# Where educational translation is likely to help: hypotheses to test

Research protocol, 4 September 2026. These are prospective expectations for the next analysis, written after seeing problems in an earlier analysis. They are not an original preregistration, measured results, funding recommendations, or a ranked Top 100.

## Research question

Which combinations of language, educational material and delivery format are likely to create the most additional usable educational access per standardized unit of AI compute?

The expected pattern must be stated before inspecting the next ranking. Otherwise a missing community can be mistaken for a low-need community. A surprising result is a reason to investigate the hypothesis, data and implementation. It is not a reason to force the result to match expectations.

Existing work by this programme is a case study. Its presence, absence, completion, local files and sunk compute have no effect on global educational-need estimates or candidate inclusion.

## Testable hypotheses

| ID | Hypothesis and expected pattern | What would support or weaken it? | Consequence of a surprising absence |
|---|---|---|---|
| H01 | Large language communities with substantial stage-specific unmet need should repeatedly appear among high-reach candidates. Indonesian, Hindi, Bangla, Simplified Written Chinese and Urdu are strong initial expectations for investigation, not guaranteed winners. | Compare affected learner counts and usable local provision at the same stage. High local provision or small additional benefit can weaken a particular package case; national wealth or local project completion cannot substitute for that evidence. | Trace the candidate through every inclusion, population, overlap and scoring step. Missing data cannot settle the comparison against it. |
| H02 | Where a regional language supports schooling but advanced study moves to another language, bridge-to-tertiary resources may have especially high utility. | Need evidence for the actual transition, learners' capabilities and subject gaps. NPTEL explicitly identifies this mechanism in India. Strong advanced local resources or comfortable academic-language access would weaken the predicted benefit. | Compare Tamil, Telugu, Marathi, Bangla, Odia and other relevant communities; do not infer one language's provision from another. |
| H03 | A need mechanism found in one community should generate hypotheses for related educational systems and neighboring communities. Russian-language research or self-study gaps imply checking Ukrainian, Belarusian, Polish, Bulgarian and other relevant communities. | Shared educational histories, academic registers or resource routes make an analogy worth testing. Conflict, displacement, different publication markets and different language policies may increase or decrease the gap. Genealogical relationship alone supplies neither an equal score nor a population estimate. | If Russian is present and Ukrainian absent, investigate source coverage, population definitions and implementation before interpreting the result. Repeat this test globally. |
| H04 | Large national or census language categories may conceal distinct communities with different access needs. | Compare the category definition with exact mother-tongue rows, scripts, schooling and bilingualism. Examples include Hindi-group varieties and Bengali categories containing Sylheti and Chatgaya. Shared classification is not measured comprehension. | Retain the group observation and its possible constituent populations. Resolve overlap or use bounds; never silently discard a constituent or credit it with the whole group. |
