the map and table OCR noise (garbled glyph runs) from the texts and article rows, 1330 lines on 103 pages, and keeps the real text (headings, numbered items, article text, tariff-code lists); some garbled characters remain inside kept lines and numbers in such tables are not reliable. Annex pages with noise lines removed (law id, page: lines):

• kh-pressocm-3216: p4 (7), p6 (3)
• kh-pressocm-3306: p4 (9)
• kh-pressocm-3309: p4 (13), p5 (10)
• kh-pressocm-3312: p6 (41), p8 (48)
• kh-pressocm-3315: p8 (31)
• kh-pressocm-3321: p4 (54), p5 (9)
• kh-pressocm-3333: p4 (8)
• kh-pressocm-3336: p5 (12)
• kh-pressocm-3339: p8 (26), p9 (25), p10 (26), p13 (24)
• kh-pressocm-3342: p11 (63), p12 (52), p14 (29)
• kh-pressocm-3345: p6 (12)
• kh-pressocm-3348: p5 (31)
• kh-pressocm-3351: p6 (8)
• kh-pressocm-3357: p5 (21), p6 (11), p7 (12)
• kh-pressocm-3360: p6 (38), p7 (35), p10 (25), p11 (26), p12 (33), p13 (35), p16 (30), p17 (26)
• kh-pressocm-3366: p4 (41)
• kh-pressocm-3370: p4 (10)
• kh-pressocm-3373: p4 (25)
• kh-pressocm-3379: p6 (15)
• kh-pressocm-3382: p4 (12)
• kh-pressocm-3394: p8 (10)
• kh-pressocm-3400: p4 (26), p5 (18), p6 (24), p7 (9)
• kh-pressocm-3403: p6 (16), p7 (19)
• kh-pressocm-3409: p5 (5), p7 (12), p12 (4), p14 (5), p23 (4), p25 (3), p27 (2), p28 (1), p31 (2), p32 (2), p41 (2), p42 (2)
• kh-pressocm-3415: p12 (2), p13 (4), p16 (1)
• kh-pressocm-3424: p5 (20), p8 (41)
• kh-pressocm-35380: p2 (1)
• kh-pressocm-55399: p1 (8)
• kh-pressocm-82774: p2 (3), p4 (3), p6 (3), p7 (4), p9 (1), p10 (4), p12 (4), p13 (1)
• kh-pressocm-86067: p21 (3), p27 (6)
• kh-pressocm-93769: p12 (5), p13 (13), p14 (8), p17 (4), p21 (10)
• kh-pressocm-106739: p18 (1), p19 (2), p25 (5), p26 (4)
• kh-pressocm-107981: p12 (9), p14 (16)
• kh-pressocm-116532: p21 (2), p27 (3), p29 (1), p41 (2), p45 (1), p51 (4), p53 (2), p58 (3), p60 (5), p61 (2), p65 (3), p68 (1), p69 (2), p71 (1)
• kh-pressocm-119518: p25 (5), p27 (2), p29 (2), p35 (1)

Residual glyph noise (r3): on all kept pages a third line filter (T) removes OCR glyph noise that the first two let through (stamp and table-border fragments, '7777' runs read off dotted blank-fill lines): 212 lines on 152 pages, each listed with its rule in manifests/changed_ids_kh_ocr.json (residual_noise_lines_removed_by_page). A line is never removed when it holds a real Khmer word, an article heading, a document number or a table amount. Of the 198 residual lines CoS listed in 66 r2 article rows, 83 are removed and 115 stay, each with its reason in manifests/annex_noise_review_kh_ocr.json: 80 table amount row, 16 short fragment, 8 real words with fill, 5 abbreviation, 4 numbered item, 2 document number.

Word-numbered article headings (r3): 37 article rows in 12 amending laws start at a heading whose number is written as a word (មាត្រាមួយ .-, មាត្រាពីរ ._, ...); they get ids <law_id>-art-w1, -art-w2, ... and metadata_json.heading_variant: word_numbered (with number_word as printed; heading_spelling when the word មាត្រា is misread; one number word misread by the OCR, មាត្រាមី for មាត្រាបី in kh-pressocm-83215, is recorded as number_word_ocr_misread). Each row starts at its heading and keeps the introductory line with it. Choice: the articles of the amended law that an operative article quotes (digit headings such as 'មាត្រា៣៣ ថ្មី.-') stay verbatim inside the w-row that quotes them and are not split (65 quoted headings, listed in article_heading_lines_not_split with the reason 'quoted amended article inside w'); digit and word numbering are checked separately.

