Repository path: workshop/experiments/E-20260812i-footing-channel/design.md · rendered 2026-09-09

E-20260812i-footing-channel — the channel the mark rides on, and the marks nobody's source asked for

FROZEN BEFORE THE COMPARATOR WAS OPENED. ARM-footing step 1, study limb. The translation limb — T-unter-prishibeyev-R06-v1 and T-unter-prishibeyev-R04-v1, Chekhov's «Унтер Пришибеев» whole — was frozen and committed at 9d0e01e and b15aa66 before this page was written, and this page was written and committed before any English rendering of the story was read.

1. Question

framework/v0.2 §2's statement S1 counts in one direction — the source marks; does the English carry? — and reports an absence across five pairs. RS-20260812h, a T1 census on Tagore's «ডাকঘর» the day before this design, counted in both directions and found two things S1 cannot see:

1. The channel predicts the carriage. Where the Bengali's mark rode on a word (ছোঁড়া boy, বাঁদর monkey), both English hands carried a device at 5 of 5; where it rode on morphology alone (তোর for তোমার, the particle রে), both lost it at 3 of 3.
2. Both hands marked a site the source leaves neutral — the same speech, a century apart.

That census was exploratory: its labels were written after the comparator was read. This design asks the same two questions on an unrelated pair with the criteria, the site list and the predictions all fixed in advance.

Q1 (channel). Does an English hand carry a footing marker at source places whose mark is lexical and lose it at places whose mark is morphological only?

Q2 (insertion). At how many places does an English hand mark footing where the source does not? This project has never counted it.

2. Materials

Source. Anton Chekhov, «Унтер Пришибеев» (1885), Russian Wikisource ПСС text, SHA-256 6f51d7215ae4e49d1cf3ae67dc8e49a3b5c744a1f192f7d7f672094159b376f3, 23 paragraphs, 1,225 words. A courtroom scene chosen because its subject is a footing collision: a magistrate addressing a retired NCO in neutral вы, the NCO answering in ваше высокородие, the -с clitic and honorific plural verbs, and the same NCO's remembered ты and epithets to a village constable.

Hands. Four, in two classes.

The lead translated knowing this session's general subject. Both lead hands are therefore reported descriptively only. P1 and P3 are told nothing but "translate this story into English"; the prompt is stored verbatim under materials/.

3. The source-side census — FIXED HERE, BEFORE THE COMPARATOR

3.1 What counts as a marked place

A place is a clause or sentence in which the speaker marks their footing toward the person they are addressing, by departing from the neutral baseline. The baseline in this text is вы without supplement between the magistrate and Prishibeyev, and вы/plural between any speaker and a group.

UP markers: ваше высокородие / вашескородие; господин + title addressed to the hearer; the -с clitic (словоерс); изволите + infinitive; a plural verb for a single addressee or referent of respect; Слушаю as a soldier's acknowledgement.

DOWN markers: ты and its forms to a non-intimate adult; vocative epithets (дурак этакой, такой-сякой, брат, кавалер).

Not markers, registered so the count cannot drift: a bare office-title used as a vocative (Урядник!) — neutral; generic ты in a proverb (на твоей же душе грех) — not addressed to anyone; second-person plural imperatives to a crowd (расходись, скалите) — neutral.

One counting rule, registered because otherwise ты swamps everything: a ты-footing toward one addressee inside one continuous stretch of speech is one place, at its first occurrence. A later place in the same stretch enters the census only if it carries an additional lexical marker.

3.2 Channel

• LEX — the place carries at least one marker that is a word (an honorific noun, a title, an epithet, изволить).
• MOR — every marker at the place is bound or inflectional: the -с clitic, ты-forms, a plural verb used honorifically.
• BOTH — the place carries markers of each kind. Excluded from Q1's contrast and reported separately.

3.3 The site list — 27 marked places

Built from the Russian alone, before the comparator was opened. ¶ is the source paragraph.