| H05 | Educational-resource shortages differ by stage. A language can have extensive school resources and substantial postgraduate or research gaps. | Compare curriculum coverage and source availability separately for school, vocational, university and research use. Japanese, Chinese, Russian, English, French and other established academic languages remain candidates for specific missing materials. | Investigate the proposed advanced package, rather than excluding the entire language as already served. |
| H06 | A modest number of well-designed interlanguage editions could sometimes serve more learners per compute than separate editions. The advantage may be greater in formula-rich mathematics for readers with adequate prerequisites. | Compare exact reader communities, scripts, mathematical background, prose burden, semantic-error patterns and standardized production cost. Evaluate definitions, negation and proof dependencies as well as formulas. Existing literature, source analysis and model/mechanical tests can support provisional scenarios without waiting for new participants. | Investigate bridge design and domain fit. Missing a comprehension estimate is not evidence of zero reach; family population is not evidence of full reach. |
| H07 | Availability of some translated resources will often leave important curriculum and delivery gaps. | Compare actual subjects, prerequisite sequences, exercises, answers, offline availability, accessibility and licensing. One catalogue is not a census of all provision; a listed file is not measured learner access. | Ask what is missing and what would add value, rather than treating either a catalogue entry or its absence as a final need estimate. |
| H08 | Signed-language and accessibility interventions can create substantial access that written-language counts miss. | Use sign-language-specific populations and educational evidence, not total deafness counts. Compare text, signed explanation, captions, MathML, audio and tactile routes by endpoint. | An entirely spoken/written result set requires an explicit modality audit. A missing written tokenizer is not a reason to remove signed learners. |
| H09 | Offline, phone-usable, exercise-rich resources can benefit motivated learners without teachers, while classroom resources can help a different or overlapping audience. | Estimate autonomous study, classroom use and research/reference use separately. Test file size, navigation, prerequisite support, answers and device requirements. Do not assume every person owns a smartphone or that a download guarantees learning. | Investigate independent-learning routes whenever a model implicitly requires teachers or continuous internet. |
| H10 | Conflict, displacement, statelessness and cross-border language use can change both need and delivery routes without changing the language's registry homeland. | Use dated displacement and education evidence, separate current location from language identity and avoid double-counting people across countries. | Census-country absence triggers a route search, not exclusion. Ukraine is one test; other displaced communities require the same treatment. |
| H11 | The best source packages may differ materially across communities: foundational learning, science, computing, statistics, finance, management and advanced research may each dominate somewhere. | Link package recommendations to local educational or practical needs and independent existing provision. OpenStax and Open Logic provide comparison baselines, not an exclusive subject universe. | A uniform proposal to translate the same book everywhere requires justification. |
| H12 | Raw population, tokenization and missingness can each dominate a numerical model for the wrong reason. | Run factor-removal, source-removal, cost-standardization, common-prior sensitivity and counterfactual missing-data tests. Keep gross, cached and fresh tokens distinct. | Investigate any ranking that is nearly a population table or that drops a community after evidence is removed. |