Article headings: 124 article rows in 46 laws start at an OCR misspelling of មាត្រា (មាគ្រា 70, មាតា 11, ហត្រា 5, បាត្រា 5, មាតគ្រា 2, ខាត្រា 2, បាគ្រា 2, មហត្រា 1, ...); the text is verbatim and the spelling is recorded in the row's metadata_json.heading_variant (heading_spelling on w-rows). 63 lines that start with an article heading are cross-references or running-text mentions (e.g. 'មាត្រា៨២ និងមាត្រា៩០ នៃច្បាប់...') and are not split (listed in manifests/changed_ids_kh_ocr.json, article_heading_lines_not_split). 46 article rows repeat an article number within a law (second instrument, copy or annex in the same post, or a misread number) and get <id>--<sha8> as in fix-v5: sha8 is the first 8 hex digits of sha256 over the UTF-8 string law_id + "|" + title + "|" + text of the row as published (title empty when null); the first row of an id in law and text order keeps the plain id.

Article numbering (gate article_sequence): numbers rise within each law except 24 restarts at មាត្រា១ in 11 laws (second instrument, copy or annex in the same post: kh-pressocm-105207 (3), kh-pressocm-3333 (1), kh-pressocm-3345 (1), kh-pressocm-3388 (1), kh-pressocm-3403 (1), kh-pressocm-3406 (1), kh-pressocm-3436 (1), kh-pressocm-3473 (2), kh-pressocm-93149 (1), kh-pressocm-93178 (1), kh-pressocm-98906 (11)), 89 gaps in 59 laws (an article whose heading the OCR lost, a dropped page, or a misread number; listed in manifests/article_split_kh_ocr.json; r4: gaps are counted on the effective numbers, i.e. article_number_corrected where a row has one, while restarts and the exceptions below are on the numbers as read; gate gap_contents checks that no gap's missing article heading sits, in any spelling, in any row of the law (r5; numberless headings count with their inferred number); kh-pressocm-9628 10 -> 12: the heading 'មាត្រា ១១' is printed on page image 02a6c080...#p5 (below the chapter 2 title) but is not in the OCR output; the article text is at the end of row art-10; kh-pressocm-9628 16 -> 18: the heading 'មាត្រា ១៧' is printed at the top of page image 02a6c080...#p7 but is not in the OCR output; the article text is at the end of row art-16; kh-pressocm-35363 4 -> 6: the heading 'មាត្រា៥ .' is printed at the top of page image 3ba74ccd...#p3 but is not in the OCR output; the article text is at the end of row art-4 (the numberless 'មាត្រាង .' below the chapter 3 title is article 6); kh-pressocm-3464 1 -> 3: the heading 'មាត្រា ២.-' (execution clause) is printed on page image 20d79b53...#p10 but is not in the OCR output; the article text is at the end of row art-1 (the numberless 'មាត្រាព.-' is article 3)) and these reviewed exceptions:

• kh-pressocm-110395: 39 -> 32 (ocr_number_misread: previous heading is ៣១ printed, OCR read ៣9 (Arabic 9 for ១))
• kh-pressocm-110395: 49 -> 42 (ocr_number_misread: previous heading is ៤១, OCR read ៤9)
• kh-pressocm-3255: 4 -> 4 (ocr_number_misread: the last article (execution clause) is ៥ printed, OCR read ៤)
• kh-pressocm-3476: 25 -> 22 (ocr_number_misread: heading is ២៦ printed (sequence 25, 26, 27), OCR read ២២)
• kh-pressocm-3483: 9 -> 6 (ocr_number_misread: previous heading is ៥ printed (sequence 4, 5, 6), OCR read ៩)
• kh-pressocm-35363: 159 -> 16 (ocr_number_misread: previous heading is ១៥ printed, OCR read ១៥៩ (stray glyph))
• kh-pressocm-35799: 17 -> 14 (ocr_number_misread: previous heading is ១៣ printed (sequence 12, 13, 14), OCR read ១៧)
• kh-pressocm-67153: 0 -> 0 (ocr_number_misread: first article is ១ printed, OCR read 0 (Arabic zero))
• kh-pressocm-70620: 19 -> 12 (ocr_number_misread: previous heading is ១១ printed, OCR read ១9)
• kh-pressocm-9628: 145 -> 16 (ocr_number_misread: previous heading is ១៥ printed (page image 02a6c080...#p6: articles 12-16 on that page), OCR read ១៤៥; r4: the garbled heading lines of 12, 13, 14 and 16 on that page (មា‹គា ១២. ...) are rescued from line filter L and split)
• kh-pressocm-9628: 28 -> 25 (ocr_number_misread: previous heading is ២៤ printed, OCR read ២៨)
• kh-pressocm-9628: 32 -> 32 (ocr_number_misread: previous heading is ៣១ printed, OCR read ៣2)
• kh-pressocm-9628: 62 -> 62 (ocr_number_misread: previous heading is ៦១ printed, OCR read ៦2)
• kh-pressocm-9628: 64 -> 64 (source_repeat: the last page (printed page ៧៣២, promulgation and signature block) repeats article ៦៤ (abrogation clause, heading read ទាក្រា ៦៤) after article ៦៤ on page ៧៣១; both kept as published)
• kh-pressocm-93149: 20 -> 15 (source_page_order: the post publishes the page images out of order (articles 15-17 come after 20); text kept in published order)
• kh-pressocm-94234: 14 -> 4 (source_page_order: page images published out of order (articles 4-10 after 14))
• kh-pressocm-94234: 20 -> 15 (source_page_order: page images published out of order (15-16 after 20))

