1. Limits of this study
The 49 applications were chosen for being real, actively maintained products rather than libraries or teaching material: Excalidraw, Cal.com, Plane, Outline, Immich, n8n and nineteen others in TypeScript; Netflix Dispatch, Langflow, Open WebUI, Prefect, Wagtail, Saleor and eighteen others in Python. Library source was measured first and discarded as unrepresentative — libraries do type work that applications never touch, and extends ranked fifteenth in a library corpus and outside the top twenty-five in application code.
Frequency is not comprehension. Knowing the thirteen most common symbols in Python does not mean you can read Python, any more than knowing the thirteen most common English words means you can read English. What the ranking establishes is the size and shape of the vocabulary, not the time it takes to acquire it.
Author-invented names are excluded, which is a judgement rather than a measurement. They are 27–30% of all tokens and carry no meaning outside the file they appear in.
Quote marks are excluded along with the string contents they wrap. The text inside a literal is the author’s own content, and a quote mark is a mark whose reading function a reader already has — neither is vocabulary to be learned.
Python’s name share is slightly inflated: the tokeniser strips triple-quoted docstrings but not all inline prose, so some English words are counted as names. This does not affect the coverage figures, which count only non-name tokens.
The 99% figure rises with corpus size — more repositories mean a longer tail. The 80–90% band is stable across corpora and is the meaningful number.
2. Results
Table 1 ranks the most frequent symbols. Select a column heading to sort.