ocr-projects-archive / ocr-tech /docs /DATA_LICENSING.md
prosperious's picture
Add 1000 files (OCR_INDRA/.gitignore ...)
f9bc0c1 verified
|
Raw History Blame Contribute Delete
31.5 kB

Data licensing ledger

Started 2026-07-14 from the adversarial license audit (see RESEARCH.md "Verification wave"). Every data source used for training or claims MUST have a row here. Legend: OK / RISK (unresolved) / NO.

Source License (verified) Train Publish numbers Redistribute Commercial
Santali Wikipedia dump CC BY-SA 4.0/GFDL OK OK OK (attribution+SA) OK
Aale Chhatka 2023 issue Creative Commons Public Domain Mark 1.0 applied on publisher/uploader item (not CC0; mark is not a license) OK by project policy OK by project policy OK by project policy OK by project policy
Aale Chhatka Sept 2023 UNSPECIFIED (same publisher) RISK RISK NO RISK
NVIDIA OCR-Synthetic-Multilingual-v1 (en/validation/validation_000.h5; 6,802 horizontal train + 419 test line crops on k2 ~/ocr-deva-ft/data/nvidia-en-lines, provenance.jsonl per crop) CC BY 4.0 (dataset card, verified 2026-09-24) OK (attribution) OK OK with attribution OK
Devanagari OFL font set for synthetic rendering (48 families: Google Fonts ofl/ + embedded-OFL Aarav, Annapurna SIL Nepal; list in the 2026-09-24 fine-tune font_selection.json) SIL OFL 1.1 (verified per family) OK (rendered images) OK fonts not redistributed OK
Gorkhapatra-derived real chunks chunks-v2 and chunks-v4 (24 editions; chunks-v4 trained the deployed devanagari-ft-v3; tools/build_gorkhapatra_devanagari_chunks.py) as the Gorkhapatra row: owner determination 2026-07-22 covers internal training and derived metrics OK-internal OK (metrics) NO (crops) owner decision required before a model trained on it is offered commercially
himalaya-ai/nepali-deva-ocr-eval (500 rows from gauravgiri/nepali-ocr-dataset; photographed handwriting) none recorded NO (eval-only probe) RISK (internal diagnostic only) NO NO
IIIT-HW-Dev handwritten Hindi words (HF c3rl/IIIT-INDIC-HW-WORDS-Hindi; Kaggle himanshukumarrajak/indic-data and re-uploads) research and educational use only (CVIT IIIT Hyderabad) OK-research only OK-research only NO NO (a model trained on it stays a research model until CVIT permits otherwise)
DHCD, Devanagari Handwritten Character Dataset (UCI id 389) CC BY 4.0 OK (attribution) OK OK with attribution OK
sweekardahal handwritten Nepali word crops (Kaggle sweekardahal/handwritten-labeled-dataset-for-text-recognition) CC0 (Kaggle dataset page, checked 2026-09-24) OK OK OK OK
sweekardahal Nepali handwritten page photos, detection boxes (Kaggle sweekardahal/nepali-handwritten-images-for-text-detection) CC0 (Kaggle dataset page, checked 2026-09-24) OK OK OK OK
IndicScan handwritten Devanagari words (Kaggle rushdaismailaslami/handwritten-devanagari-character-and-word-dataset) CC BY 4.0 (Kaggle dataset page, checked 2026-09-24) OK (attribution; used as test only) OK OK with attribution OK
ocr.gov.np public notices (3 PDFs, tools/deva_rec_torch/page_probe/SOURCES.txt) + our hand transcription of 43 lines Government of Nepal public notices; no licence stated NO (eval-only) OK (derived metrics only) NO (PDFs not redistributed) NO
IIIT Mozhi printed lines (HF mirrors darknight054/indic-mozhi-ocr, AtharvImmverse/indic-mozhi-ocr) none stated on CVIT project page (checked 2026-09-24) NO until CVIT confirms NO NO NO
chiki-chemed primer CC BY-NC-ND 4.0 NO (catalogue only — verified unused) — NO NO
Videha Tirhuta text Uploader-applied item-level CC BY-SA 4.0 (IA metadata verified); item descriptions also reserve individual-author rights, retained as a caveat OK by project policy OK by project policy OK by project policy (attribution+SA; do not imply broader author-rights waiver) OK by project policy
