Repository path: workshop/experiments/E-20260804-displaced-marking-fr/materials/census.md · rendered 2026-09-09

The census — every decision row in the frozen «La Nuit» log, classified

Frozen before any FORCED, DECOY or POSITIVE rendering was written and before the predictions were fixed. The translation and its log were committed at caf6b2e, complete, before this page existed; nothing below could reach back into them.

Extraction — the whole log, not a filter over it

census.py parses the frozen log into its numbered rows and asserts they run D1…D31 without a gap. The population is therefore all 31 rows, and recall over the log is total by construction — which is the one thing E-20260802e's census could not claim about itself. What it is not is recall over the translation: decisions made without noticing them do not enter a log at all (R04 §Known limitations), and no method here reaches those.

The same script re-runs E-20260802e's frozen three-filter extractor over this log, and the gap is the finding about the instrument. On la-nuit the frozen filters return 6 windows at filter 1, 6 at filter 2, and 2 at filter 3. This page's hand audit of all 31 rows finds 12 Class A rows. So the mechanical census recovers 2 of 12, and its own docstring's warning — "a lower bound on their own population" — understates by a factor of six on a log it was not tuned against. That is recorded here and in the result's limits; it is method work and it is not this session's question.

Classification rule — E-20260802e's, quoted verbatim, not amended

• 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.

Classification is the lead's, as it was there.

The 31 rows

Class A 12 · Class B 3 · Neither 16 · total 31.

The A:B ratio runs the other way from RS-20260802e's. That census, over the project's whole prior record, found 8 : 6–8. This one log gives 12 : 3. One translator, one text, one session, and two censuses built by different methods — the numbers are not comparable and are set side by side only so that nobody later reads 12:3 as a replication of 8:6–8. It is not one.

The eight distinct source devices, and the test population

The twelve Class A rows are eight distinct source devices: three rows are the pronoun on in three functions (D12, D19, D27) and three are grammatical gender (D22, D23, D30). Rows sharing a device are not independent sites, so the test population is one site per device, eight sites, and the within-device picks are named here with their reasons before any rendering was written:

Not one of the eight can be marked with an address noun, and that is a property of the material rather than a rule imposed on it: «La Nuit» is a solitary monologue with no addressee anywhere in its 1,873 words — the only second person in the story is a beggar-woman's one line and the narrator's one question to a ragpicker, neither of which is a site here. ARM-framework-v01 step 3 inherited the constraint that a stress test must not exercise the address noun again (framework/v0.1 §4: five of R1's six licensed device categories have never been isolated). This text cannot exercise it.