H03 is a symmetry test for investigation, not a demand for identical regional outcomes. For example, evidence could support a larger Ukrainian-language resource gap than a Russian-language gap, while the Russian-language audience is larger. These are separate quantities to estimate, not a geopolitical ordering to choose in advance.

## Provisional expectation for a 100-candidate shortlist

The table names 100 candidates that I currently expect to be worth comparing for a high-impact portfolio. Numbers are identifiers, not ranks. The first group contains many strong high-reach expectations; later groups contain regional, modality and bridge hypotheses whose marginal benefit could outperform a larger language's particular package. Group size is not a funding quota.

Each row names an edition or comparison profile, not a claim that all people under that label are one population. Multi-country and multi-script rows must expand into separate evidence cells. Different scripts do not create additional people. A shared-register proposal is not counted as an existing language or as guaranteed coverage.

Stage codes: **F** foundational/primary; **S** secondary; **V** vocational/practical; **U** university; **R** postgraduate/research. These are stages to investigate, not findings about what every learner needs.

### A. Broad-reach and major educational-language candidates

| ID | Candidate edition/profile | Population or script distinctions to preserve | Initial stages | Hypotheses |
|---:|---|---|---|---|
| 001 | Indonesian | Latin; Indonesia; L1 and additional-language users separately | F/S/V/U/R | H01,H07,H09 |
| 002 | Hindi | Devanagari; exact Hindi versus census-group mother tongues | F/S/V/U/R | H01,H02,H04 |
| 003 | Bangla | Bengali script; Bangladesh and India separately | F/S/V/U/R | H01,H02,H04 |
| 004 | Standard Written Chinese, Simplified | Mainland written-literacy profiles; not all Chinese varieties or all residents | S/V/U/R | H01,H05,H07 |
| 005 | Urdu | Perso-Arabic/Nastaliq; Pakistan and Indian communities separately | F/S/V/U/R | H01,H02,H09 |
| 006 | Telugu | Telugu script; regional school-to-tertiary routes | S/V/U/R | H01,H02,H07 |
| 007 | Tamil | Tamil script; India and Sri Lanka separately | S/V/U/R | H01,H02,H10 |
| 008 | Marathi | Devanagari; Maharashtra learning routes | S/V/U/R | H01,H02,H07 |
| 009 | Vietnamese | Vietnamese Latin orthography | S/V/U/R | H01,H05,H11 |
| 010 | Hausa | Latin first profile; Ajami separately; Nigeria and Niger separately | F/S/V/U | H01,H09,H10 |
| 011 | Standard Kiswahili | Latin; national standards and L1/L2 learning routes separately | F/S/V/U | H01,H02,H10 |
| 012 | Modern Standard Arabic | Learned written register; country, schooling and vernacular profiles separately | S/V/U/R | H01,H04,H06 |
| 013 | Brazilian Portuguese | Brazil; independent provision and disadvantaged learners | F/S/V/U/R | H01,H07,H09 |
| 014 | Spanish, Latin American educational profiles | Country-specific terminology and underserved routes, not a continent-wide count | F/S/V/U/R | H01,H07,H11 |
| 015 | English, underserved learner profiles | Territory and capability cells; targeted open/offline/accessible gaps | F/S/V/U/R | H05,H08,H09 |
| 016 | French, underserved learner profiles | African and other contexts separately; French ability cannot cover all residents | S/V/U/R | H02,H05,H10 |
| 017 | Turkish | Latin; Türkiye and diaspora separately | S/V/U/R | H01,H05,H10 |
| 018 | Iranian Persian | Perso-Arabic; distinguish Dari and Tajik | S/V/U/R | H01,H05,H06 |
| 019 | Thai | Thai script; standard versus other Thai-area languages | S/V/U/R | H01,H04,H05 |
| 020 | Burmese | Myanmar script; distinguish other Myanmar languages | F/S/V/U | H01,H04,H10 |
| 021 | Filipino | Latin standard; Tagalog and other Philippine language access separately | S/V/U/R | H01,H02,H04 |
| 022 | Standard Malay, Malaysia | Latin profile; Jawi and other regional standards separately | S/V/U/R | H01,H05,H06 |
| 023 | Gujarati | Gujarati script | S/V/U/R | H02,H07,H11 |
| 024 | Kannada | Kannada script | S/V/U/R | H02,H07,H11 |
| 025 | Odia | Odia script; standard versus census-group scope | S/V/U/R | H02,H04,H07 |
| 026 | Malayalam | Malayalam script | S/V/U/R | H02,H05,H07 |
| 027 | Eastern Punjabi | Gurmukhi; do not substitute Western Punjabi/Shahmukhi counts | F/S/V/U | H02,H04,H10 |
| 028 | Russian | Cyrillic; resident and cross-border academic users separately | S/V/U/R | H03,H05,H10 |
| 029 | Japanese | Japanese writing; targeted subject and advanced-source gaps | U/R | H05,H07,H12 |
| 030 | Korean, South Korean standard | Hangul; North Korean and diaspora profiles separately | U/R | H05,H07,H10 |