Misread article numbers: the 13 rows behind the ocr_number_misread exceptions keep their ids (as read) and carry metadata_json.article_number_corrected with article_number_correction_reason: kh-pressocm-110395-art-39 -> 31, kh-pressocm-110395-art-49 -> 41, kh-pressocm-3255-art-4--345cf836 -> 5, kh-pressocm-3476-art-22--77caaafc -> 26, kh-pressocm-3483-art-9 -> 5, kh-pressocm-35363-art-159 -> 15, kh-pressocm-35799-art-17 -> 13, kh-pressocm-67153-art-0 -> 1, kh-pressocm-70620-art-19 -> 11, kh-pressocm-9628-art-145 -> 15, kh-pressocm-9628-art-28 -> 24, kh-pressocm-9628-art-32 -> 31, kh-pressocm-9628-art-62 -> 61.

Garbled article headings (r4): 12 article rows start at a heading the OCR garbled beyond the misspellings above (letters dropped or replaced, quote or dot glyphs before the number, the digit ៩ read as ទ័): kh-pressocm-3483-art-25 ('មារតតា ២៥'), kh-pressocm-9628-art-9 ('មាត្រា ទ័'), kh-pressocm-9628-art-12 ('មា‹គា ១២'), kh-pressocm-9628-art-13 ('មា4គ1 ១៣'), kh-pressocm-9628-art-14 ('បា‹តា ១៤'), kh-pressocm-9628-art-16 ('មា,ុតា ១៦'), kh-pressocm-9628-art-21 ('មាត្រា “ ២១'), kh-pressocm-9628-art-28--2d178550 ('មារុតា ២៨'), kh-pressocm-9628-art-35 ('មារុតា ៣៥'), kh-pressocm-9628-art-44 ('មាត្រា. ៤៤'), kh-pressocm-9628-art-63 ('មារុតា ៦៣'), kh-pressocm-9628-art-64 ('មារុតា ៦៤'). They are split only as the next article or a small gap (at most 3 above the current number) and record metadata_json.heading_variant (the heading as printed), heading_number_read, heading_detection: relaxed and, for the look-alike digit, article_number_corrected with its reason. 4 such heading lines that the first line filter had dropped as glyph noise are kept in place in the text (listed in manifests/changed_ids_kh_ocr.json, heading_lines_rescued). Numbered table rows of the form ' | <4-digit code>' (the annex list of pollution sources in kh-pressocm-70868) are data and are never removed by the residual filter.

Headings without a number (r5): 4 article rows start at a heading line whose number the OCR lost or turned into a letter ('មាត្រា', 'មាត្រាង .', 'មាត្រាព.-', 'មាត្រា.'); the number is inferred and recorded with the text verbatim in metadata_json.heading_variant, heading_detection: numberless, article_number_inferred and inference_basis: kh-pressocm-123672-art-25 (sequence: between articles 24 and 26; page image 360c8b01769e0cd3.jpg (img006) shows មាត្រា២៥ under the seal); kh-pressocm-3464-art-3 (page image 20d79b5366d77a72.pdf p10 shows មាត្រា ៣.- (entry into force); article 2 (execution clause) has its heading on that page but it is not in the OCR output); kh-pressocm-3467-art-5 (internal reference: the distribution list has 'ដូចមាត្រា៥'; sequence: last article after 4; page image ba5826718e3287ce.pdf p2 shows មាត្រា៥.); kh-pressocm-35363-art-6 (page image 3ba74ccddb56c1d3.pdf p3 shows មាត្រា៦ ._ below the chapter 3 title (article 5 is the heading at the top of that page, which is not in the OCR output)). Gate numberless_heading: no such line is left inside a row (0 reviewed exceptions).

Checksums: manifests/kh_ocr_SHA256SUMS.json lists the sha256 of every other file of this batch's tree (MANIFEST.md included).