Videha Tirhuta silver line pack (outputs/videha-tirhuta-silver-2026-07-17/; 7 exact-hash-bound issues) Uploader-applied item-level CC BY-SA 4.0 on the two Internet Archive parent items; per-issue source URLs and hashes are recorded in the pack provenance; retain the existing individual-author-rights caveat OK by project policy OK by project policy OK by project policy (attribution+SA; redistribute with pack provenance and do not imply a broader author-rights waiver) OK by project policy
Videha embedded "janaki" font MPP Janaki 1.000 metadata says “All rights reserved”; no license grant decode-only — NO NO
Zenodo 6967421 Newa MSS CC-BY-4.0 OK OK (attribution recorded in docs/ATTRIBUTIONS.md) OK+attr OK+attr
nepal-lipi.com text (285 lines) all-rights-reserved, research-only NO (v2 payload VIOLATES — rebuild required) NO NO NO
NepalLipiUnicode.ttf bare copyright, no grant RISK (in v2 payload renders — rebuild) RISK NO NO
Nithya Ranjana 1.000 DU (8a2ebe62…b92b) and NU (22a12a91…1528) Exact official Ek Type release; embedded and standalone SIL OFL 1.1; upstream identifies DU as Devanagari-based and NU as Newa-based candidate only — no synthetic payload admitted NO — no OCR metric or model OK subject to OFL OK subject to OFL
Commons “Bhujimol script characters (Unicode)” SVG (bf79c86f…d8655d) Uploader Z3lvs marks own work CC0 1.0; exact Commons API metadata and source bytes frozen NO — basic-glyph reviewer reference only; not a font or mapped training asset NO OK under CC0 OK under CC0
Wellcome MS Indic epsilon 1 Bhujimol manuscript Authoritative IIIF manifest applies Public Domain Mark 1.0 and explicitly permits copying, modification, distribution, performance, and commercial use without permission; attribution: Wellcome Collection NO — complete remote-review candidate; 217 canvases, zero qualified page decisions or labels NO NO until a frozen reviewed pack exists NO until admitted
GRETIL Aṣṭasāhasrikā e-text (Vaidya 1960 edition) Page states “FOR REFERENCE PURPOSES ONLY” and defers copyright/usage to the source file; no dataset-training or redistribution grant was verified NO — different-edition work-level reviewer reference only; no folio/line crosswalk NO NO NO
Internet Archive Asa Archives item 1147 (Bhāsvati) Uploader-applied item-level CC0 1.0 metadata; uploader authority over the Asa Archives scan is not yet verified NO — catalogue only pending origin, exact-file SHA-256, qualified script review, and labels NO NO NO
Wikimedia Haramekhala 1374 Bhujimol image Public Domain Mark 1.0 on a single Asa Archives-derived image; mark is not a license and origin chain is unresolved NO — catalogue only NO NO NO
British Library EAP790/21/121 Bhujimol manuscript Catalogue says research-only; EAP image route carries CC BY-NC 4.0 branding NO — catalogue only NO NO NO
Commons Cambridge MS Add.2137 Bhujimol photograph Exact Commons JPEG is self-licensed by its photographer under CC BY-SA 4.0; attribution and share-alike apply. Cambridge's separate complete IIIF scan is CC BY-NC 4.0 and is provenance/script evidence only. NO — one selected composite, zero qualified pixel decisions or labels NO NO until an exact hash-bound reviewed pack exists NO until admitted
Cambridge MS Add.2137 catalog TEI and listed Nāradasmṛti editions Catalog metadata/reference only; the TEI states that the Newari legal text and Newari translation are unpublished, and no complete line transcription is supplied NO — partial work-level Sanskrit reference only; not companion ground truth NO NO NO
Smithsonian S1991.161 image FS-6314_06 Exact image is marked CC0 on the object page, alongside generic usage-conditions text NO — catalogue only; one IIIF canvas and no labels NO NO NO
