Repository path: workshop/experiments/E-20260802e-displaced-marking/materials/census.md · rendered 2026-09-09

The census — every site in the project's frozen logs where a grammatical marking is declared lost

Frozen before any rendering was written and before the design's predictions were fixed. Extraction is mechanical (analysis/extract_census.py, committed with this file, reproducible from the repository alone). Classification is the lead's and is what §5 of the design puts to independent readers.

Extraction — mechanical, three filters in series

1. Claim filter. Over all workshop/translations/*/*/translation.md, a frozen regex for unavailability assertions ("English has no …", "no target counterpart", "cannot be carried", "lost outright", "no compensation", …). 126 claims.
2. Grammatical-category filter. The claim's 660-character window names a grammatical category from a frozen lexicon (pronoun · person · number · case · mood · aspect · tense · gender · honorific · diminutive · morphology · particle · article · word order · voice · clitic · impersonal · reflexive · auxiliary · copula …). 61 claims survive. The 65 excluded are lexical gaps — no single English word for удаль, for roba, for длань — where "use another category" is the ordinary move and carries no news.
3. Loss filter. The window states that nothing was carried: no compensation · not compensated · lost outright · the loss is total · nothing to repair with · no available alternative · declared loss · flattened with no …. 13 windows survive, collapsing to 8 distinct sites (four windows are errata quoting two of the sites back).

Recall is not established, and one site proves it. T-kiseru-R04-v1 D15 — "English politeness has no way to be this deferential" — is a Class A site by every criterion and filter 3 misses it, because the intensifier carries markdown asterisks the regex does not allow. It was found by hand and is included, and it is recorded here rather than quietly folded in: the mechanical census is a lower bound on its own population, by at least one. Any figure below is stated over the sites named, never over "all sites in the project".

Classification rule (the lead's; frozen; §5 tests whether it is followable)

Given a window that passes filters 1 and 2:

• Class A — the log states the source's marking was not carried at all: no device in any category was substituted for it.
• Class B — the log states the source's category is absent and records a compensation taken in a different grammatical category.
• Neither — the gap is lexical, or the window is about the project's own instruments rather than about a translation decision.

Class A — declared total losses (the test population)

Eight sites. Two have been tested by an independent comparator and both were wrong. Six have never been tested, and this experiment tests all six — a census of the remaining population, not a sample of it.

Class B — the same gap, a compensation found in another category

Recorded because it is the standing evidence that the move exists, and because it was made without any rule telling the translator to make it.

Six Class B sites against eight Class A. The project's own logs therefore already contain the move R1 prescribes, at roughly the same frequency as the declaration that it is impossible — and nothing in the logs connects the two.