Leftover contents fragments: 32 non-blank lines (orphan contents entries, page numbers and running page headers) sit between two cut contents blocks no more than 15 lines apart and were not cut (they are outside the approved plan), in 2 laws:

• kh-e1-9e-98-e1-9f-81-e1-9e-9a-e1-9f-80-e1-9e-93-e1-9e-80-e1-9f-92-e1-9e-9a-e1-9e-98-e1-9e-9a-e1-9e-8a-e1-9f-92-e1-9e-8b-e1-9e-94-e1-9f-92-e1-9e-94-e1-9e-9c-e1-9f: 17 lines (parent lines 229, 230, 357, 358, 480, 481, 5827, 5861, 9559, 9601, 16057, 16093, 16120, 16162, 16163, 16164, 16166)
• kh-faolex-cam145554: 15 lines (parent lines 2622, 2624, 2726, 2820, 2852, 2853, 2854, 2937, 2939, 2965, 2966, 2967, 2968, 2969, 2970)

Fix 2026-10-17 (Limon laws: structural headings at row ends)

Batch fix-20261017-kh-limon-tails on parent 158516831d82c8e8a0dd1f3ffde6e94a035b0502. The 10 other laws in legacy Limon encoding (Civil Code kh-law_030_1207_kh, Code of Civil Procedure kh-law_021_0706_kh, the six Civil Code commentary volumes and two shorter laws); laws.text unchanged; laws 274, articles 12,027 (438 changed, none added or removed; 11 law rows changed: metadata only). Every other row is byte-identical (manifests/rowhash_proof_kh_limon_tails.json); ids in manifests/changed_ids_kh_limon_tails.json; code in scrapers/kh_limon_tails_20261017/.

• 899 Khmer structural heading lines (541 headings: book / chapter / part / section, with their wrapped title lines) sat at the end of 438 rows. They are removed verbatim from the row ends (each row now ends at its last non-blank line before them) and recorded as structural metadata of the law row (metadata.kh_limon_tails_20261017.structural_headings, 611 headings with their position: row_end, inside_row or between_rows). The earlier list of 136 rows (structural_headings_kh_headings.json) is included; the rest were missed by its classifier (heading set right under the last body line, title on the next line).
• Sole divisions ('CMBUkeTal' = the only chapter of its book) carry no number in the text: number stays null and each such entry gets sole_note, implied_ordinal 1 and its parent heading; this includes the 4 sole chapters of the bilingual commentary (offset 1384100: the sole chapter of Book 8).
• scrapers/kh_limon_tails_20261017/requirements_kh_limon_tails_20261017.txt pins the direct packages and their full installed dependency closure (urllib3, certifi, numpy, …).

Gates: verbatim_slices, conservation, unrowed_delta, structural_tail, structural_offsets, structural_complete, sole_documented, row_prefix, law_row_delta, limon_coverage, merged_heading, heading_only, article_count, pins_transitive, fleet / pressocm / r5 OCR gates.

Checksums: manifests/kh_limon_tails_SHA256SUMS.json lists the sha256 of every other file of this tree (MANIFEST.md included).

Fix 2026-10-17 (bilingual commentary: structural headings)

Batch fix-20261017-kh-headings on parent f7fcb714eecdd46ba0703162af6238bcb7313bc2. Law -e1-9e-9a-e1-9e-98-e1-9e (bilingual Civil Procedure commentary); laws.text unchanged; laws 274, articles 12,027 (12,031 -> 12,027: 70 changed, 4 removed, none added; 1 law row changed: article_count and metadata). Every other row is byte-identical (manifests/rowhash_proof_kh_headings.json); ids in manifests/changed_ids_kh_headings.json; code in scrapers/kh_headings_20261017/.

• 90 Khmer structural heading lines (89 headings: book / chapter / part / section of the commentary that follows) sat at the end of 70 rows (mostly English rows: the quoted code article that ends a notes block). They are removed verbatim from the row ends (each row now ends at its last non-blank line before them) and recorded as structural metadata of the law row (metadata.kh_headings_20261017.structural_headings, 96 headings: these 89, the 5 of the earlier batch and 2 inside row -article-63--cd2d039f, Book 8 and its sole chapter, which the Khmer introduction of that book follows; that row is unchanged).
• 4 of these rows held only a note line ('Article N (…) > see the note for Article M') before the heading; as in the earlier batch, the note line now ends, verbatim, the row before it in laws.text (only a line break between) and the note row is removed (changed_ids, now_in).
• Offsets: every entry's offset is the first character of the heading text in laws.text; indent is the verbatim leading indentation (laws.text[line_start:offset]) and text the verbatim slice (a wrapped heading keeps its second line's indentation). The 5 entries of metadata.kh_bilingual_notes_20261017.structural_headings (offset at the line start, indentation dropped) are corrected in place.