British Library EAP1296/2/217 Ranjana manuscript EAP file page carries CC BY-NC 4.0 branding NO — catalogue only NO NO NO
Wellcome MS Indic delta 33 selected Ranjana images Four exact digital-image records carry CC BY 4.0; the physical manuscript remains closed-store only and has no complete IIIF presentation manifest NO — catalogue only; selected overlapping/partial views and no labels NO NO NO
BDRC MW1NPL132 Ranjana manuscript BDRC's official access policy classifies scanned original manuscripts as public domain and permits sharing/adaptation for any purpose; exact scan W1NPL132 is AccessOpen/StatusReleased. The IIIF manifest has no license field, and the work-record CC0 marker is metadata-only. NO — complete remote-review candidate; 21 canvases, zero qualified page decisions or labels NO NO until a separately frozen reviewed pack exists NO until admitted
BDRC MW1NPL23 mixed-script Ranjana manuscript BDRC's official access policy classifies scanned original manuscripts as public domain and permits sharing/adaptation for any purpose; exact scan W1NPL23 is AccessOpen/StatusReleased. The physical record assigns both ScriptNewa and ScriptRanj; the manifest has no license field. NO — remote-review candidate; 25 canvases, zero qualified page decisions or labels NO NO until a separately frozen reviewed pack exists NO until admitted
BDRC MW1NPL99 NepaleseHooked/Bhujimol manuscript BDRC's official access policy classifies scanned original manuscripts as public domain and permits sharing/adaptation for any purpose; exact scan W1NPL99 is AccessOpen/StatusReleased. The physical record assigns ScriptNepaleseHooked; the project maps that institutional term to its Bhujimol review profile, subject to qualified page review. The IIIF manifest has no license field, and the work-record CC0 marker is metadata-only. NO — complete remote-review candidate; 73 canvases, zero qualified page decisions or labels NO NO until a separately frozen reviewed pack exists NO until admitted
Nepalbhasa Manuscripts in UK archive Publicly viewable with transcription/translation, but no explicit reuse, training, redistribution, or commercial license found NO — permission required NO NO NO
Sikkim Herald editions hosted on sikkim.gov.in Custom State Portal permission: material may be reproduced free of charge in any format/media without specific permission, if accurate, non-derogatory/non-misleading, and prominently source-acknowledged; explicitly identified third-party material is excluded. Not public domain and not a standard open license. The newer ipr.sikkim.gov.in listing is discovery-only unless its file is mirrored on the State Portal or separately permitted. OK by project policy for State Portal-hosted government-published Herald pages after third-party screening OK with source/permission attribution and honest claim boundaries OK with the stated conditions; exclude third-party material OK by the permission text (no NC restriction), subject to the same conditions
Unicode L2/22-096 embedded khema 2019 reference font No standalone font redistribution grant captured; used only inside the proposal for semantic and glyph-shape attestation decode/reference only; not a training or rendering asset — NO NO
Gorkhapatra Nepal Copyright Act 2059 §3 "government work" exclusion — UNVERIFIED for Gorkhapatra Sansthan (statutory corporation); owner determination 2026-07-22 (see below) OK-internal OK (derived metrics/counts) NO (pixels/crops) OK (metrics only)
OLCKOptimum fonts (extracted) OFL 1.1 (embedded license text) OK OK OK OK
SakalBharati, Noto fonts OFL OK OK OK OK
MithilaUni.ttf no license text, no provenance quarantine — do not use until documented — NO NO

