Fixed Text Artifact Corpus and Reproducibility Report

Research question

Can the local formatting engine remove selected mechanical artifacts deterministically while leaving clean visible wording unchanged under safe settings? This report tests that narrow behavior. It does not measure AI authorship or compare language models.

Corpus and license

Download AnyFormatLocal Fixed Text Artifact Corpus v1. The maintainer-authored synthetic fixtures are released under CC0-1.0 and contain no user text.

The corpus records prompt, model, model version, and generation date fields. Prompt and model values are explicitly null because no generative model produced these fixtures.

Method

Each fixture supplies an input string, selected rule IDs, and an exact expected output. The automated test loads the public JSON, runs the same local-format-cleaner engine used by the product, and compares strings exactly. Change counts come from engine change records; character counts use JavaScript string length.

Results

All eight version-one fixtures pass the checked-in reproducibility test. The set covers hidden Unicode, non-breaking spaces, tabs, repeated spaces, excess blank lines, opt-in Markdown removal, safe punctuation preservation, and multilingual text. This result supports only those fixed cases and the tested engine version.

Limitations

The corpus is synthetic and intentionally small. It does not represent every editor, language, Unicode code point, or document format. Passing it does not mean arbitrary text is error-free, human-written, or safe to publish. See the methodology for individual rule risks and use the local cleaner to review changes.

Reproduce or report a correction

Run the repository test suite to execute the corpus test. Report an incorrect fixture, result, or description to support@ai-text-cleaner.com; accepted corrections require a new corpus version rather than silently changing v1.