How the blending engine works

The blender splits each name at its vowel boundaries, joins the opening of one to the ending of the other at every combination of those boundaries, then measures four properties of each candidate that a program can genuinely check in spelling. It never decides whether a blend is good.

This page documents all of it, including the parts that are judgement calls and the parts the engine cannot see at all. If something here is wrong or unclear, the contact page is the fastest way to get it fixed.

Step one: finding the split points

Most combiners cut a name in half. That is why they produce so much unusable output, since the midpoint of a name rarely falls where a syllable ends. This engine walks the name instead, finds each group of vowels, and marks a split after the vowel group and after each consonant that follows it.

The letter y counts as a vowel except when it opens a name or directly follows another vowel. That single rule is why Taylor splits after Tay rather than treating ay as one unit, and it is the difference between finding Tayvis and never finding it.

Run on taylor, the engine returns split points at 2, 3, 4, 5, giving these openings:

• ta
• tay
• tayl
• taylo

The same points are used from the other end to produce the pieces a name can contribute as an ending. Every opening is then paired with every ending from the other name, in both directions, and results shorter than three or longer than fourteen characters are dropped.

Step two: source coverage

Coverage is the share of a name that survives into the blend, measured in characters from the contiguous piece that name contributed. If Taylor gives Tay, coverage is three of six, so 50 percent. If Alex gives Ale, coverage is three of four, so 75 percent.

This is the score that matters most, and it is worth being precise about what it does not say. It counts characters, not recognisability. Jordan contributing dan scores the same as Jordan contributing Jor, even though Jor is the part most readers would recognise. That gap is a real limit of the measure and it is documented in the limits report.

Step three: balance

Balance is one minus the difference between the two coverage figures. Two names each giving up half of themselves score a perfect balance. A blend where one name contributes 75 percent and the other 25 percent scores 0.5, and reads as one name with something attached rather than as a combination of two.

The report describes balance in words rather than numbers: even, close, uneven, lopsided. A precise decimal would imply a precision the measure does not have.

Step four: spelling flags

Flags mark spellings that tend to cause trouble for a reader seeing the blend for the first time. A flag is not a disqualification. Flagged candidates still appear in the results, because a hard-to-spell blend is a fact about the candidate rather than a verdict on it.

• Doubled letter at the join. The same character on both sides of the seam, which makes the boundary ambiguous.
• Three identical letters in a row. Almost always a spelling people correct without noticing.
• A run of three or more consonants. Readers stumble over these, and they are the most common cause of a blend being pronounced two different ways.
• Three or more vowels in a row. Leaves the stress unclear.
• Unusual length. Over eleven characters is long for a tag or a handle. Under four means very little of either name is left.
• Identical to an input. The blend came out the same as one of the names you typed, which means it is not a blend.

Step five: the join pattern

The join is the seam where the two pieces meet, and the report names the two characters that meet there. Four cases are possible, and they behave differently when somebody reads the blend cold.

• Vowel meets consonant. The most forgiving seam, and the one that most often survives a first read.
• Consonant meets vowel. Usually clean, and keeps the two halves distinct.
• Consonant meets consonant. Can build a cluster the reader has to work through.
• Vowel meets vowel. The boundary goes soft and the blend invites more than one pronunciation.

How the ordering works

Candidates are sorted by a single number: average coverage weighted at 0.55, balance at 0.45, minus a small penalty for drifting from the length that suits the chosen context and a small penalty per spelling flag. Context changes the ideal length only. It never changes what the report says about a candidate.

That weighting is a judgement call and we would rather state it than hide it. Weighting coverage above balance means a blend that keeps more total letters can outrank a blend that splits more evenly. Taylor and Travis is a live example: Traylor and Travlor both rank above Tayvis, because each keeps more characters overall, while the world uses Tayvis. The engine is not wrong, it is measuring something different from what people choose on, and that gap is the reason this site does not claim a best result.

What the engine does not measure

• Pronunciation. The engine reads spelling, not sound, so it cannot tell you how a blend will be said.
• Phonetic flow. Whether a blend is pleasant to say is a judgement the software does not make.
• Cultural associations. A blend can be structurally clean and mean something unfortunate in another language or community.
• Community usage. If a fandom already uses a different name for the pairing, that convention beats any structural score.

These four are not oversights waiting to be fixed in a later version. Three of them are outside what spelling can tell you, and the fourth, community usage, is a fact about people rather than about names. The review guide covers how to check them yourself, and the blind name test turns the most important one into something you can actually run.

Determinism and versioning

The engine has no randomness in it. The same two names and context return the same candidates in the same order every time, which is why a shared link reproduces a result exactly and why the regression checks mean anything. This page carries the engine version, and it changes when the scoring changes.