Gates: verbatim_slices, conservation, unrowed_delta, structural_tail (no row of the law ends with a structural heading), structural_offsets, structural_complete, row_prefix, law_row_delta, limon_coverage, merged_heading, heading_only, en_mid_sentence, article_count, fleet / pressocm / r5 OCR gates.

Checksums: manifests/kh_headings_SHA256SUMS.json lists the sha256 of every other file of this tree (MANIFEST.md included).

Fix 2026-10-17 (bilingual commentary: unrowed lines)

Batch fix-20261017-kh-bilingual-notes on parent 7df0e7cab48e85849316afca9c69046ca6b95ada. The bilingual Civil Procedure commentary (-e1-9e-9a-e1-9e-98-e1-9e) had 74 non-blank lines in no row after the merged-split batch; each is decided in manifests/unrowed_lines_review_kh_bilingual.json; laws.text is unchanged. Articles 12,045 -> 12,031 (+2 added, 14 changed, 16 removed); laws unchanged (274; 1 law row changed: article_count and a metadata record). Every other row is byte-identical and in its place (manifests/rowhash_proof_kh_bilingual.json); ids in manifests/changed_ids_kh_bilingual.json; code in scrapers/kh_bilingual_notes/.

• Quoted Japanese articles: Civil Execution Act article 121 (notes of Khmer article 465) and Rules of Civil Procedure article 190 (notes of Khmer article 315) get one row each (-article-121--…, -article-190--…), replacing the false-split rows that held part of them (-article-54--0b799872, -article-315--18696329); article 121 in the notes of Khmer article 458 and Civil Provisional Relief Act article 49 get their unrowed line and the false-split row after it joined verbatim (-article-54--26075f3a, -article-136--0e35957a removed).
• 11 English rows cut at a line starting 'Article N' that continues the line before (10 inside a sentence of a quoted article, e.g. -article-275 inside German article 296; 1 note line 'Article 73 > see the note for the preceding article') are joined verbatim to the row before them and removed; two of them also held a Khmer part / section heading after that text, which is now structural metadata. Each removed id is listed with the row that holds its text (changed_ids, now_in). Gate en_mid_sentence: no English row of this law starts mid-sentence unless reviewed (unrowed_lines_review_kh_bilingual.json, en_mid_sentence_reviewed).
• -article-496--25d9097e held only the wrapped tail of a note line and the Khmer Part 4 heading: the note lines and the tail now end the row of Khmer article 508, the heading is structural metadata.
• 2 quotes repeat, verbatim, a quote that already has a row (Civil Execution Act articles 57 and 135): class quoted_article_repeat, no second row.
• 51 lines are English note lines ('Article N (…) > see the note for Article M') and 1 line labels the quoted German rules: classed, no row.
• Khmer structural headings (5: Book 1 and Chapter 1 before article 1, Part 3, Section 4 and Part 4) are recorded verbatim as structural metadata of the law row (metadata.kh_bilingual_notes_20261017.structural_headings).

Gates: conservation (no row text lost except the structural headings moved to metadata; only the reviewed lines added), unrowed_classified (the lines still in no row are exactly the classed ones), limon_coverage, merged_heading, heading_only, en_heading, en_mid_sentence, id_scheme, repeat_verbatim, article_count, fleet / pressocm / r5 OCR gates.

Checksums: manifests/kh_bilingual_SHA256SUMS.json lists the sha256 of every other file of this tree (MANIFEST.md included).

Fix 2026-10-17 (merged headings and Limon split)

Batch fix-20261017-kh-merged-split on parent e31da4e82b386cdeedef654a972ae3e51e0617d3. Article rows that held the heading of a later article are cut at that heading; laws.text is unchanged. Articles 11,184 -> 12,045 (+889 added, 908 changed, 28 removed); laws unchanged (274; 15 law rows changed: article_count and a metadata record). Every other row is byte-identical and in its place (per-row hashes in manifests/rowhash_proof_kh_merged.json); ids in manifests/changed_ids_kh_merged.json; per-law decisions in manifests/article_split_kh_merged.json; code in scrapers/kh_merged_split/.