### B. Regional and language-transition candidates

| ID | Candidate edition/profile | Population or script distinctions to preserve | Initial stages | Hypotheses |
|---:|---|---|---|---|
| 031 | Amharic | Ethiopic; do not substitute Ethiopia's total population | F/S/V/U | H02,H04,H11 |
| 032 | West-Central Oromo educational profile | Latin; other Oromo varieties separately | F/S/V/U | H02,H04,H11 |
| 033 | Somali | Latin standard; regional and displaced learners separately | F/S/V/U | H09,H10,H11 |
| 034 | Yoruba | Standard Latin with full marks; national and diaspora routes separately | F/S/V/U | H02,H09,H11 |
| 035 | Igbo | Standard Latin with tone/orthographic choices specified | F/S/V/U | H02,H09,H11 |
| 036 | Kinyarwanda | Latin; independent local supply audit | F/S/V/U | H02,H07,H11 |
| 037 | Kirundi | Latin; do not copy Kinyarwanda provision or reach | F/S/V/U | H02,H03,H11 |
| 038 | Lingala | School/written and urban spoken registers separately | F/S/V/U | H02,H04,H11 |
| 039 | Central Kanuri | Latin and Ajami routes separately | F/S/V | H04,H09,H10 |
| 040 | Wolof | Latin first profile; Wolofal and Garay separately | F/S/V/U | H02,H08,H11 |
| 041 | Bambara | Latin and N'Ko routes separately | F/S/V/U | H02,H06,H11 |
| 042 | Pulaar | Latin and Adlam profiles; Senegal/Mauritania and other contexts separately | F/S/V | H04,H08,H10 |
| 043 | Nigerian Fulfulde | Exact regional profile; no whole-Fulani population assignment | F/S/V | H04,H08,H10 |
| 044 | Zulu | Latin; school and higher-education transition | F/S/V/U | H02,H07,H11 |
| 045 | Northern Sotho / Sepedi standard | Latin; census and standard identity boundaries retained | F/S/V/U | H02,H04,H11 |
| 046 | Haitian Creole | Standard Haitian orthography; Haiti and displaced communities separately | F/S/V/U | H02,H09,H10 |
| 047 | Nepali | Devanagari; native and additional-language profiles separately | F/S/V/U | H02,H04,H11 |
| 048 | Sinhala | Sinhala script; specific Sri Lankan educational routes | S/V/U/R | H02,H05,H11 |
| 049 | Northern Pashto | Perso-Arabic; Afghanistan/Pakistan and Southern Pashto separately | F/S/V/U | H04,H09,H10 |
| 050 | Dari | Perso-Arabic; Afghanistan and displaced learners separately | F/S/V/U | H06,H09,H10 |
| 051 | Sindhi, Pakistan standard | Perso-Arabic; Indian/Devanagari profiles separately | F/S/V/U | H02,H04,H10 |
| 052 | Saraiki | Perso-Arabic; no automatic Urdu or Punjabi coverage | F/S/V | H02,H04,H09 |
| 053 | Bhojpuri | Devanagari profile; Indian and Nepalese populations separately | F/S/V | H02,H04,H09 |
| 054 | Maithili | Devanagari profile; Tirhuta and territory distinctions retained | F/S/V/U | H02,H04,H10 |
| 055 | Chhattisgarhi | Devanagari; distinguish Hindi census grouping from comprehension | F/S/V | H02,H04,H09 |
| 056 | Haryanvi | Devanagari; actual Hindi overlap to estimate | F/S/V | H02,H04,H12 |
| 057 | Santali | Ol Chiki first profile; other script-literacy routes separately | F/S/V | H04,H08,H09 |
| 058 | Sylheti | Bengali-script and Sylheti-Nagri profiles separately | F/S/V | H04,H08,H10 |
| 059 | Chittagonian | Spoken and written-register design separately; no automatic Bangla coverage | F/S/V | H04,H08,H10 |
| 060 | Javanese | Latin first profile; registers and Indonesian overlap separately | F/S/V/U | H03,H04,H12 |
| 061 | Sundanese | Latin first profile; Indigenous-script route separately | F/S/V | H03,H04,H12 |
| 062 | Cebuano | Latin; Filipino/English proficiency measured separately | F/S/V/U | H02,H04,H11 |
| 063 | Central Khmer | Khmer script; Cambodia and other communities separately | F/S/V/U | H02,H09,H11 |
| 064 | Lao | Lao script; no automatic Thai coverage | F/S/V/U | H03,H06,H11 |
| 065 | Northern Uzbek / Uzbekistan standard | Latin and Cyrillic profiles; Southern Uzbek separately | S/V/U | H02,H06,H10 |
| 066 | Kazakh | Cyrillic first profile; actual script cohorts separately | S/V/U/R | H02,H06,H10 |
| 067 | Tajik | Cyrillic; do not assign Persian/Dari literacy automatically | S/V/U | H02,H06,H10 |
| 068 | Uyghur | Perso-Arabic first profile; diaspora script routes separately | F/S/V/U | H04,H08,H10 |
| 069 | Central Kurdish / Sorani | Perso-Arabic; country and educational systems separately | F/S/V/U | H04,H09,H10 |
| 070 | Northern Kurdish / Kurmanji | Latin first profile; other scripts and regional varieties separately | F/S/V/U | H04,H09,H10 |

