# The Load-Bearing Vocabulary of Claude

Louis Abraham samples about 1,000 GitHub pull requests a day and groups their descriptions by the words they use, with no labels and no list of suspect words. One group appeared in 2026 and kept growing, and its most characteristic words read like a Claude session transcript: *quietly, nobody, plainly, re-derived, halves, outright, load-bearing, survived, rests, mutation-checked*. The page names it "the load-bearing vocabulary of Claude" [[load-bearing-vocabulary-claude]].

The clipped page is a rendered dashboard and carries little prose: the corpus size at capture (609 days, 473,669 pull requests, 53,976,283 words), the word the board had selected, and the word list. The selected word was *quietly*, 36.43 times more frequent in the cluster than outside it, at 86 per million words at its peak against 14 per million across the whole corpus. The method below comes from the project's README at https://github.com/louisabraham/load-bearing, which is not saved in `sources/`.

## How it is measured

GH Archive could not be used. Since mid-2025 the public event feed carries almost nothing but push events, which hold no text, and an earlier version of the project built on it undercounted `load-bearing` by a factor of 158. The collector uses GitHub's search API instead: ten five-minute windows a day, one drawn at random from each 2.4-hour block, seeded on the date so the corpus is reproducible. Four bot apps and empty bodies are excluded in the query itself. Accounts whose login marks them as bots are dropped, identical word sets within a week are collapsed, and no author contributes more than three descriptions a week. A word enters the vocabulary once 50 distinct accounts have written it, which is why `load-bearing` (848 accounts) stays and a term used 242 times by two accounts does not.

The model is k-means with KL divergence in place of squared distance: each of ten "ways of writing" is a distribution over words, and each description goes to the nearest one. There is no time parameter in the fit, so any rise over time comes from how the descriptions were assigned rather than from anything the model was told. At the README's snapshot (to 2026-08-17), one component went from 0.7% of descriptions at the start of 2025 to 39% by mid-2026, still growing by about 1.2 points a week. Words are ranked by the ratio of their frequency inside the component to outside it, with a pseudo-count so rare words cannot top the list by accident. `load-bearing` appears 929 times inside and 82 outside, 39×, at the top.

The author is open about the soft spots. `K = 10` was chosen by looking at the outcome: below ten the component is mixed with unrelated vocabulary, above fourteen it splits. The random seed moves the headline number. A single unconditioned fit finds the arriving component in 31 of 32 runs. The same fit also works as a naive Bayes classifier, published as `detector.html`, which the README says "can say a text is written like that cluster and never who wrote it".

## What the name claims

Nothing in the method knows which descriptions an LLM wrote. The corpus excludes bot accounts, so these are descriptions posted under human accounts, and the attribution to Claude is a reading of the vocabulary and its timing. The page links one piece of corroboration, [anthropics/claude-code#53454](https://github.com/anthropics/claude-code/issues/53454), opened in April 2026 by a user on Opus 4.7 whose Claude Code had started using "load-bearing" in chat, documents and commit messages, and kept doing so after being told to stop. That is strong circumstantial evidence, not a measurement of authorship, and other models trained on similar data could write the same way.

## Why it matters here

It shows a writing tell arriving in public, measured week by week, which is what [[ai-detector-arms-race]] predicted: surface tells change with each model release, so a fixed list goes stale. The vault's own list, [[meta/tropes]], has "quietly" as its first entry, among magic adverbs used to convey subtle significance; *load-bearing*, *mutation-checked* and *re-derived* are not on it, and this cluster suggests they belong in the next revision. The vault is maintained by Claude Code, and at least 25 pages in `wiki/` already use "load-bearing", so the vocabulary is in this vault's own prose as well.

The approach sits next to [[ai-comment-classifier]], which built a calibrated human-vs-LLM classifier for code comments from public data. It differs in having no labels at all. What it cannot settle is the question [[credibility-as-slop-test]] and [[checking-for-human-writing]] argue is the real one: a description written in this vocabulary may still be accurate and checked by the person who posted it. Where it touches code review, it gives [[agent-principal-agent-problem]] a number: the effort signal a pull request description used to carry is now missing from more than a third of them.