Repository path: workshop/experiments/E-20260729b-graded-drift/design/sites-alfred.md · rendered 2026-09-09

Set A — the reflex census of Alfred's Preface, frozen

Written from materials/alfred-preface-oe.txt and dictionary glosses alone. No English rendering of this text was read before this file was committed. Machine-readable form: sites-alfred.json, 80 sites — 71 CLEAR, 9 ARGUABLE.

The inclusion rule, stated before the list was made

A word of the passage is a site when all three hold:

• (a) Reflex. The Old English word has a modern English form that is its direct historical continuation, not a later borrowing, a re-formation, or a homograph from another etymon.
• (b) Free-standing. That modern form is a word of modern English in its own right, not only a bound morpheme or an element surviving inside a fixed compound.
• (c) Writable. A translator rendering this passage could plausibly write that modern form at that point — whether or not doing so would be a good choice.

Excluded by (a): notu (the modern note is from Latin nota via French, not from notu); fyrst "period of time" (the modern first is from fyrest, the superlative of fore). Both were tempting and both fail the etymon test; they are named here because a rule that never excludes anything is not a rule.

Excluded by (b): hād (survives only as the suffix -hood; the free noun hood is OE hōd, a different word).

Excluded by (a) or by having no modern reflex at all, though each is a word a reader would notice: wealhstōd "interpreter", æstel, andgit "understanding", ǣ "law", onstāl, lārēow "teacher", ēðel "homeland", sidu "custom", māðum "treasure", fultum "help", wilnung "desire", ðīowotdōm "service", geðīode "language". geðīode is the most consequential exclusion in this list — the text's own word for language has no modern descendant, which is a fact about the drift this experiment is measuring and not a defect of the census.

Where the same lemma occurs more than once in the passage, one site is entered, at the first occurrence, and the passage form recorded is the form at that occurrence.

Two things this census is not

1. It is not the passage. RS-20260729-drift-window-verify §7 measured the lead's Beowulf census against an independent one at Jaccard 0.4375 — a rule written down explicitly, applied by two readers, landing 44% together. No independent census is run here. Every Set A figure is a statement about this frozen site set, and the result page says so wherever it reports one.
2. It is not used for the crux test. G3 — the dip test and the mixture comparison — is computed on Set B, the 52 Beowulf lemmas censused at S053, whose site list was frozen before this question was asked. This census was written after the design that specifies the dip test, by the person who specified it, so it is disqualified from that statistic by construction. Set A carries G5 (three-rater reliability on a second text, which does not pass through the lead) and the exploratory G6/G7.

Cross-text repeats

Six lemmas appear in both sets, which was not designed for and is recorded because it is useful: folc → folk, sibb → sib, wēnan → ween, rīce → rich, cweðan → quoth, mōd → mood, hierde/hyrde → herd, and rǣdan/ārǣdan → read. The last two are the sharpest: hyrde → herd is the lemma on which S053's two readers gave nearly the same sentence for opposite classes, and rǣdan → read is the lemma at which the anchor's 0 of 12 was corrected to 1 of 12. Both now have a graded score from three raters, given in a different context, with no Beowulf named.

Set A take-scoring rule, frozen (design amendment A6)

Take/refuse on Set A is decided mechanically, not by a rater:

A rendering TAKES the reflex at site X if the modern reflex form of X, or an inflected or compounded form of it, occurs in that rendering within the sentence that renders the sentence containing X. Matching is case-insensitive on whitespace-normalised text, against the frozen reflex string and a stated inflection list per site. Otherwise the rendering REFUSES.

The lead's own row is not blind — the lead wrote this census and scored it before translating — and is excluded from every statistic that pools translators (F4).