Vitrine IV
What the ratio is predicted by
Every seated piece was counted on both sides. The measurement is old; what is new is that it now has a predictor, and the predictor is not in the Turkish.
The criterion
The project’s own published counting criterion, printed so the figures can be re-run by another hand:
tokens matching [\w’'-]+ over the body, front matter excluded, H1 title line removed; list numerals count, the title does not
- English, over the sixty
- 50,080 words
- Turkish, over the sixty
- 35,351 words
- The book’s ratio
- 0.706
The question was malformed
For most of the stage a piece was convicted when its ratio left the band of its family — documents in one band, narration in another. Two explanations for the bands were tested and both failed. Marker tokens that cross at one by construction, such as field numerals and build directives, move the spread of the family that carries most of them from 0.061 to 0.058: they explain the form before story one exactly and the book not at all. Length gives r = −0.26, too weak to carry a band.
What holds is the English’s own function-word density. English the, a, of, to, is, and have no Turkish word: they are absorbed into suffixes or they vanish. So a piece dense in them must compress, and a piece of bare labels cannot.
- The least squares
ratio = 0.927 − 0.598 × density- Correlation, over 60 pieces
- r = −0.718
- Within a single family of 7 pieces
- r = −0.874 — stronger inside the family than across all of them, which is what shows the family is not the variable
- Residual standard deviation
- 0.034
- Beyond two standard deviations of their own prediction
- 3 of 60 — DS-042 (−0.096), DS-043 (−0.091), DS-026 (0.069)
The sixty pieces plotted are the 59 seated pieces that carry a ratio and the form before story one. DS-038 carries none, by the law in vitrine I: it is the same text on both sides, so its ratio is one by construction and measures nothing. The returned field is too short to measure and is excluded by the instrument’s own floor.
And every conviction before this one measured the wrong thing
The register’s own row, on the band that had been convicting pieces for the whole stage:
the families were a PROXY and the bands convict a piece for the DENSITY OF ITS ENGLISH. Two hypotheses tested and both failed — incompressible marker tokens (explains the PROLOGUE exactly and the book not at all) and length (r = −0.27). What explains it is the English’s own function-word density: r = −0.732 over 60 pieces, and −0.872 WITHIN family D — the within-family correlation being STRONGER than the across-family one is the proof that the family is not the variable
docs/VERIFIER-LOG.md, row V-80. The instrument is tools/vh_tr_ratios.py, and the only figure a pass may convict on is the residual.
Two figures for the length hypothesis stand on this page and they differ in the second decimal: −0.26 above, derived at this build over the 60 pieces plotted, and −0.27 inside the row, derived at the filing over the set that filing used. Neither is adjusted to agree with the other and neither is dropped. Both scopes are stated instead, which is the only thing that makes two numbers for one test readable rather than a contradiction — and which is the subject of vitrine VII.