• Legacy Limon-encoded Khmer (9 laws, +811 rows): every heading 'maRta N>-' of the published live-51 splitter, and amended-article headings 'maRta N fμI>-' that fill a gap of that sequence, now starts a row. The bilingual Civil Procedure commentary (-e1-9e-9a-e1-9e-98-e1-9e, 789 English rows listed by article number) gets one row per Khmer article (id <law>-article-N--<sha8>), placed after the English row it was cut from.
• Japanese translations (3 laws, +47 rows): every whole-line heading '第N条（title）' in sequence starts a row.
• r5 OCR rows (3 laws, +3 rows): kh-pressocm-35394-art-7, kh-pressocm-86067-art-12, kh-pressocm-91696-art-6 were behind 1-3 junk glyphs inside the previous row (the glyphs stay there).
• 75 rows of the first release were false splits at a reference to their own article (e.g. '第７７条が適用されないと解される…', 'maRta 989 ¬…¦ xagelIenH…'); each keeps its id and now starts at the heading of its article, and the text before the heading stays with the article it belongs to.
• Removed: 28 rows of the bilingual commentary whose only text before a Khmer article was their title line: the tail of an English sentence ('…the provisions in' / 'Article 38.': the old splitter cut at a line starting with a reference) or an English note line of the notes block before it ('Civil Execution Act (Japan)' / 'Article 59 (…) > see the note for Article 431'). Cut alone they would be heading-only rows, so the line is joined back to the row before it (the sentence or notes block it belongs to) and the Khmer article gets its own row; the old ids are listed in manifests/changed_ids_kh_merged.json (articles_removed, with the row each line was joined to).
• Not cut: -e1-9e-98-e1-9e-9ee5c0bd (Japanese, 15 headings): a two-column layout interleaved line by line in the text layer; a verbatim cut would move text of another article (left for a column-aware re-extraction).

Gates: conservation (every non-blank character of the old rows is in a new row; the only text added to rows is unrowed heading text and 6 unrowed English note-line gaps joined with removed rows), coverage (every heading of the family starts a row), sequence (numbers rise by 1 in text order; gaps reviewed), merged_heading (live detector, form feeds read as blanks: no target or changed row holds the next heading), id_scheme, heading_only, inner_heading, r5 OCR gates, check-E noise lines unchanged. The 106 headings listed in manifests/merged_headings_scan_kh_live51.json and their outcome are in manifests/merged_headings_review_kh_merged.json.

Checksums: manifests/kh_merged_SHA256SUMS.json lists the sha256 of every other file of this tree (MANIFEST.md included).

Fix 2026-10-17 (live-51 article split)

Batch fix-20261017-kh-live51-split on parent 59d4d01131d5a40e4c16119d4051c3dac76573a1. Article rows for 15 laws of the first release that had none, a re-split of 2 laws whose rows held several articles, one row split, and noise review of the OCR rows. Articles 10,356 -> 11,184 (+829 added, 10 changed, 1 removed); laws unchanged in number (274); 23 law rows changed (article_count, metadata; text only where noise lines were cleared). Every other row is unchanged (per-row hashes in manifests/rowhash_proof_kh_live51.json); ids in manifests/changed_ids_kh_live51.json; per-law decisions in manifests/article_split_kh_live51.json; code in scrapers/kh_live51_split/.

Per law (rows before -> after): kh-faolex-cam1068 0 -> 39 (en_article); kh-faolex-cam1071 0 -> 4 (en_article); kh-faolex-cam117198 0 -> 167 (en_article); kh-faolex-cam124593 0 -> 79 (en_article); kh-faolex-cam134075 0 -> 45 (en_article); kh-faolex-cam134079 0 -> 8 (en_article); kh-faolex-cam134594 0 -> 13 (en_article); kh-faolex-cam145554 0 -> 3 (en_article); kh-faolex-cam151539 0 -> 119 (en_article); kh-faolex-cam153589 0 -> 48 (en_article); kh-faolex-cam159612 0 -> 84 (en_article); kh-faolex-cam161035 0 -> 8 (km_prakar); kh-prakas_086_1208_moj_kh 0 -> 7 (km_prakar); kh-prakas_098_1209_moj_kh 0 -> 22 (km_prakar); kh-e1-9e-94-e1-9f-92-e1-9e-9a-e1-9e-80-e1-9e-b6-e1-9e-9f-e1-9f-a9-e1-9f-a8-20-e1-9e-9f-e1-9f-92-e1-9e-8f-e1-9e-b8-e1-9e-96-e1-9e-b8-e1-9e-93-e1-9e-b8-e1-9e-8f-e1 0 -> 22 (limon_prakar); kh-law_007_0511_kh 3 -> 82 (km_matra); kh-e1-9e-9f-e1-9f-81-e1-9e-85-e1-9e-80-e1-9f-92-e1-9e-8f-e1-9e-b8-e1-9e-80-e1-9f-86-e1-9e-8e-e1-9e-8f-e1-9f-8b-e1-9e-85-e1-9f-86-e1-9e-96-e1-9f-84-e1-9f-87-e1-9e 2 -> 82 (limon_matra); kh-law_030_1207_kh 1189 -> 1190 (row split).

