Beta limitations
What the current language profiles guarantee—and what they do not.
MyanLex is in pre-release development. This page describes current boundaries; it is not an announcement that a hosted beta or stable package is available.
Regression tests are not an accuracy measurement
The current corpus is source-verified, not independently linguistically reviewed. Passing every regression case means the implementation matches those expected results. It does not establish real-world accuracy, complete language coverage, or preservation of meaning for arbitrary input. Independent review and broader evaluation remain outstanding.
| Operation | Current guarantee / scope | Important limitation |
|---|---|---|
| Detection | Advisory evidence for Unicode, Zawgyi, mixed, unknown or non-Myanmar input | Shared-subset text is ambiguous. Encoding switches inside a Myanmar run may be missed; this is not language identification. |
| Conversion | Explicit source and target; optional validateSource: true rejects detected mixed, mismatched or uncertain direction-changing input | No automatic mixed-text repair. Passing the guard does not prove correctness. Unicode-to-Zawgyi is not lossless for all Myanmar-script languages. |
| Normalization | Separate NFC operation, idempotent for supported input/runtime | Not spelling repair, encoding conversion or an input-to-output offset map. |
| Syllabification | Lossless supported Burmese written-syllable segmentation | Not dictionary words or guaranteed phonological syllables; other Myanmar-script languages are not silently treated as Burmese. |
| Orthography | Non-mutating structural diagnostics | valid: true does not mean correct spelling or meaningful Burmese. |
| Transliteration | Mechanical ALA-LC 2011 mapping; unsupported text preserved and reported | Not pronunciation, translation, lexical word division or contextual capitalization. |
| Tokenization | Lossless supported script/lexical groups with code-point spans | Burmese tokens are written syllables, not words. Script reporting and emoji grouping are scoped, not comprehensive language/grapheme detection. |
Safe integration
Preserve original input. Choose conversion direction explicitly and enable
source validation for pasted text. When it is rejected, inspect detection
results and review the source; do not repeatedly convert in both directions.
from: "auto" is not supported.
Offsets are zero-based, half-open Unicode code-point positions, not JavaScript UTF-16 indexes, byte positions or grapheme indexes. Normalization and conversion can change text length and ordering; recalculate downstream spans on their actual input.
See language operations for per-feature details. Production readiness also requires security, deployment and operational verification; linguistic regression results cannot establish those guarantees.