R0027/2026-03-26/Q002/H2¶
Statement¶
Linguistic structure is not the primary challenge for non-English prompt engineering. The challenges are computational (training data volume, tokenization) and structural differences between languages are secondary.
Status¶
Current: Partially supported
H2 is partially supported in an unexpected way. The evidence confirms that computational factors (tokenization, training data) account for the vast majority (72-87%) of performance failures, with direct linguistic nuances contributing only ~2%. However, H2 overstates the case by suggesting structural differences are secondary — they are secondary in direct impact but primary in causing the tokenization inefficiency that drives most failures.
Supporting Evidence¶
| Evidence | Summary |
|---|---|
| SRC04-E01 | 72-87% of failures from model limitations, only ~2% from language nuances |
| SRC03-E01 | Tokenization fertility (a computational property) predicts accuracy |
Contradicting Evidence¶
| Evidence | Summary |
|---|---|
| SRC01-E01 | Linguistic features do significantly influence prompt effectiveness |
| SRC01-E02 | Specific structural features create specific challenges |
| SRC05-E01 | Agglutinative structure directly breaks tokenization |
Reasoning¶
H2 correctly identifies the dominant mechanism (computational) but incorrectly frames linguistic structure as secondary. The relationship is causal: linguistic structure → tokenization inefficiency → performance degradation. Linguistic structure is the upstream cause of the computational challenge.
Relationship to Other Hypotheses¶
H2 captures the proximate cause (computation) while missing the ultimate cause (linguistic structure). H3 integrates both perspectives more accurately.