Repository path: workshop/experiments/E-20260801b-census-author/mode-rule.md · rendered 2026-09-09

The elicitation-mode instrument — frozen 2026-08-01 (S078) BEFORE any span was selected or any word translated

Why this file exists and why it is committed first. workshop/experiments/README.md's standing critic disposition 2: the instrument is frozen before the design that uses it, not authored alongside it. This session's translation limb exists to measure one thing — whether the way an option census is elicited moves the statistic the study limb compares across authors — and an instrument written after the sites were chosen could be shaped to the answer.

What was and was not in view when this file was committed. The source work had been chosen (Adam Szymański, «Stolarz Kowalski», 1911, Polish Wikisource) and its Polish text fetched to disk. Roughly the first ninety words of its opening paragraph and the first 160 characters of each of its sixteen rendered paragraph blocks had been printed to the transcript in the course of establishing that the text was complete and how long it was. No span had been selected, no site had been identified, and no word had been translated. No English rendering of any part of the work has been read at any point, and none will be before both translations and both censuses are frozen.

1. The two elicitation modes

census-rule.md (S073, frozen) §1 defines a census site as a locus "at which the translator, at the moment of choosing, weighed more than one English rendering… recorded then, in the same pass as the translation, and never reconstructed afterwards from the finished text."

That definition describes one of the two modes below. It is the mode every option census in this repository was written in, and it is not the mode the three non-lead censuses of RS-20260731e-option-census were written in. That page's §7 named the difference and did not measure it:

"three models enumerating on demand are not a translator writing a log at the moment of choosing, and the two elicitation modes differ in more ways than the manipulation."

Mode MOMENT. Exactly census-rule.md §1. The translator records each site as it is decided, in the same pass as the translation, writing down every rendering it was prepared to write at the instant it chose. The census exists before the finished sentence does.

Mode ONDEMAND. The translator translates the passage with no census of any kind written, and no locus marked. The translation is frozen. Only then, reading the frozen source and the frozen English side by side, does the translator enumerate — at each site the selection rule below picks — every English rendering it judges to have been live at that locus. The census is a reconstruction from the finished text, which is what a subject who did not do the translating can produce and is what the three seats in E-20260731e produced.

The two modes differ in exactly the way this project's lead censuses differ from its non-lead ones. Nothing else about them differs: same translator, same regime (R04), same work, same session, same site fields, same class rule.

2. Site fields and the class rule

Both unchanged from census-rule.md §1 and §2, quoted here so this file is self-contained and so no later session has to reconcile two versions:

• CONSTRAINED — a proper name, a numeral, or a common noun/verb/adjective with a single standard English equivalent in a general Polish–English dictionary and no cultural qualifier.
• OPEN — a culture-bound item, a dialect or sociolect form, or a figure whose rendering requires choosing among the eight-wide handling set of CL-20260726-handling-set.
• MIXED — neither cleanly.

options is a set, not a ranking. chosen is always a member. Nothing recorded under this instrument is a quality judgment about any rendering.

3. Site selection, and it is the same rule in both arms

In MOMENT: every locus at which more than one rendering was weighed, in source order; the first 20 are the arm's items.

In ONDEMAND: reading source and frozen English side by side in source order, every locus at which the translator judges more than one rendering to have been live; the first 20 are the arm's items.

If an arm's passage yields fewer than 20 sites the passage is extended forward in the source until it does, and the extension is declared. If it yields more, the surplus is recorded and not used.

The rule is identical in both arms and the mode is the only thing that differs. What the rule cannot make identical is who is applying it and when, and that is the manipulation, not a defect.

4. The passage split, fixed here before the span is chosen

One contiguous span of the work, split into two adjacent passages A and B of approximately equal length, at a sentence boundary named in design.md once the span is fixed.

Passage A is arm ONDEMAND. Passage B is arm MOMENT. The order of work is fixed and is part of the instrument:

1. Translate A under R04, writing nothing down but the translation. Freeze it in its own commit.
2. Write A's ONDEMAND census from the frozen A. Freeze it in its own commit.
3. Translate B under R04, writing the MOMENT census as it goes. Freeze both in one commit.

Step 2 precedes step 3 deliberately, so that no live census exists in the session when the reconstruction is made. The reverse order would let the MOMENT habit shape the ONDEMAND list, which is the contamination this arm cannot afford.

5. What this instrument does not do, declared before the run

• It does not make ONDEMAND an uncontaminated reconstruction. The translator knew, while translating A, that a census would be asked for afterwards. It wrote nothing down, and options weighed and then forgotten are lost by construction — but it is not a translator who had never heard of a census. This biases ONDEMAND toward MOMENT, i.e. against finding a mode effect, and that is the conservative direction for the study limb's purposes.
• A and B are different text. Any A/B difference is mode confounded with passage. The design reports per-site n and both arms' class distributions so a reader can see how far apart the passages are, and the arm's claim is bounded by that.
• B is translated after A, so B has A's context and A does not have B's. Declared, not repaired: repairing it would require translating one of them cold, which changes the regime.
• It says nothing about which mode is right. The project's published figures are MOMENT figures and its only non-lead figures are ONDEMAND figures. The question is how far apart they are, not which to prefer.