Repository navigation
Count em-dashes in prose, not in a document's own markup - #80
Merged
Merged
Conversation
`rhet.em-dash` counted every "--" in the raw text. In a Markdown document that charged the writer for things nobody typed as punctuation: - frontmatter delimiters (---) - table separator rows (|---|---|---| counts three) - command-line flags inside fenced code blocks (--json, --against) - a line of hyphens typed above a bibliography The worst case is not ours. pelic-021970 in the calibration corpus is an essay by a second-language learner who drew one separator line of forty hyphens above their references. The engine counted twenty em-dashes and printed "LLMs lean on the em-dash as a rhythm crutch" on their report. That finding is now gone. What moved, measured before and after: file dashes density severity README.md 139 -> 73 3.9 -> 2.0 High -> Medium Docs/Calibration/README.md 40 -> 15 2.6 -> 1.0 Medium -> Low SKILL.md 47 -> 31 2.0 -> 1.4 Medium -> Low CONTRIBUTING.md 23 -> 15 1.7 -> 1.1 Medium -> Low pelic-021970.txt 20 -> 0 2.7 -> none Medium -> no finding What did not move: the published calibration. Re-run over all 296 texts, `published-calibration.json` is identical except its date — threshold 30, 2 of 296 flagged, both bounds, every language row, the noisiest-rule shares. This removes a false signal; it does not move the number. Two hyphens between words is still how people type an em-dash without the key for it, and that still counts. Three or more is a separator, never a dash between clauses. Each of the six guards was checked by mutation: all six fail against the old counter. Found from the outside, using the skill on a real email and then the rule pack against ourselves. The brief that reported it concluded "nothing calculates wrong" and proposed rewriting our own strings; the strings are a separate and much smaller question, and rewriting SKILL.md to lower its score is the one thing this project tells everyone never to do. It drops to Low on its own. Known and left alone: words inside code blocks still enter the density denominator, which understates it for technical documents. That lives in the shared word count, not in this rule. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015PEbbiYSNPw7jE3LrPNhyF
This was referenced Sep 7, 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.
The defect
EmDashAnalyzercounted every--in the raw text as an em-dash. In any document with markup that charges the writer for things nobody typed as punctuation:---)|---|---|---|counts as three)--json,--report,--against)The case that decides it is not ours
pelic-021970.txtin the calibration corpus is an essay by a second-language learner who drew one separator line of forty hyphens above their references. The engine counted twenty em-dashes and printed "Em-dash overuse (20 in 745 words, 2.7/100). LLMs lean on the em-dash as a rhythm crutch" — on a report a teacher might hand back to that student.It is the same shape as #75: a rule charging a learner for their register, except here the student did not even write the thing being counted.
What moved
README.mdDocs/Calibration/README.mdSKILL.mdCONTRIBUTING.mdpelic-021970.txtWhat did not move
The published calibration. Re-run over all 296 texts:
published-calibration.jsonis identical except its date — threshold 30, 2 of 296 flagged,rateLow,rateHigh, every language row, the noisiest-rule shares. Corpus fingerprint unchanged (78bda061bde3dc99).That matters for the house rule about tuning the instrument: this removes a false signal, it does not move the number.
What still counts
Two hyphens between words is how people type an em-dash without the key for it, and it still counts. Three or more is a separator, never a dash between clauses. Dashes inside fenced blocks and inline code spans are markup.
Tests
Six guards, each verified by mutation — all six fail against the old counter — plus two regression guards that the real dash and the ASCII stand-in still count. 424 tests green.
Left alone on purpose
Words inside code blocks still enter the density denominator, which understates density for technical documents. That is a defect in the shared word count, not in this rule, and changing it would touch every analyzer and the corpus. Filed separately.
Provenance
Found from outside the repo, using the skill on a real email and then the rule pack against ourselves (
AI_Tasks/Pendientes/OPUS_05_…). That brief concluded "nothing calculates wrong" and proposed rewriting our own suggestion strings. The strings are a separate and much smaller question; rewritingSKILL.mdto lower its score is the one thing this project tells everyone never to do — and it drops to Low here on its own, as a consequence rather than a goal.🤖 Generated with Claude Code
https://claude.ai/code/session_015PEbbiYSNPw7jE3LrPNhyF