14 LEX · 13 MOR · 0 BOTH. Both cells clear F1's floor of 5.

3.4 A secondary class, reported and not pooled: reference honorifics

Four places mark the speaker's stance toward a person who is not the addressee. They are listed because one of them is the single most extreme grammatical marker in the text and has no English slot whatever.

3.5 The unmarked address units — where an insertion can be found

Sixteen stretches in which one speaker addresses one hearer and the Russian carries no marker: ¶1, ¶4, ¶7a (to the witnesses), ¶7b (to Prishibeyev), ¶8 (Zhigin → the crowd), ¶8 (Zhigin → Prishibeyev, twice), ¶9, ¶10 (the elder → Prishibeyev), ¶11, ¶12, ¶13, ¶15, ¶16, ¶19, ¶21, ¶23.

4. Procedure

1. Freeze this page and commit it. Done before step 2.
2. Dispatch the two panel hands (P1, P3), one call each, prompt stored verbatim, temperature 0, max_tokens 5,000.
3. Only then open GAR and align all four hands to the source paragraphs.
4. Census A. For each of the 27 marked places × each hand: does the English carry a marker of the same footing at the corresponding place? Labels, fixed here and taken from RS-20260812h §3: LEX the English marker corresponds to a lexical marker the source has there; REL the English adds a marker where the source's mark is morphological only; LOST no English marker; REV the English marker points the other way. Carriage = the place is not LOST.
5. Census B. For each hand, list every English footing marker in the 16 unmarked units. Each is an insertion.
6. Independence. tools/dependence_check.py for every non-GAR hand against GAR.
7. Verification recomputes every reported number from the stored codings by an independent path.

5. Predictions, registered

• P1 (channel). Pooled over the hypothesis-blind hands, carriage at LEX places − carriage at MOR places ≥ +0.50. RS-20260812h gives +1.00 on Bengali in both its hands.
• P2 (insertion). Each hypothesis-blind hand marks footing at ≥ 1 of the 16 unmarked units; and pooled over the blind hands, insertions are ≥ 10% of all places that hand marks.
• P3 (convergence, descriptive, no bar). Report which unmarked units more than one blind hand marks. RS-20260812h found two hands inserting in the same speech a century apart; whether that reproduces is worth seeing and this design is not powered to test it.

6. Failure criteria

• F1 — fewer than 5 LEX or fewer than 5 MOR places. Checked at freeze: 14 and 13. Not fired.
• F2 — a non-GAR hand shares a 12-gram with GAR, or has a longest common run ≥ 15 tokens against it. That hand is not independent of GAR, its agreement with GAR is not counted as replication, and the fact is reported on the hand's own page.
• F3 — a panel hand truncates, refuses, or returns under 1,200 English words. It is dropped and named, and the primary rests on the hands that remain.
• F4 — GAR turns out to be an abridgement or a different story. Every primary is withheld.
• F5 — the two lead hands, already outside the primary, are additionally dropped from the descriptive report if F2's condition holds for them.

7. What this design cannot do, stated before it is run

1. The coding is the lead's, unanchored, and one of the four hands is the lead's own. Every number is internal-judgment-only. A blind independent coding would be a stronger instrument; putting it to panel seats would be a study of rater reproducibility, which the subject rule makes method work, and it is not run.
2. One story, one pair. With RS-20260812h that is two pairs, which is a replication and not a generalisation.
3. P1 and P3 are 2026 language models, plausibly carrying GAR in their weights; F2 is what stands between that and a false replication.
4. The insertion census counts markers, not their rightness. Whether an inserted boy or my man is a good idea is a judgment about quality, Tier D is not passed, and no such claim is made.

8. Pre-flight cost estimate

Built from max_tokens, per note (abc), and inflated 4× on the routing caution in config/models.md §Pricing caution. Today's UTC headroom before this session: $1.298861274.