Owner determination: Gorkhapatra derived-metric publication (2026-07-22)

Recorded decision by the project owner (ashish@ampixa.com): Gorkhapatra is Nepal's state-owned national daily (Gorkhapatra Sansthan), and on that basis the owner accepts publication of DERIVED METRICS AND COUNTS (CER scores, line counts, split sizes) computed from its epaper material — including in the LimbuOCR paper — as an acceptable use. Scope limits:

  • This is an owner risk-acceptance decision, not a verified legal opinion; the §3 "government work" statutory question above remains unverified.
  • It does NOT cover redistribution of page pixels, rendered images, or crop packs; those remain withheld pending a formal license statement.
  • Training-use status for internal models is unchanged (internal OK).

Claim corrections (2026-07-14)

  • Ol Chiki public headline (2nd correction, same day): the label audit found 67/895 crops (all in the public-marked issue) are BLANK — a pack-builder rendering bug that added guaranteed-error lines. Blank-free numbers: public-marked-issue-only cp CER 0.1018 / exact 0.191 (n=366) is a historical diagnostic only; combined blank-free 0.0826 (n=828) stays internal-diagnostic pending the September issue's license. The earlier 0.1929/0.1275 figures are superseded (blank-inflated). These figures are not claim-grade: crop, label, punctuation, leakage, and asset-independence review remains pending. The public-marked-source-only v3 candidate now has 358 rows after 75 blank renders were rejected; it remains unreviewed, unfrozen, and claim-ineligible.
  • Newa v2 curriculum payload (2026-07-14) is license-contaminated (31.5% of synth text + ~half of renders from research-only sources) and is also blocked on recipe grounds (see SOTA review) — DO NOT TRAIN from it; rebuild on Zenodo train-split text + OFL fonts only.
  • The recovered Newar-Devanagari normalized pack is not admitted (2026-07-15 r12). An independent omac audit exactly reproduced the two narrow whitespace rewrites over 6,763 rows and decoded all 6,763 referenced PNGs with zero missing, blank, or cross-split image hashes. That mechanical success does not repair the source: the recovered label rows omit PDF URL, PDF hash, page, bbox, date, and edition; Gorkhapatra remains RISK; and no qualified native pixel-label review exists. The split additionally has 18 exact and 575 total 16+-codepoint train-to-eval overlap rows, plus 136 dev-to-test substring rows and four duplicate-image groups in train. A deterministic geometry sample found adjacent-line ink in two of the three widest crops. Do not train, score, publish CER, or redistribute this pack. Evidence: outputs/newar-devanagari-normalized-claim-boundary-audit-2026-07-15-r12/ and outputs/newar-devanagari-normalized-visual-sample-review-2026-07-15-r12/.
  • Four-profile Newari routing is not four-profile data admission (2026-07-15 r30; supersedes r1 capability state). Newar/Prachalit, Bhujimol, Ranjana, and Devanagari have explicit runtime and registry profiles. The recovered deva-v2 export is now exact-hash bound as one locally operational diagnostic model, while every profile remains claim-ineligible and no four-script OCR claim exists. The model's narrative historical score is not admitted: its recovered Newari evaluation pack fails the rights, provenance, native-review, leakage, and split-independence gates documented above. Bhujimol has no admitted font or real pixel-label set. Ranjana remains unencoded in Unicode 17.0. The later r36 custody audit supersedes only this paragraph's font-license finding; it does not change the r30 data/model result. Evidence: outputs/newari-four-writing-system-support-audit-2026-07-15-r30/.
  • The recovered Devanagari model now has a runtime smoke, not an accuracy result (2026-07-15 r31). The exact r30-bound export loaded in a pinned CPU-only Paddle 3.3.1 / PaddleOCR 3.7.0 runtime on omac and returned one nonempty Devanagari prediction for a single OFL-font synthetic Newari line. The rejected historical evaluation pack was not used, no CER or other quality metric was computed, and the imperfect prediction is retained only as mechanism evidence. The model remains excluded from redistribution and every quality claim. Evidence: outputs/newari-devanagari-runtime-smoke-audit-2026-07-15-r31/.
  • The failed-fit Newar/Prachalit checkpoint now has exact export custody, not a quality result (2026-07-15 r33). The surviving native-v1 checkpoint was exported on the 64 GiB omac host with CPU-only Paddle 3.3.1 from an unchanged hash-pinned PaddleOCR checkout. The export binds three inference files, the exact 106-symbol dictionary, the Apache 2.0 PaddleOCR license, and the existing CC BY 4.0 Zenodo 6967421 attribution. It remains a known failed fit: the original PPLCNetV3 route exposed 40 CTC timesteps while 95.7% of its training labels exceeded 40 characters. Its historical norm_edit_dis is not admitted as codepoint CER, no runtime smoke has yet passed, the Cambridge labels remain qualified-review pending, and the model binary is retained locally rather than redistributed through Git. This raises exact local diagnostic-model custody to two of four Newari profiles; it does not establish Prachalit quality or a four-script OCR claim. Evidence: outputs/newari-four-writing-system-support-audit-2026-07-15-r33/.
