Working Note 003
Six adjectives got seven of the nine checks
I measured my own notes and turned the counts into a voice document. Against a controlled test it beat the adjective list already in the repo on two checks out of nine, and both were habits a style guide would have talked me out of. The other seven the adjectives already had.
Finding
Describing a voice and measuring one are not the same size of improvement. Six adjectives sitting in this repo since the beginning turned out to carry most of what I thought only counting could give me. What counting added was narrow: two patterns out of nine, both of them things standard writing advice actively discourages.
That is a smaller result than the one I set out to report, and I only have it because the first run of the experiment was broken in a way I caused.
Why I looked at this
Not as a writing exercise. I draft these notes with AI and then edit the voice back in by hand, sentence by sentence, and I wanted to stop paying that tax. The repo already had a voice section — clear, honest, practical, reflective, slightly informal, systems-minded, followed by a list of things to avoid.
I assumed it was too vague to constrain anything — that assumption is the part that did not survive.
The question
Does measuring published work produce a better voice instruction than describing it — and if it does, which parts of the instruction are actually doing the work?
What I expected
That the adjectives were decorative, and that measurement would beat them across the board.
What I tried
Counted two notes with a script. Median sentence sixteen words, a third of sentences under twelve, twenty-seven bold spans of which every one was either a complete declarative sentence opening a paragraph or a project name, seven contractions against twenty-two expanded negations, one em dash per hundred and ten words.
Wrote those counts up as a voice document. Then tested it properly — the part I had skipped in both previous notes.
Four arms, same source material, same section headings, only the guidance differing: no guidance, the six adjectives, the measured document, and a fourth arm of competent generic writing advice matched to the measured document’s length. Each ran in a separate context that had seen nothing else. A nine-check rubric, written before any arm ran, scored them mechanically.
What happened
Five things, and the first invalidates the run it belongs to.
The first run was confounded, and I built the confound myself. I made the note’s subject the voice measurement, so the shared brief stated the target patterns as facts. Every arm received the treatment. The one subject guaranteed to leak was the one I chose.
Re-run on an unrelated subject, the arms separated. No guidance scored four. The adjectives scored seven. The measured document scored nine.
Matching the word count did not close the gap. The fourth arm — competent generic advice at four times the adjectives’ length — scored six, slightly below the sixty words it was meant to overwhelm, so volume was not the explanation.
The advantage sat in two checks. Excluding bold usage and em dash rate, the adjectives and the measured document tie at seven out of seven. Everything else I had written down was general craft that sixty words already delivered.
Both differentiating patterns are ones a style guide argues against. Standard advice says use bold sparingly and convert most dashes to commas. My notes do the opposite, at twelve to fifteen bold spans each and triple the recommended dash rate.
Cutting the document by a third changed nothing. Removing the general craft left eight hundred and ninety-six words, and it scored nine again.
Where it broke
I scored a file while it was still being written. The adjective arm was mid-write when I measured it, and I reported a number that was wrong by two points before the completion notice arrived and I checked the hash. For a note about measuring things carefully, this is the failure that should embarrass me most.
The instrument was wrong three times, and I found the third inside this note. A check scored the proportion of bold spans rather than the count, so an arm using bold twice passed identically to one using it eleven times. A search pattern then reported a missing limitation that was present three times over, phrased differently. Scoring this note, I found the sentence splitter had been merging every bolded claim with the sentence after it — which inflated one measure across every note I have written. The tooling was at fault each time, and each time it flattered me.
The winning arm scored above me. Nine of nine, against my own eight. An imitation that hits stated targets more reliably than the person does is demonstrating that it can follow numbers — which is not the same as sounding like anyone.
The generic arm is not a clean control. Its advice actively opposed my practice on two of the three checks it failed, because I wrote it and standard advice genuinely says those things. It tests whether good generic guidance wins, not whether length alone does.
There is no held-out sample. Both existing notes went into building the document, so nothing tested whether it generalises to a note it was not fitted on. That is the third consecutive note with a missing control — and this time I knew in advance, and ran out of material rather than out of care.
Current view
Provisional, and considerably smaller than what I believed at lunchtime.
Measurement earns its keep on idiosyncrasy, not on craft. The general qualities of writing — rhythm, concreteness, honest hedging — are available from any decent style guide, or from six well-chosen adjectives, and writing them down again bought nothing I could detect. What counting recovered was the handful of habits that are mine rather than good, and which generic advice would have edited out.
If you are writing a style instruction from your own published work, measure the artefact, then throw away everything the measurement agrees with received wisdom about. The residue is the part worth keeping, and it will be shorter than you expect.
The honest status of the document is that it is followable. Whether it sounds like me is untested, and this note is the first evidence either way.
What I’d try next
Score this note against the same rubric. It is the held-out sample the last three notes have been missing, and it exists now only because it got written.
Then the harder version: give someone the two source notes and this one with the titles removed, and ask which was written under the document. A rubric can tell me the em dashes landed. It cannot tell me the thing I actually want to know.
Related: 001 had a missing control, 002 had a missing comparison, and this one has a missing holdout. The pattern is stable enough now to be the finding.