Catch the half of the conversation nobody meant to submit - #65
Merged
Conversation
A competitor's pattern set — amanmaqsood/prose-humanizer, MIT, a writing skill rather than a detector — turned out to hold 70 signals our English pack does not fire on. Six of them are worth having and forty-six are not, and the difference is the point. Every other rule here is a judgement about style, and a formal writer can lose to any of them. These six are not that. "I hope this helps" at the foot of an essay is not a register a student chose; it is an assistant's closing line, pasted in with the answer, and it says where the file has been rather than who is talented — the same kind of claim the character scanner makes. That is also why a corpus of research articles could admit them: no amount of pre-2022 academic prose tells us how often a human blogger writes "the possibilities are endless", but none of it will ever contain "as of my last training update". Screened before they were written, not after somebody complained. Each regex was counted against the calibration corpus — 249,455 words of English and 39,712 of Spanish, all published before generative models existed — and all six scored zero in both. Re-running the calibration afterwards moved nothing: same corpus fingerprint, same 0 of 90 at 25/100, same 4.1% upper bound. They cost nothing to carry. Twelve other candidates from the same source were rejected by that same screen, and they are the interesting half. Its tier-1 vocabulary list contains underpin (5 of 65 texts), optimize (6), elucidate (5), paradigm (4), exemplify (3) and illuminate (3) — ordinary research English. They fire too rarely for the human-rate gate to catch them, which needs eight texts, and too often to be called a tell. Adding them would have spent the false-positive budget of every formal writer in the corpus. This is issue #61 in the other language, found before shipping rather than after. Six more were left out for a reason worth writing down: "this essay will explore", "here's the kicker", "the possibilities are endless". These are plausible tells whose population is student essays and blog posts, and the corpus contains neither. Measuring on one population and applying to another is the mistake #59 already exists for. They need their own corpus first. Both languages, because the residue is bilingual too. 306 tests, 23 new — including the ones that check the rules leave a tutor's own "let me know" and a historian's mid-sentence "of course" alone. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015PEbbiYSNPw7jE3LrPNhyF
This was referenced Aug 24, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Six new rules in both packs, mined from a competitor's pattern set and screened against the
calibration corpus before they were written down.
Where they came from
amanmaqsood/prose-humanizer (MIT) is a writing
skill, not a detector — it says so itself, and it refuses to guess authorship. Its
rules/patterns.jsonholds 22 pattern groups. Probing each of them against our own analyzer, rather than grepping the
pack file, found 70 signals our English pack does not fire on.
Six were worth having. The difference between those six and the rest is the whole content of this PR.
What went in
chat.model-self-referencechat.training-cutoffchat.signoffchat.capability-disclaimerchat.answer-preamblechat.eager-openerEvery other rule in this project is a judgement about style, and a formal writer can lose to any of
them. These are not. "I hope this helps" at the foot of an essay is not a register a student
chose; it is an assistant's closing line, pasted in with the answer. It says where the file has been
rather than who is talented — the same kind of claim the character scanner makes.
That is also why a corpus of research articles and encyclopedia prose could admit them at all.
What the screen said
Every regex counted against the calibration corpus first — 249,455 words of English, 39,712 of
Spanish, all published before generative models existed.
123fa5b9ebca3f29,same 0 of 90 flagged at 25/100, same 4.1% upper bound. Only the run date changed.
thresholdsgave them no human rate, correctly — a rule that fired on nothing has nothing measured.What the screen rejected, which is the interesting half
Twelve candidates fire on writing published before generative models existed. Their tier-1
vocabulary list contains, in our 65 English texts:
optimizeunderpinelucidateparadigmexemplifyilluminatePlus
garner,synergy,interplay,uncover,groundbreaking, and "due to the fact that".These fire too rarely for the human-rate gate to catch them — it needs eight texts before it will
set a threshold — and too often to be called a tell. Adding them would have spent the false-positive
budget of every formal writer in the corpus. This is #61 in the other language, found before
shipping instead of after.
What was left out on principle
this essay will explore,here's the kicker,the possibilities are endless,only time will tell,without further ado,moving forward. All zero in the corpus, all plausible tells — andtheir population is student essays and blog posts, which the corpus does not contain. Measuring
on one population and applying to another is the mistake #59 already exists for. They need their
own corpus first, and that is worth opening as its own issue.
Checks
306 tests, 23 new — including the negative ones: a tutor's own "let me know when the survey is
finished", a historian's mid-sentence "of course", "here is the revised timetable the committee
agreed", and their Spanish equivalents.
🤖 Generated with Claude Code
https://claude.ai/code/session_015PEbbiYSNPw7jE3LrPNhyF