Commit d91b607f by PLN (Algolia)

docs: cadence measures baseline on a 15-text Gutenberg human corpus

Antithesis and bookends separate the article fixture from human prose by
an order of magnitude; staccato and sentence-cv do not (dialogue-heavy
novels are staccato and bursty). Report-only, no thresholds.
parent b58d2f8b
# Cadence measures — human corpus baseline
Fetched 2026-11-11: the 15 most popular English public-domain novels on
Gutenberg (gutendex, `copyright: false`), plain-text bodies trimmed to the
text between the `*** START/END OF THE PROJECT GUTENBERG EBOOK ***` markers
so license blocks don't skew the bookends. This is the human baseline the
project requires before any cadence rule could ever fire — same discipline
as the < 2% per-rule false-positive gate for catalog v3. **No thresholds are
set from this run; the measures stay report-only.**
| # | Text (author, year) | Words | Parags | staccato | triads /1000w | antith /1000w | bookends | sent-CV |
|---|---|---:|---:|---:|---:|---:|---:|---:|
| 1 | Complete Works of Shakespeare (1623) | 963 460 | 40 921 | 15.2% | 2.03 | 0.01 | 0.00 | 1.32 |
| 2 | Alice in Wonderland (Carroll, 1865) | 26 525 | 817 | 45.7% | 0.53 | 0.00 | 0.00 | 0.91 |
| 3 | Jane Eyre (Brontë, 1847) | 185 390 | 4 115 | 37.0% | 1.38 | 0.02 | 0.00 | 0.90 |
| 4 | Pride and Prejudice (Austen, 1813) | 127 360 | 2 509 | 33.7% | 0.60 | 0.02 | 0.00 | 0.88 |
| 5 | Middlemarch (Eliot, 1871) | 316 381 | 4 786 | 17.1% | 0.55 | 0.01 | 0.00 | 0.86 |
| 6 | Romeo and Juliet (Shakespeare, 1597) | 25 958 | 1 098 | 18.1% | 1.58 | 0.04 | 0.00 | 1.26 |
| 7 | Sherlock Holmes (Doyle, 1892) | 104 506 | 2 546 | 38.2% | 0.73 | 0.00 | 0.01 | 0.80 |
| 8 | Dorian Gray (Wilde, 1890) | 78 979 | 1 525 | 27.7% | 0.68 | 0.04 | 0.00 | 1.01 |
| 9 | Crime and Punishment (Dostoevsky, 1866) | 203 612 | 3 984 | 35.1% | 1.36 | 0.00 | 0.00 | 0.88 |
| 10 | Moby Dick (Melville, 1851) | 212 796 | 2 814 | 16.7% | 1.42 | 0.00 | 0.00 | 0.96 |
| 11 | Dracula (Stoker, 1897) | 161 321 | 2 190 | 30.7% | 0.66 | 0.01 | 0.00 | 0.75 |
| 12 | Little Women (Alcott, 1868) | 192 557 | 4 691 | 36.4% | 1.82 | 0.03 | 0.03 | 0.80 |
| 13 | Dr Jekyll and Mr Hyde (Stevenson, 1886) | 25 631 | 362 | 22.9% | 1.09 | 0.00 | 0.02 | 0.85 |
| 14 | The Blue Castle (Montgomery, 1926) | 68 274 | 1 575 | 27.1% | 0.59 | 0.00 | 0.00 | 0.80 |
| 15 | Frankenstein (Shelley, 1818) | 75 042 | 797 | 12.4% | 1.17 | 0.00 | 0.00 | 0.65 |
| — | **Human medians** | | | **27.7%** | **1.09** | **0.01** | **0.00** | **0.88** |
| — | AI article excerpt (test fixture) | 295 | 12 | 33.3% | 0.00 | **10.17** | **0.11** | 0.52 |
## What the baseline honestly says
- **Antithesis and bookends separate hard.** Human max antithesis is
0.04/1000w against the article's 10.17 — three orders of magnitude. Human
max bookends is 0.03 against the article's 0.11. These two carry the
signal, and any future threshold would sit comfortably above every human
text here.
- **Staccato does not discriminate at all.** Dialogue-heavy novels are
naturally staccato: Alice hits 45.7%, Sherlock Holmes 38.2% — both above
the article's 33.3%. A staccato "alarm" would flag Austen before it flags
a chatbot. The measure stays descriptive.
- **Sentence-CV runs the other way.** The article (0.52) is *below* every
human text (0.65–1.32): fiction is bursty because dialogue alternates
with long narration. Low CV is a hint, not a verdict, and fiction is the
wrong control for it.
- **Triads overlap completely** (human 0.53–2.03/1000w; the article excerpt
scores 0 because it contains none). The full article's ~a dozen triads
would land inside the human range.
Consequence for calibration: if cadence thresholds ever exist, they belong
to antithesis and bookends first; staccato, triads and CV stay manual
reading aids. Fiction is also the wrong-only corpus — essays and journalism
would shift these medians, so the corpus should widen before any number is
trusted.
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment