/declauding

Where entries 43 to 47 come from

Entries 1 to 42 were promoted from drafts, one specimen at a time, under the rule in SKILL.md: a phrase earns an entry after it shows up twice. Entries 43 to 47 came the other way round, from a corpus that had already counted the phrases. This file records what was measured, so a later reader can check the entries against the evidence instead of taking them on trust.

The source

Louis Abraham's load-bearing samples GitHub pull request descriptions — ten five-minute windows a day drawn by a date-seeded RNG, bodies fetched through the search API, bot logins and mass-posters filtered out — and clusters them by the words they are written with. Ten clusters, hard assignment, KL k-means with no time parameter anywhere in the fit, so a cluster's weekly share is attribution after the fact rather than a trend the model was free to draw.

The numbers below are from analysis.js generated 2026-08-28: 595 days in 85 whole weeks, 2025-01-06 to 2026-08-17, 461,121 descriptions, 51,079,244 word appearances, 19,798 words past the 50-distinct-authors floor. k=10, SEED=6, 8 fits, cheapest published.

One of the ten clusters was 0.70% of the first eight weeks and 39.5% of the last four. The least-squares line over the last twelve weeks is +1.24 points a week. Its highest-lift words, by the ratio of inside-frequency to outside-frequency:

load-bearing 39x   plainly 34x   quietly 30x   refusal 28x   survived 28x
re-derived 27x     halves 27x    asserted 25x  nobody 25x    genuinely 24x
deliberately 24x   premise 23x   refuses 23x   outright 23x  byte-identical 23x
ruling 22x         genuine 22x   handed 22x    carries 21x   died 20x

The register this skill produces

That is not the slop vocabulary of entry 20. There is no delve, no tapestry, no robust. It is flat, concrete, verdict-shaped technical prose: the register this skill's Overcorrection section describes as the target, where the writing is plain-and-sure, facts are short, and failures are stated dryly.

The register this skill produces is the fastest-growing cluster in the corpus. Entries 43 to 47 exist because a subtractive pass that removes 42 staging tics and lands the draft in a register 39% of GitHub now writes has traded one detectable shape for another.

Putting the staging back would be worse. What the entries ask instead is that flat certainty be checked the way staging is, by the same generative test one register over: am I stating the finding, or performing having settled it? An author reaches for these shapes when a sentence has to sound settled.

The measured rates and their limits

Rate of the cluster's top-150 vocabulary as a percentage of body-prose word tokens, measured the way declaude_lint.py measures it — headings, tables, code fences and, where marked, quoted specimens excluded:

The stdlib figure is a floor rather than a fair control. API reference prose is a different genre, so some of the separation is genre and not authorship.

The third row is the one that settles what the rule can claim. Louis Abraham's README scores 1.51, above this skill's SKILL.md, on nothing x9, carries x3, alone x2, never x2 and half x2. Quoted specimens are already excluded, so they do not account for it. He writes this way, and writes it well.

So the rate is a register locator, not an authorship detector. It answers one question: is this draft written in the cluster's register. The corpus-register density line in declaude_lint.py says so in its note. That wording carries weight. A reader who takes the line for a detector will start cutting nothing and measured out of correct sentences, which is entry 23's failure mode with a number attached to it.

That is also why there is no blocklist here. Every word in the list is a word a person writes. Entries 43 to 47 fire on the shape each family builds, and each one carries an earned column because each family carries claims.

Reproducing it

git clone --depth 1 https://github.com/louisabraham/load-bearing
python3 - <<'EOF'
import json, re
s = open('load-bearing/analysis.js').read()
d = json.loads(s[s.index('{'):s.rindex('}') + 1])
lead = set(d['components'][0]['word_list'][:150])
words = re.findall(r"[A-Za-z0-9/_-]*[A-Za-z][A-Za-z0-9/_-]*", open('DRAFT.md').read().lower())
print(100 * sum(w in lead for w in words) / len(words))
EOF

The list ships in declaude_lint.py as CORPUS_LEAD, so re-running that against a fresh analysis.js is also how the list gets updated. The cluster is one fit's answer, and load-bearing's own §5 says the seed moves the headline, so treat a shifted list as a shifted sample and not a correction to this one.

The top of the table

The two highest rows are this skill's own files: references/register.md at 3.17, this one at 2.65. Both are documents whose subject is the vocabulary, so the result is expected, and it is also the demonstration. A rate reports where the prose sits. Authorship and quality are outside what it measures.