# Testing expectations without turning them into exclusions

Operational research protocol, 4 September 2026. Read with the frozen `HYPOTHESES_AND_EXPECTED_CANDIDATES_v1.md` and active `METHODS_GLOBAL_COMPARABILITY_v3.md`. This adds test definitions; it does not retrospectively change the 100 expected profiles or manufacture an empirical ranking.

## What the expected Top 100 is for

The working expectation is that the final portfolio will contain a substantial high-reach group including Indonesian, Hindi, Bangla, Simplified Written Chinese and Urdu; other strong initial investigations include Telugu, Tamil, Marathi, Hausa and Kiswahili. The 100-profile document also anticipates regional-language, displacement, research, signing and interlanguage opportunities. Those are research expectations, not reserved places or funding quotas.

The surprise worth investigating is not merely a different order. It is a result without a plausible explanation: a large additional-language route counted only by mother tongue; one language retained and its relevant comparator absent; all research needs inferred away from school provision; or all shared-register benefits forced to zero.

The 100 identifiers are not ranks. The five thematic sections are not priority bands. A candidate outside the frozen table can outperform every candidate inside it. Neither an entry's section nor the order in which its evidence is collected can be an input to benefit or compute estimates.

## Keep every stage eligible

The original F/S/V/U/R labels identify initial research leads, not allowed stages. Every retained profile remains eligible for foundational, secondary, vocational/practical, university and research interventions. Foundational includes preschool/early learning as a distinct substage where applicable. A specific intervention still needs prerequisites and a coherent endpoint; a research-language profile need not claim to teach pre-literacy in order for that community's early-learning route to remain represented.

For each profile and stage, retain an explicit disposition: investigated, not yet investigated, or outside the specifically defined intervention's scope with a substantive reason. Missing measurements, an omitted initial-stage tag, small population, national income and local programme completion are never such reasons. The default is not yet investigated, not zero benefit. An out-of-scope intervention does not erase the community's other interventions.

The bounded executable checkpoint constructs all five stage cells for the current 100 profiles and for synthetic missing-data fixtures. It checks that changing, deleting or reversing the initial-stage annotations cannot change those cells. This proves a property of this constructor, not that the eventual global ranking implementation has passed its full inclusion audit. The same tests must be run through the actual allocation pipeline before publication.

## Distinguish three kinds of statement

1. **Empirical predictions** have a fixed scope, observable contrast and stated failure rule. Record supported, weakened, contradicted or unresolved; inadequate evidence is unresolved, not support.
2. **Exploratory expectations** suggest what may matter before effect sizes or comparable inputs are known. The overall expected Top 100 and H01's broad high-reach expectation currently belong here. Freeze the forecast now; freeze numerical scenario definitions before the next order, and report whether those scenarios reproduce it. Do not claim a forecast is confirmed merely because a chosen prior produces it.
3. **Diagnostic obligations** govern the method. H03's investigation symmetry, H04's identity retention, and H12's missingness/local-programme invariance are not claims that related populations have equal needs. Unequal outcomes can be correct; unexplained disappearance cannot.

## Bounded empirical tests

These comparisons use direct source evidence where available, explicit transport where justified and labelled scenarios otherwise. All proposed packages remain hypotheses rather than authorized translation work.

| Test | Scope and expected direction | Support or contradiction rule | Current disposition |
|---|---|---|---|
| T01: additional-language reach | Indonesian, Urdu and Kiswahili, separately: contemporary ability/home-use routes may exceed an L1-only audience. Compare each source within its actual frame. | Support requires source evidence of additional-language reach or defensible joint bounds; an L1-only measure cannot settle it. A source-defined route with no additional-language users contradicts that route's prediction. Do not add incompatible surveys. | Indonesian has a supported bounded example; Urdu and Kiswahili need additional construct-matched evidence. No common numeric uplift is assigned. |
| T02: subject-language transition | Indian regional-language tertiary entry, Tanzania secondary entry and Niger primary transition, separately: local-language explanation plus target-medium subject practice can address a documented transition not answered by conversational-language instruction alone. | Support for the mechanism requires evidence of the actual transition and a capability/material gap; evidence of no transition or already comfortable subject access weakens that route. A translated package's learning effect is a separate hypothesis. | Initial institutional evidence supports several transition mechanisms; package effects remain unresolved. |
| T03: Ukrainian/Russian differentiated packages | Compare Ukrainian and Russian provision for school catch-up, university prerequisites and research/reference use. Expect relevant opportunities in both, but potentially different strongest packages. | A claim that one is covered by the other requires explicit community, subject, language-capability and access evidence. For a proposed package, complete equally usable independent provision can weaken additionality. A different audience size alone neither confirms nor contradicts the package mechanism. | Both have evidenced routes; the present comparison does not establish a global efficiency order. |
| T04: stage-specific provision | Japanese, Simplified Written Chinese, Ukrainian and Russian: compare independently available school quantitative material with an exact advanced subject/source package. | The prediction of an advanced gap is weakened for each package by an existing accessible equivalent with the required content and prerequisites. School provision alone is not evidence either way. Record failed package hypotheses instead of moving to a new topic without disclosure. | Package-level comparison remains open; no language-wide absence-of-material claim is made. |
| T05: mathematical bridge advantage | Compare a specified interlanguage's formula-rich and prose-heavy explanatory passages for the same prepared reader profile; retain definitions, negation, quantifiers and proof dependencies. | The expectation is greater retained task benefit for formula-rich passages, at matched content/quality. Equal or lower retention weakens it. Literature or model/mechanical evidence supports only its stated endpoint, not observed human learning. Cost advantage additionally requires r > C_bridge/C_separate. | Explicit design/retention scenario; no fabricated comprehension rate. |
| T06: portable access | Compare an offline lesson package with an online-only counterpart on the same low-storage device, after one complete download and disconnection. | Navigation, instructional content, exercises and answers must remain available offline. Broken dependencies contradict the operational access claim. This test does not by itself establish learning gain or smartphone ownership. | Delivery experiment defined; not yet run on a new teaching package. |
| T07: practical-subject additionality | For Bangla, Hausa and Kiswahili, compare a specified numeracy/enterprise package against independent provision for the same tasks. | Evidence of unmet task/prerequisite/format needs supports the package hypothesis; an equally usable existing route weakens it. An institution offering a general business degree proves neither universal access nor absence of need. | Source-backed package proposals exist; incremental learning benefit remains unmeasured. |