### C. Comparators, research access and regional-depth candidates

| ID | Candidate edition/profile | Population or script distinctions to preserve | Initial stages | Hypotheses |
|---:|---|---|---|---|
| 071 | Ukrainian | Cyrillic; residents and displaced learners separately | F/S/V/U/R | H03,H05,H10 |
| 072 | Polish | Latin; advanced gaps and refugee-learning routes separately | S/V/U/R | H03,H05,H10 |
| 073 | Romanian | Latin; Moldova and Romania separately | S/V/U/R | H03,H05,H07 |
| 074 | Bulgarian | Cyrillic; no Russian or Interslavic coverage assumption | S/V/U/R | H03,H05,H06 |
| 075 | Belarusian | Cyrillic first profile; actual preferred/usable educational language | S/V/U/R | H03,H05,H10 |
| 076 | Serbian standard | Cyrillic and Latin; other BCMS standards separately | S/V/U/R | H03,H05,H06 |
| 077 | Tigrinya | Ethiopic; Ethiopia, Eritrea and displaced learners separately | F/S/V/U | H03,H09,H10 |
| 078 | Kabyle | Latin first profile; Arabic/French overlap separately | F/S/V/U | H04,H08,H11 |
| 079 | Central Atlas Tamazight | Exact variety; Tifinagh/Latin routes and Moroccan standard separately | F/S/V | H04,H08,H11 |
| 080 | Tachelhit | Exact variety; Tifinagh/Latin/Arabic literacy separately | F/S/V | H04,H08,H11 |
| 081 | Ayacucho Quechua | Specified regional educational orthography | F/S/V | H04,H09,H11 |
| 082 | Cusco Quechua | Specified regional educational orthography; not all Quechua | F/S/V | H04,H06,H11 |
| 083 | Paraguayan Guarani | Standard orthography; bilingual and predominantly Guarani profiles | F/S/V/U | H02,H04,H11 |
| 084 | Central Nahuatl | Exact variety and local orthographic route; not all Nahuatl | F/S/V | H04,H09,H11 |
| 085 | K'iche' | Standard plus regional comprehension profiles | F/S/V | H04,H09,H11 |
| 086 | Tok Pisin | Standard written register; L1/L2 and schooling routes separately | F/S/V/U | H02,H04,H11 |
| 087 | Bislama | Standard written register; not all Vanuatu language users | F/S/V | H02,H04,H11 |
| 088 | Kalaallisut / West Greenlandic | Greenlandic standard; other Inuit languages separately | S/V/U/R | H04,H05,H11 |

### D. Signed-language hypotheses

| ID | Candidate edition/profile | Population or script distinctions to preserve | Initial stages | Hypotheses |
|---:|---|---|---|---|
| 089 | American Sign Language | US and other actual communities separately; not universal signing | F/S/V/U/R | H08,H09,H11 |
| 090 | Brazilian Sign Language / Libras | Regional signing variation and educational stages | F/S/V/U | H08,H09,H11 |
| 091 | Indian Sign Language | Regional variation and actual sign-user population; no national deafness substitution | F/S/V/U | H04,H08,H11 |
| 092 | Kenyan Sign Language | Actual Kenyan signing communities, not Kiswahili coverage | F/S/V/U | H04,H08,H11 |
| 093 | South African Sign Language | Regional variation; English/Zulu written access separately | F/S/V/U | H04,H08,H11 |

### E. Shared-register and interlanguage hypotheses

| ID | Candidate edition/profile | Population or script distinctions to preserve | Initial stages | Hypotheses |
|---:|---|---|---|---|
| 094 | Interslavic mathematical register | Parallel Latin/Cyrillic; East/West/South Slavic reader profiles separately | U/R | H03,H06,H12 |
| 095 | Hindi-Urdu shared mathematical register | Devanagari/Nastaliq; register and literacy-specific reach | S/V/U | H02,H06,H12 |
| 096 | Malay-Indonesian shared technical profile | Exact national terminology choices; no wholesale Austronesian coverage | S/V/U | H03,H06,H12 |
| 097 | Persian-Dari-Tajik shared technical profile | Perso-Arabic/Cyrillic; lexical and schooling differences explicit | S/V/U/R | H03,H06,H12 |
| 098 | Oghuz-focused Inter-Turkic design hypothesis | Turkish/Azerbaijani/Turkmen profiles; not all Turkic languages | U/R | H03,H06,H12 |
| 099 | N'Ko literary Manding educational profile | Exact Manding communities and N'Ko literacy; no family-wide reader count | F/S/V/U | H04,H06,H08 |
| 100 | Romance mathematical intercomprehension profile | Exact Portuguese/Spanish/French/Italian/Romanian routes; natural-text and constructed alternatives compared | U/R | H05,H06,H12 |

## Challengers remain eligible

This shortlist is a prediction to challenge, not the universe from which the result must be selected. Chinese Traditional profiles, Western Punjabi, Southern Pashto, Egyptian and other spoken Arabic varieties, Xhosa, Chichewa, Luganda, Akan standards, Fijian, Samoan, Tetum, Mongolian, Armenian, Georgian, Czech, Slovak, Croatian, Bosnian, German, Italian, Ukrainian and other signed languages, Romani varieties, additional Quechua/Nahuatl/Mayan/Inuit varieties and many others can enter the final portfolio if supported. They are not assigned a score of zero or a rank below 100 by being outside this provisional table.

