What is STRR and how does it complement fertility as a tokenizer fairness metric?

**STRR**, the single-token retention rate, is the fraction of a language's words that the tokenizer represents as one token instead of splitting into pieces. If a language has 10,000 words and 3,000 of them are single tokens, its STRR is 0.3. This is different from fertility, which is the average number of tokens per word measured over a sample of real text. Fertility tells you the typical cost; STRR tells you how much of the whole vocabulary got dedicated single-token coverage. The reason both matter is that fertility is dominated by the words people use most. A language can have a handful of very common words that are single tokens, which keeps the average low, while thousands of less-common words still shatter into many pieces. The average looks fine, but coverage is shallow. STRR catches exactly that gap, because it looks at the vocabulary as a whole and is not weighted by how often each word appears. So fertility answers "what does typical text cost?" and STRR answers "how broadly does the tokenizer actually cover this language?" Together they give a fuller fairness picture than either alone.

**STRR**, the single-token retention rate, measures one thing. It is the fraction of a language's vocabulary that maps to exactly one token. The numerator counts words tokenized as a single token; the denominator is the total word list. Fertility, by contrast, is the average tokens per word over a representative corpus. The crucial difference is weighting. Fertility is frequency-weighted, so a few very common single-token words dominate the average. STRR is unweighted over the word list, so every word counts equally. That difference is exactly why the two are complementary. A tokenizer can post a low, healthy-looking fertility because the most frequent words happen to be single tokens. Meanwhile the long tail of content words fragments into three or more pieces. The corpus average never sees that tail clearly. STRR does, because it asks what fraction of the whole vocabulary got single-token coverage, regardless of frequency. So fertility tells you the typical cost a user pays on ordinary text, and STRR tells you the breadth of the tokenizer's coverage across the language. Reporting both is the honest fairness scorecard: one for average cost, one for tail coverage.