Skip to content
Text docs0.2

Quality gates

Latest stable Based on Text release 0.2.1

The release quality evidence is a deterministic repository gate. It proves selected stable behaviors; it is not a broad statistical claim about NLP accuracy.

AreaStable evidence
Korean mixed-text tokenizationKoreanTextProcessorTest surface-token fixtures
Japanese mixed-text tokenizationJapaneseProcessorTest fixtures
mixed-language detectionLanguageDetectorExtensionsTest
request boundary failuresTokenizeMessageTest, BlockMessageTest, and processor facade tests
safety adapter behaviorrunnable example tests with sanitized failure assertions

The fixtures emphasize user-visible token surfaces. They intentionally avoid asserting every internal morphological choice, leaving room for model and dictionary improvements without hiding a user-facing regression.

Terminal window
./gradlew :tokenizer-korean:test \
--tests "io.bluetape4k.tokenizer.korean.KoreanTextProcessorTest"
./gradlew :tokenizer-japanese:test \
--tests "io.bluetape4k.tokenizer.japanese.JapaneseProcessorTest"
./gradlew :lingua:test \
--tests "io.bluetape4k.lingua.LanguageDetectorExtensionsTest"

The report was validated on macOS with JDK 21 or newer and the checked-in Gradle wrapper; Gradle 9.5.1 was observed in the recorded local output.

The report does not:

  • compare against third-party NLP systems;
  • publish precision, recall, or F1 over a large external corpus;
  • guarantee performance or memory on another host;
  • establish that a default blockword dictionary matches every product policy;
  • replace application tests for supported languages and input distribution.

State these gaps when using the evidence in a release decision. A deterministic gate is valuable because it is reproducible, not because it answers every quality question.

Add a fixture when a user-visible token, route, masking result, or sanitized error becomes part of your product contract. Add scored corpus evaluation only with a versioned corpus, explicit metric, reproducible command, and interpretation threshold.

See testing for application coverage and Aho-Corasick benchmarks for performance evidence.