T01–T07 do not exhaust all useful hypotheses. They prevent the initial mechanisms from remaining untestable slogans. Before an additional empirical claim is promoted to a finding, it receives the same scope, comparator and failure-rule fields. A later source or better explanation may revise a hypothesis, but the original prediction and its outcome stay in the record.

## Mechanism-based comparison groups

These groups trigger related questions; they are not equivalence classes or population pools.

| Comparison group | Why compare it? | What must not be copied? |
|---|---|---|
| Russian / Ukrainian / Belarusian / Polish / Bulgarian / Serbian profiles | Shared regional academic and educational histories can motivate comparable stage/supply questions. | Population, present disruption, resources, script ability, language preference or rank. |
| Indonesian / Malay / Javanese / Sundanese | National/additional-language reach, related-register reuse and first-language routes overlap in different ways. | A national language's coverage of every native-language learner. |
| Hindi / Urdu / Punjabi / Bhojpuri / Maithili and other separately identified communities | Learned registers, census aggregation, scripts and school-language transitions can conceal distinct access routes. | Census-group totals, Nastaliq/Devanagari/Gurmukhi literacy or universal mutual intelligibility. |
| Bangla / Sylheti / Chittagonian, with Bangladesh and India separate | Broad census labels can mask different spoken/written and educational profiles. | Exact-variety comprehension from a broad category label. |
| Hausa Nigeria / Hausa Niger | A shared language connects different instruction-language systems and provision. | English/French transition rules or a national rate from the other country. |
| Kiswahili Tanzania / Kenya / separately scoped regional routes | A cross-border language can support different L1/L2 and schooling pathways. | A regional speaker estimate into each country's audience. |
| Persian / Dari / Tajik; Turkish / Azerbaijani / Turkmen profiles | Related registers may support source reuse or shared designs. | Scripts, local terminology or whole-family comprehension. |
| Japanese / Chinese / Korean / English / French research profiles | Extensive schooling can coexist with specific advanced or accessible-format gaps. | A wealth label as evidence of no need. |
| Exact signed languages and their surrounding written-language routes | Modality barriers are not captured by spoken/written population tables. | A country's deafness total or written-language proficiency as sign-user counts. |
| Displaced communities across host countries | The same people may need continuity and host-medium subject access. | Refugee totals as one language count, or duplicate counts across residence/history. |

New evidence can add comparison groups. Genealogy is one lead among educational policy, shared publication markets, migration and modalities; it is not itself a need score.

## What happens when an expectation fails

For every expected candidate absent from a reported selection, record its candidate/stage/package and the exact reason. Distinguish identity mismatch; initially missing population, need, supply, overlap or cost data; stage filtering; implementation error; and evidence-supported lower benefit under the stated scenario. Only the last is an empirical allocation result. An unresolved comparison stays visible without a numeric bottom rank.

A surprising paired result triggers two checks:

- Trace both candidates from identity through source inputs, inference, candidate generation and output.
- Run the implicated rule across the **whole retained candidate-stage universe**, not just the noticed pair. For example, an inner join on population data requires checking every initially unmeasured identity, not adding one Ukrainian row.

The possible conclusions are evidence-supported asymmetry, an unresolved comparison with explicit assumptions, or an implementation change followed by a bounded rerun. None requires matching the expected result, assigning equal scores, a new human study or a human approval step.

## Required initially missing and counterfactual fixtures

- Identity present, population absent from the start.
- Population present, stage-specific supply absent from the start.
- Signed-language profile present, written tokenizer and production-cost measurement absent.
- Unresolved exact identity retained with no fabricated registry match.
- Existing observation deleted after candidate generation.
- Initial stage annotations changed or removed.
- Local-programme completion/cost/files changed while global inputs are fixed.

All must retain an output disposition. Missing data must not be translated to zero, common midpoint or a lower ordinal position. Defensible ranges may inform a comparison; where even a useful range is unidentified, report that state rather than inventing one. Test outputs need not pretend that every wholly unmeasured candidate is an established potential winner.

## Checkpoint boundary

The current executable tests concern the 100-profile stage constructor and synthetic fixtures. The existing universal registry/source accounting is a separate checkpoint. Neither is a claim that all global stage evidence, package comparisons, numeric forecasts or publication QA are complete. The next substantive step is to populate and compare the stage/package evidence, including negative results, then run these tests through the actual allocation pipeline.