Headings: English 'Article N' / 'ARTICLE N' / 'Art. N' (FAOLEX texts; a line whose tail is a cross-reference, e.g. 'Article 5 of this law', or whose previous line ends in a reference word, is not split; amended articles 'Article 11 new' are split), Khmer មាត្រា N (the r5 OCR splitter, headings behind table-border glyphs included), ប្រការ N (prakas, with OCR misspellings accepted only as the next number or a gap of up to 3) and legacy Limon 'maRta N>-' / 'Rbkar N>-'. Contents-leader lines are never headings; a heading more than 10 above the current number is not split (quoted or misread). Rows are verbatim slices of laws.text from a heading to the next. OCR-garbled headings whose number could not be read are split only where the number is confirmed (FAOLEX page images re-read, or the article titles of the English and Japanese versions of the same law) and record metadata_json.metadata.split.heading_review; misread numbers keep the id as read and record article_number_corrected with the reason.

Ids: <law_id>-article-<N> as the live rows (a law id too long for 180 characters is cut and gets '-' + sha1(law_id)[:8]); a second body with the same id gets <id>--<sha8>, sha8 = sha256(law_id|title|text)[:8]. Ids of existing rows are kept where the article number is unchanged. kh-faolex-cam117198 (Constitution) has a reviewed restart 160 -> 1 for the Additional Constitutional Law printed after it (articles 1-7); all 7 rows of that second run carry the --<sha8> suffix.

A heading that starts inside a body line is split there: kh-faolex-cam124593 'ARTICLE 17:' ends the last body line of article 16 in the source text layer (laws.text line 254); the row -article-17 starts at the heading (metadata_json.metadata.split.heading_review). The OCR tail of kh-faolex-cam1068-article-44 after the signature block (34 lines of seal / stamp glyph fragments) is kept out of the row by the law's end line; laws.text is unchanged and the lines are listed in manifests/annex_noise_review_kh_live51.json.

Merged-heading scan (all 11,184 article rows): no row added or changed by this batch, and no row of a law split here, holds another article's heading. 106 headings inside 103 older rows that this batch does not touch (Japanese and legacy Limon-encoded Khmer translations of the first release, and 3 OCR rows of the 2026-10-17 OCR batch) are listed with their line in manifests/merged_headings_scan_kh_live51.json for a later batch.

Gaps (reviewed): kh-law_007_0511_kh 0->2: the article 1 heading is not in the OCR text: laws.text lines 48-54 hold the chapter 1 heading and the body of article 1 ('ច្បាប់នេះ បញ្ញត្តិអំពីកាលបរិច្ឆេទនៃការអនុវត្តក្រមរដ្ឋប្បវេណី...') with no 'មាត្រា ១' line; that text stays in laws.text only (before the first row); kh-law_007_0511_kh 20->22: the article 21 heading ('Execution of Registration') was lost at the page break (laws.text line 311 reads 'To'); the body of article 21 (lines 313-317) stays at the end of the row of article 20; kh-faolex-cam1068 1->7: articles 2-6 are not in the source: the scan goes from article 1 (page 1) to Chapter 2, article 7 (page 2) (page image of the FAOLEX scan (re-read at 200 dpi)); kh-e1-9e-9f-e1-9f-81-e1-9e-85-e1-9e-80-e1-9f-92-e1-9e-8f-e1-9e-b8-e1-9e-80-e1-9f-86-e1-9e-8e-e1-9e-8f-e1-9f-8b-e1-9e-85-e1-9f-86-e1-9e-96-e1-9f-84-e1-9f-87-e1-9e 19->22: articles 20-21 are not in the text (no heading and no commentary between article 19 and article 22).

OCR noise review (CoS r5 check E): 68 lines in 34 OCR article rows have < 50 % Khmer characters or a '777' / 6+ digit run; 6 pure glyph-noise lines are removed from the row and from laws.text, the other 62 are real text and stay, each with its class in manifests/annex_noise_review_kh_live51.json (21 coordinate reference, 11 identifier number, 9 real words with fill, 7 numbers in text, 6 table amount row, 5 real words, 2 tariff codes, 1 document number).

Checksums: manifests/kh_live51_SHA256SUMS.json lists the sha256 of every other file of this tree (MANIFEST.md included).

