Reversibility Is Not Detection: Measuring Round-Trip Pseudonymization in an Inflected Language (Replication Package)

Replication package for "Reversibility Is Not Detection: Measuring Round-Trip Pseudonymization in an Inflected Language". Reversible pseudonymization replaces identifiers with realistic surrogates and later restores the originals. The article measures whether that restoration reproduces the original text character for character in Polish, where surrogates must be inflected to stay grammatical. It shows that high detection recall does not guarantee recoverability. Contents:- data/corpus/: the controlled corpus of 1,202 model-generated Polish documents with per-mention annotations (grammatical case, role, person identifier). It contains no personal data: names are sampled from public registry frequency lists (data/surnames/, data/first-names/, CC0 1.0).- data/generator/: the corpus generator: prompts, bucket definitions, prompt assembly and sampling of persons, homonyms and content (no API keys).- data/rejected_records/: generator records rejected at validation, with the errors that caused each rejection.- data/roundtrip_raw.jsonl: raw request and response bodies of the pseudonymization service on the corpus, including the state objects that carry its substitution maps.- data/T*.csv: every result table used in the article.- code/: analysis scripts, including the Section 8.3 pilot script (no API keys).- reference_outputs/: frozen copies of all outputs for byte-level comparison.- reproduce.sh: one command that reruns the analyses and compares every output with its reference copy (see REPRODUCE.md). The system's responses on MultiGraSCCo are not included because they carry the corpus text; the package carries their SHA-256 checksums. KPWr 1.3 and MultiGraSCCo are cited, not redistributed: the package ships only their SHA-256 checksums, and person names from KPWr in derived files are replaced by tokens. See EXTERNAL_CORPORA.md. Funding: European Union, European Funds for Lubelskie 2021-2027, Priority I, Action 1.3, project FELU.01.03-IP.01-0021/24.