Skip to main content
Shipname

Reproducible software checks

Name Blend Engine Regression Checks

Twenty deliberately varied pairs run through the same deterministic engine used by Shipname. This page catches software regressions; it is not a human preference study and does not validate whether the displayed candidates are good names.

Cases
20
Invariants met
20/20
Engine
2026.08.24-2
Reviewed
2026-07-29

What these checks can prove

Each test pair declares a minimum structural score and a maximum warning count before the engine runs. Meeting those invariants proves only that the current code produced a candidate with the expected source fragments and warning behavior. The thresholds are maintained by this project to detect changes in its own software.

These checks do not prove pronunciation, recognition, community use, preference, cultural suitability, or consent. A candidate can meet every invariant and still be rejected by a person. The limits report documents examples where that distinction matters.

Current run

Generated at request time from engine 2026.08.24-2.

Download machine-readable results
PairCategoryTop candidateStructural scoreWarningsTest invariantCheck
Taylor + TravisSimilar opening consonants create several plausible boundaries.Common LatinTayvis99082; ≤0Met
Alex + JordanTests a short first input against a longer second input.Common LatinAledan99082; ≤0Met
Sofia + MateoBoth inputs contain open vowel sequences.Common LatinSofteo97082; ≤0Met
Emma + NoahTests two compact, vowel-heavy names.Common LatinEmmoah99082; ≤0Met
Chloe + MasonThe written and spoken boundaries may not agree.Common LatinMasloe97080; ≤0Met
Daniel + OliviaTests two medium-length names with many possible cuts.Common LatinDanvia97082; ≤0Met
Bo + JoTwo-letter inputs leave almost no room for source-preserving cuts.ShortBojo77075; ≤1Met
Ann + LeeTests compact inputs and repeated-letter handling.ShortAnee94078; ≤1Met
Kai + AvaBoth names are short and vowel-heavy.ShortKava94078; ≤1Met
Elizabeth + JonathanTests whether the engine avoids simply joining two long names.LongJonabeth99082; ≤0Met
Christopher + AlexandraA large cut space should still produce compact leaders.LongAlexopher92082; ≤0Met
Aaliyah + MuhammadTests repeated letters and differing source lengths.LongMuhyah99080; ≤1Met
Mary-Jane + O'ConnorSeparators are normalized before candidate construction.PunctuationMarnor97080; ≤0Met
Anne-Marie + Jean-LucTests two compound names after punctuation normalization.PunctuationJeanarie97080; ≤0Met
Zoë + ChloéLatin diacritics are retained and flagged for human pronunciation review.DiacriticsZoloé93175; ≤1Met
José + MaríaTests source preservation when accented letters are present.DiacriticsMarosé99175; ≤1Met
Łukasz + ZofiaThe non-ASCII join requires review by a relevant-language speaker.DiacriticsZofasz99075; ≤1Met
小明 + 小红Character slicing is deterministic, but pronunciation is not scored.Non-Latin小明小红76165; ≤1Met
はるか + れんKana inputs expose the boundary of the Latin readability heuristic.Non-Latinはるかれん76165; ≤1Met
민준 + 서연Hangul output is generated but explicitly left for human language review.Non-Latin민준서연80165; ≤1Met

Reusable research asset

Download or cite this run

Both downloads are generated from the same production code path as the table above. Each row includes the input pair, declared invariant, engine version, current top candidate, warning count, and check result. The files are software-test data, not evidence of pronunciation or human preference.

Suggested citation

Shipname Editorial Team. (2026). Shipname Engine Regression Checks (Version 2026.08.24-2, reviewed 2026-07-29). https://shipname.net/research/name-pair-benchmark

Known limits

The engine scores visible character structure, not cultural meaning, gender, relationship suitability, trademark status, handle availability, or pronunciation in a specific language. Non-Latin results are candidates for human review, not linguistic validation. New engine versions may change rankings; the version and review date above make those changes auditable.

An invariant is not a recommendation

The table checks software invariants. It cannot decide pronunciation, identity, community convention, or whether a compact result is actually useful. The consolidated limits report explains eight failure modes and the human checks needed before use.

Read the name-pair limits report →Contribute a structured human review →