  • The failed-fit Newar/Prachalit export now has a runtime smoke, not an accuracy result (2026-07-15 r34). The exact r33 export loaded on the 64 GiB omac host in a network-denied, CPU-only Paddle 3.3.1 / PaddleOCR 3.7.0 runtime and returned one structurally valid prediction for one 1,200 x 180 synthetic line rendered with Noto Sans Newa 2.007. The font's embedded SIL OFL 1.1 metadata and SHA-256 are bound; the font binary is not redistributed. The executed runner recorded U+11Cxx source-text metadata in error. Its raw JSON and exact runner bytes remain unchanged, while the reviewed U+114xx Newa render provenance and corrected maintained runner are separately bound. The model read one image and made one prediction, with zero training, GPU, external-network, evaluation-row, output-to-source comparison, or scoring actions. Therefore no CER, accuracy, redistribution admission, Prachalit quality claim, or four-script OCR claim follows. Evidence: outputs/newari-prachalit-runtime-smoke-audit-2026-07-15-r34/.
  • Nithya Ranjana font licensing and mapping are now audited, not OCR capability (2026-07-15 r36). The local DU binary exactly matches Ek Type's official 1.000 release (8a2ebe62…b92b), and the official NU binary was acquired from the same release (22a12a91…1528). The release ZIP, all seven members, tag commit, release metadata, embedded license records, standalone OFL texts, cmap counts, and GSUB script tags are hash-bound. Ek Type states that Ranjana is unencoded and provides DU on Devanagari Unicode and NU on Newa Unicode. DU is therefore reference-only for this project's Newa output route; NU is the compatible Ranjana-style render candidate. Neither font is a Bhujimol font or proof of historical manuscript-domain fidelity. No synthetic training payload, real pixel, label, model, prediction, score, or CER is admitted. Evidence: outputs/newari-historical-font-custody-audit-2026-07-15-r36/.
  • The CC0 Bhujimol vector chart is a reviewer reference, not a font (2026-07-15 r37). Wikimedia Commons records uploader Z3lvs, own work, CC0 1.0, and a 2026-05-24 creation date for the exact 181,286-byte SVG. Structural inspection finds 111 outline paths and no text element, embedded font, path ID/codepoint map, shaping rule, script, external resource, or raster image. It can orient a qualified reviewer to basic shapes, but cannot render arbitrary Newa text, attest conjunct behavior, validate a manuscript hand, or supply labels. The bounded font search still produced no admitted Bhujimol font. No manuscript pixel, training row, model, prediction, score, or CER is admitted. Evidence: outputs/newari-bhujimol-reference-chart-audit-2026-07-15-r37/.
  • Bhujimol/Ranjana discovery is still catalogue-only (2026-07-15 r2). Six source candidates now have rights evidence and blockers recorded. The strongest Bhujimol lead is Internet Archive Asa Archives item 1147: uploader-applied CC0 metadata, 63 scanned pages, and 325 machine-detected line boxes. Those boxes are not labels, and uploader authority, institutional accession, exact-file SHA-256, qualified script identity, pixel faithfulness, and transcription mapping remain unresolved. The strongest Ranjana rights lead is Smithsonian image FS-6314_06, specifically marked CC0, but its IIIF manifest contains only one canvas even though the physical object has 77 folios. British Library EAP candidates are noncommercial or research-only; Nepalbhasa Manuscripts in UK has no captured reuse grant. Accordingly, zero pixels, labels, review manifests, models, or numbers are admitted. Evidence: outputs/newari-source-candidate-audit-2026-07-15-r2/.
  • Asa item 1147 now has a catalog identity, not a rights clearance (2026-07-15 r3). The uploader's CC0 catalog spreadsheet (SHA-256 8484dbe95358b83d7d78d5d1752fafac5feeecfc007ccf5d1fe8b6ea3ce9bffd) crosswalks the scan to DPN 1147, Bhāsvati, NRN ज्योति - २७१, RN 2919, Sanskrit/Nepalbhasa, N.S. 528, 33.3 × 4.1 cm. The catalog script field is blank. The same uploader created the hidden Internet Archive collection and uploaded both the spreadsheet and scan, but no inspected profile or metadata proves that the uploader is an authorized Asa Saphu Kuthi representative or had authority to apply CC0. The exact questions needed to close this gap are staged, not sent, in docs/NEWARI_ASA_1147_RIGHTS_REQUEST.md. The source remains excluded from every data and model path. Evidence: outputs/newari-source-candidate-audit-2026-07-15-r3/.
  • Wellcome MS Indic epsilon 1 reopens the Bhujimol real-pixel route (2026-07-15 r4). The official Wellcome manuscript record identifies shelfmark MS Indic epsilon 1; Wellcome's editorial account explicitly calls its script Bhujimol. The complete IIIF manifest (SHA-256 596296750114948b39da2252b5a7b7c0738287dadcf7aed32939bfc29980c6d4) has 217 unique canvases/services and applies Public Domain Mark 1.0 with explicit commercial copying/modification/distribution permission. Five 2000-pixel-wide visual probes spanning canvases 1, 50, 100, 150, and 210 decoded; four sampled leaves contained dense multi-block text and the first was a cover/label view. This is source-scale and geometry evidence only, not qualified script review or ground truth. The r9 handoff acquired and hash-recorded all 217 JPEGs on omac (58,471,370 bytes, 217 unique hashes); raw pixels remain remote and the orchestration host mirrors metadata only. The r10 audit creates a blank 217-row review worksheet. Zero qualified page decisions, lines, labels, evaluation rows, models, or metrics are admitted.