OCR batch 2026-10-17 (pressocm.gov.kh scans)

Batch ocr-20261017-kh-pressocm on parent 1a0c098cb504cbf103c01ef79d959762cb3a2479. Changelog: added 223 instruments (laws, royal krams, royal decrees, sub-decrees) and 2,352 article rows from the scanned PDFs and page images published by the Office of the Council of Ministers' Press and Quick Reaction Unit (pressocm.gov.kh) that the deepen batch held because they have no usable Khmer text layer. Ids follow the pressocm scheme (kh-pressocm-<post>, articles <law_id>-art-<N> split on មាត្រា N and its OCR misspellings, <law_id>-art-w<N> on word-numbered headings of amending laws, cross-references not split; scrapers/kh_ocr/split_kh_ocr.py). Rows of the previous release are unchanged (per-row hashes in manifests/rowhash_proof_kh_ocr.json); ids in manifests/changed_ids_kh_ocr.json.

Text is OCR: PDF pages rasterised with pdftoppm at 300 dpi (grey), page images OCR'd as published, tesseract 5.5.0 (leptonica 1.84.1) with the khm model (--psm 6, single thread). Each law row's metadata_json.text_extraction records the method, engine and version, model sha256 and thresholds, and metadata_json.pages the per-page decision, the source PDF/image sha256 and the OCR text sha256. Per-page confidence is not available (plain-text OCR output; confidence: null); a 10-page spot check with word confidences is in manifests/spotcheck_kh_ocr.json.

Noise threshold: decorative symbol runs are removed; a line is kept when it has at least 2 Khmer characters, Khmer is at least 30 % of its non-space characters and at least half its Khmer characters are in runs of 3 or more; a page is dropped when it has fewer than 50 Khmer characters, Khmer is under 50 % of its non-space characters, at least 35 % of its tokens mix Khmer with Latin letters or digits (a Latin page read by the Khmer model), it is a contents page, or fewer than 50 Khmer characters survive the line filter; a post needs at least 300 Khmer characters and one anchor word; on pages flagged noisy, lines of map/table noise (low Khmer share, scattered glyphs, or improbable Khmer character trigrams under a model counted from this batch's clean pages) are also removed; on all kept pages, lines of residual glyph noise with no real Khmer word are removed. 1,614 of 1,841 pages were kept.

Spot check (10 random pages compared with the page images): body text on typed pages is plausible at about 99 % of characters (8 of 10 pages; tesseract median word confidence 85-92); headings in the bold display font, uniform/insignia plates and table pages are far worse (about 75-90 %); recurring errors are pre-base vowels detached in bold headings, subscript confusions (្វ/្រ, ្ឈ/្ល, ្ហ/្ជ), ច/ប and ៥/៦/៩ confusions, and Latin text read as Khmer-looking glyphs; Khmer subscripts on body text are mostly right.

Fix 2026-10-17 (contents pages in law full texts)

Batch fix-20261017-kh-tocfix on parent 8dc752b68af757edc84b3dfc65143f7eaf01e42c. Changelog: removed table-of-contents / list-of-tables pages from laws.text of 24 laws (50 blocks cut, 2 kept as body text; 469,405 characters). Each block was checked line by line (contents entries with dot leaders and page numbers, contents headings, page numbers, wrapped entries) and hashed before it was cut; every block starts at its contents heading (or, without one, at its first entry), so titles, law names, promulgation/edition lines and series/author lines before the contents are kept; body text and the documents' real tables are kept. articles.parquet and every other column and row are unchanged (per-row hashes in manifests/rowhash_proof_tocfix.json). Ids, old/new text sha256 and every block decision: manifests/changed_ids_tocfix.json; plan: manifests/fixplan_toc_fulltext.json.

Deepen batch (pressocm.gov.kh)

Added 1 instruments (laws, royal krams, royal decrees, sub-decrees) published as born-digital PDFs by the Office of the Council of Ministers' Press and Quick Reaction Unit (pressocm.gov.kh, an official OCM unit), from its decree and sub-decree categories; news, circulars, decisions and prakas were filtered out. Text comes from the PDFs' own Khmer text layer (no OCR); articles are split on មាត្រា N and get the 2026-10-16 rules (HTML entities, heading-only and table-of-contents rows dropped). 226 further instrument posts are scans, held for a later OCR batch. 0 instruments already in the dataset were skipped. Ids: manifests/changed_ids_pressocm.json. Rows of the previous release are unchanged.

pressocm.gov.kh states no reuse licence; the instruments are official legislative texts reproduced for research, credited to the Office of the Council of Ministers (Press and Quick Reaction Unit). robots.txt (only /wp-admin/ disallowed) was checked through the fleet robots guard on every request and redirect hop; at least 2 s between requests; at most 1,000 requests; no login, no captcha/WAF workaround.

Provenance

Primary legal texts are public legal information. Dataset compilation is a research artifact; always verify against the official source before relying on a text. Not legal advice.