The proposed Inter-Germanic concept remains an explicit challenger. Investigate actual incremental access for Low German, Plautdietsch, Pennsylvania Dutch, Hutterisch, Frisian, Afrikaans, Yiddish, Scots, Faroese and other relevant profiles. Compare West/North or narrower designs against one pan-Germanic design. Do not use German/Dutch/Nordic English proficiency as a substitute for those communities' capabilities.

No group's need is removed by a bridge proposal. Native-language and bridge editions are competing or complementary interventions on overlapping populations. Report standalone benefit separately from portfolio benefit; never sum all potential audiences without overlap adjustment.

## How the hypotheses will be tested

1. Freeze this version before generating the next candidate order. Preserve it even when findings disagree.
2. Crosswalk all 100 candidate IDs and every challenger raised by evidence to the universal identity register and explicit script/community/format overlays. An unresolved mapping remains a visible candidate.
3. Record an expected direction, available evidence, incompatible evidence, an inference rule and the next discriminating source check for each hypothesis. No generic midpoint becomes an observed need estimate.
4. For each candidate missing from a resulting Top 100, distinguish: missing population; missing need/supply evidence; genuine modeled lower benefit; overlap; cost sensitivity; identity mismatch; or implementation error. Missingness and unresolved identity never count as a substantive finding against inclusion.
5. Search for unexpected high-value candidates outside this table. Compare possible outrankers under the same scenarios. Report what could change the recommendation, not just the selected names.
6. Test whether deleting a direct observation widens uncertainty instead of silently removing its population. Test whether changing local project files leaves every need estimate unchanged.
7. Publish the final empirical comparisons separately from this hypothesis table. Do not promote the expected shortlist to findings merely because a model reproduces it.

## Initial source basis and limits

The World Bank's 2021 *Loud and Clear* identifies language-of-instruction mismatch as a barrier and reports nearly 37% of students in low- and middle-income countries taught in a language they do not understand. This supports investigating H02; it is not a rate to multiply by every language's population. [World Bank](https://www.worldbank.org/en/topic/education/publication/loud-and-clear-effective-language-of-instruction-policies-for-learning).

UNESCO's multilingual-education overview describes a global access problem, including migration and language diversity. Its 40% population figure and the World Bank student figure have different denominators. Neither supplies a language-by-language ranking. [UNESCO](https://www.unesco.org/en/languages-education).

India's 2011 C16 distinguishes census language groups from named mother tongues; C17 reports up to two additional languages. These provide population and reporting evidence, not direct measures of comfortable academic access. [C16](https://censusindia.gov.in/nada/index.php/catalog/10191), [C17](https://censusindia.gov.in/nada/index.php/catalog/10262).

NPTEL describes difficulties moving from regional-language schooling to English technical education and provides translations in 11 languages. This is direct institutional support for H02 and evidence of existing supply relevant to H07, not proof that all needed material is available. [NPTEL](https://nptel.ac.in/translation).

ONS's detailed language classification groups Bengali with Sylheti and Chatgaya. This is a concrete example of why a census category cannot by itself establish a single edition's coverage. [ONS TS024](https://www.ons.gov.uk/datasets/TS024/editions/2021/versions/2?showAll=main_language_detailed).

The Unicode Standard describes N'Ko as a literary Manding register and script. This makes a shared written-profile hypothesis concrete, but a Manding speaker count is not a count of N'Ko-literate learners. [Unicode 17.0, section 19.4](https://www.unicode.org/versions/Unicode17.0.0/core-spec/chapter-19/).

These sources motivate mechanisms. They do not validate every candidate in the table or establish the eventual order. Country-, language-, stage- and package-specific investigation remains required.