  • Wellcome MS Indic delta 33 does not provide a complete Ranjana source (2026-07-15 r5). The physical catalogue record resolves shelfmark MS Indic delta 33, Sierra number b14901468, and closed-store availability, but exposes no digital item; the predictable IIIF presentation URL returns
    1. Four official selected-image records carry CC BY 4.0 and identify Ranjana. A prediction-free probe found one two-leaf overview, one partial-leaf environmental view, and two close-up miniature crops that overlap the overview rather than independent pages. Catalogue titles say AD 1653 while one record description says transcribed in 1635; the inconsistency is preserved. These licensed assets are useful reference evidence, but cannot supply 200 independent audited lines or a manuscript-disjoint split. Zero pixels, labels, evaluation rows, models, or metrics are admitted.
  • BDRC MW1NPL132 supplies a complete Ranjana review source, not a dataset (2026-07-15 r6). BDRC's physical-instance metadata identifies Ārya nāmasaṁgīti only as ScriptRanj and Manuscript; linked scan W1NPL132 is AccessOpen/StatusReleased, and its exact IIIF manifest has 21 unique canvases/services. BDRC's official access policy classifies scanned original manuscripts as public domain and permits sharing/adaptation for any purpose. The manifest has no license field; work metadata CC0 is kept separate from the pixel-rights basis. Five visual probes establish access and plausible running-text density only. The frozen r6 catalogue and omac handoff admit zero audited lines, labels, models, or metrics. The r7 handoff execution hash-recorded and dimension-checked all 21 JPEGs on omac; raw pixels remain remote and the local packet contains metadata only. Qualified all-page script review and an actual line count remain required.
  • Second Bhujimol and Ranjana manuscript groups are remotely acquired, not admitted (2026-07-15 r12-r16). The exact Commons photograph of Cambridge MS Add.2137 is CC BY-SA 4.0 and contains four folio photographs. The first three are the selected 28r, 28v, and 29r review units; the official Cambridge manifest identifies its hands as Bhujimol, but Cambridge's complete scan is CC BY-NC 4.0 and its pixels remain excluded. The Commons original is now fetched and SHA-256-recorded on omac. BDRC MW1NPL23 is an open, released 25-canvas manuscript governed by BDRC's scanned-manuscript public domain policy, but its institutional record assigns both ScriptNewa and ScriptRanj without page-level attribution. It therefore requires qualified all-page classification before any Ranjana line can be selected. These are distinct manuscript groups from the first Wellcome Bhujimol and BDRC Ranjana sources. The r14 execution hash-recorded all 26 source images (12,738,550 bytes, 26 unique hashes) on omac; the local r15 packet contains metadata only and exposes 28 blank review units. It admits zero local pixels, qualified decisions, audited lines, labels, models, or metrics. Evidence: outputs/newari-second-manuscript-source-catalog-2026-07-15-r12/ and outputs/newari-second-source-remote-acquisition-audit-2026-07-15-r15/. The refreshed central r16 inventory records four fetched-and-hash-recorded sources in total and still zero admitted sources. The consolidated r17 packet binds 264 unique remote images to 266 blank review units across the four manuscript groups; neither profile has a qualified decision or an audited line, so both >=200-line gates remain false.
  • Available Bhujimol work text does not remove review (2026-07-15 r43-r45). The Wellcome manuscript is identified as the Aṣṭasāhasrikā Prajñāpāramitā, but the available GRETIL text is a different edition with reference-only terms and no canvas/folio/line crosswalk. BDRC MW1NPL99 is institutionally titled “Unidentified,” so it has no known companion. Cambridge MS Add.2137 interleaves a published Sanskrit Nāradasmṛti root with Newari commentary-cum-translation; Cambridge's TEI says the Newari legal text and translation are unpublished and supplies only short catalog excerpts. The fail-closed r45 audit therefore records zero exact line-aligned transcriptions, zero admissible prefills, and zero labels. The r43/r44 UIs separately correct the Cambridge display ambiguity by rendering bounded 28r, 28v, and 29r regions while retaining the full CC BY-SA source hash and link. Work identity may guide a reviewer; it cannot be auto-accepted as manuscript ground truth. Evidence: outputs/newari-bhujimol-companion-text-audit-2026-07-15-r45/, outputs/newari-qualified-page-review-dispatch-2026-07-15-r43/, and outputs/newari-qualified-page-review-ui-2026-07-15-r44/.

Open legal questions (need human/legal resolution, not agent research)

  1. Gorkhapatra Sansthan output under Nepal Copyright Act 2059 §3 — underlies every claim-grade model to date and the live HF Space.
  2. Aale Chhatka September-2023 license — one email to the publisher.

Sikkim Herald permission correction (2026-07-15)

The earlier statutory-term analysis was incomplete because it did not inspect the Government of Sikkim State Portal's express copyright permission. The official copyright page permits reproduction of material featured on the site, free of charge, in any format or media, subject to accuracy, non-derogatory/non-misleading use, and prominent source acknowledgment. It excludes material explicitly identified as third-party. The same official portal hosts the Sikkim Herald archive, and the Government's IPR administration page identifies Sikkim Herald as the official paper published by the Information & Public Relations Department.

This is a source-specific permission, not a declaration that government works are public domain. Admission is limited to Herald editions hosted on the State Portal and carrying government-publication evidence. The newer official IPR portal is useful for discovery, but its separate subdomain does not currently expose the same legal page; its PDFs remain catalogue-only unless mirrored on sikkim.gov.in or covered by separately captured permission. Builders must reject explicitly third-party material, retain the source URL and PDF hash per edition/row, and carry the permission conditions and attribution into redistributed crops, labels, models, and published numbers. The evidence snapshot and current portal enumeration are recorded at outputs/sikkim-herald-source-rights-2026-07-15/.

The r11 source review applies that boundary to six full-hash editions and 26 pages. Reviewer 1 observed no explicit third-party marker on 25 pages and bounded both complete press photographs on Lepcha 2024-10-04 page 1 for exclusion because incidental storefront signage is visible. This does not establish government authorship or license cleanliness for every page. The frozen record remains pending_independent_reviewer_2_and_adjudication, and no line labels, training rows, evaluation rows, models, or numbers are admitted. See outputs/sikkim-herald-six-edition-manual-review-r1-2026-07-15-r11/.

The r13 Gurung Khema reopening applies the same boundary to two official State Portal Gurung Editions dated 2022-03-01 and 2022-04-22, separated by 52 days. Reviewer 1 inspected all eight pages and recorded no explicit third-party marker; this is not an authorship finding. The blinded reviewer-2 worksheet is still blank, and an independent packet audit reports PASS_READY_NOT_ADMITTED. No source string, converted string, crop, label, prediction, score, training row, or CER has been admitted. The Unicode proposal and its embedded reference font are decode/attestation evidence only and are not copied into the frozen outputs. Evidence: outputs/sikkim-herald-gurung-khema-two-edition-source-review-2026-07-15-r13/, outputs/sikkim-herald-gurung-khema-font-semantic-audit-2026-07-15-r13/, and outputs/sikkim-herald-gurung-khema-r13-adversarial-bundle-audit-2026-07-15-r13/.

Standing remediations

  • Add O'Neill/SOAS + upstream attribution wherever Zenodo-6967421 data or models appear publicly.
  • Produce a dated manifest of ampixa/ocr-tech-artifacts (HF) contents and reconcile the live limbu-ocr Space against spaces/limbu-ocr/.
  • Fonts without license text stay out of the render pool.