Repository path: workshop/experiments/E-20260730-grain-clause/variants.md · rendered 2026-09-09

C15′ and C15″ — two one-clause rewrites of C15, frozen before the census and before any translation

What this is. RS-20260729d-decision-grain §2 localised eight of the nine reader disagreements under C15 in a single clause: test 3's "an exact equivalent from a practice the two cultures share". ARM-decision-grain step 2 prescribes replacing exact with a condition that needs no shared threshold, and re-running the applicability pass. This file holds the replacement and its control.

Every one of the three rules is byte-identical to C15 outside test 3's second clause. Test 1, test 2, test 3's first clause (established borrowing), test 4 and the standing constraint are unchanged, and so is the definition sentence. Nothing else moves.

The clause being changed, and how much warrant it actually carries

candidates.md traces test 3 to four sites in A-garnett-vanka:

The clause that carries eight of nine disagreements is anchored at exactly one site. That is not an argument for changing it — the point of the arm is to test the change — but it is the honest description of what is being rewritten, and it is why the retrodiction check in design.md §5 can discriminate so weakly.

C15′ — the operationalised replacement

Its wording is the previous session's, not this one's. ARM-decision-grain step 2, written at S056 before this session saw any of S056's site-level outputs, proposes: "the English word appears in a general dictionary as a translation of the source word, without a cultural qualifier". That is the condition below, spelled out. Recorded because the session that runs the test also read the nine disagreeing sites, and the protection against tailoring the fix to them is that the fix was specified before the reading.

C15′. At a culture-bound item, four ordered tests decide the handling; apply the outcome to the whole class senses: cultural-mediation, style-correspondence, consistency evidenced on: JA→EN, RU→EN full statement: A culture-bound item is a word or phrase naming something the source culture has and the target culture does not have under the same description. At such an item, do not choose among the eight handlings by feel. Apply these four tests in order and take the first that fits. (1) Is the item's sense already stated at the site by the source itself — an apposition, a relative clause, a repeated morpheme that is itself translatable? Then retain the item (transliterate a name, keep the term) and translate the source's own explanatory material. Add no gloss of your own and do not substitute; the sense arrives at no cost to you. (2) Is the item a quantity, and is its value rather than its cultural identity what the sentence uses? Then convert the measure into the target's system and do nothing else. (3) Does the target already hold an established borrowing of the item, or would a general bilingual dictionary give an ordinary English word as a translation of the item with no qualifier marking it as foreign, historical or culture-specific — no "a kind of", no "a Russian —", no "in Japan, —", and no bracketed explanation standing in for the word itself? Then use it; no fork arises. (4) Otherwise ask whether the item is load-bearing for the passage's argument or plot, or whether it is furniture. Load-bearing: retain the item and scaffold it — manufacture its sense from the surrounding context in your own words — and do not substitute a target-culture near-equivalent, however ready one is. Furniture: retain it copy-opaque and spend no gloss on it, or omit it where the sentence does not use it. Standing constraint: whichever handling a test returns for one member of a class of items, apply that handling to every member of that class in the same text.

What it is meant to remove. Exact asks two readers to share a threshold on how close a counterpart must be. The dictionary condition asks instead whether a lexicographer printed an unqualified equivalent — a fact about a reference work, not a degree of similarity.

What it is not. It is not more warranted than the clause it replaces; it is warranted by the same single site (колодка → "lasts" is exactly the case a dictionary gives unqualified). Nothing here claims the rewrite is better, only that it removes a shared threshold.

C15″ — the control, and it is a sham at clause grain

C15″ is a sham. It is written this session, and it exists because a κ rise under C15′ would be uninterpretable without it.

RS-20260729d §6 item 3 already established that the fully mechanical C16 out-agrees warranted C15 at 0.933 against 0.452. So any rewrite that makes test 3 mechanically checkable should raise agreement, whether or not it preserves the clause's warrant. A rise under C15′ alone would therefore confirm nothing about the operationalisation and everything about mechanisation. C15″ holds the mechanical burden fixed — the same dictionary lookup — and swaps the criterion for one nothing in this project's evidence connects to handling.

C15″. At a culture-bound item, four ordered tests decide the handling; apply the outcome to the whole class senses: cultural-mediation, style-correspondence, consistency evidenced on: JA→EN, RU→EN full statement: A culture-bound item is a word or phrase naming something the source culture has and the target culture does not have under the same description. At such an item, do not choose among the eight handlings by feel. Apply these four tests in order and take the first that fits. (1) Is the item's sense already stated at the site by the source itself — an apposition, a relative clause, a repeated morpheme that is itself translatable? Then retain the item (transliterate a name, keep the term) and translate the source's own explanatory material. Add no gloss of your own and do not substitute; the sense arrives at no cost to you. (2) Is the item a quantity, and is its value rather than its cultural identity what the sentence uses? Then convert the measure into the target's system and do nothing else. (3) Does the target already hold an established borrowing of the item, or would a general bilingual dictionary give an English word as a translation of the item that is spelled with fewer letters than a romanisation of the source item itself — counting letters only, and not counting spaces, hyphens or accents? Then use it; a sentence should not be slowed by the longer of two available forms. (4) Otherwise ask whether the item is load-bearing for the passage's argument or plot, or whether it is furniture. Load-bearing: retain the item and scaffold it — manufacture its sense from the surrounding context in your own words — and do not substitute a target-culture near-equivalent, however ready one is. Furniture: retain it copy-opaque and spend no gloss on it, or omit it where the sentence does not use it. Standing constraint: whichever handling a test returns for one member of a class of items, apply that handling to every member of that class in the same text.

Matched to C15′ on: the lookup (both require a dictionary equivalent to be brought to mind), the crispness of the criterion (both are facts, not degrees), the position in the rule, the presence of a reason clause, and — after the amendment below — length.

Amendment, 2026-07-30, before the design was written and before any call. As first frozen at 9830845, C15″'s test 3 ran to 58 words against C15′'s 68, because C15′ spells out four things a qualifier can look like and C15″ spelled out nothing. Length is a plausible driver of reproducibility on its own, and an unmatched 10 words would have been an uncontrolled difference between treatment and control. A counting instruction of the same kind (a spelled-out procedure) and no additional warrant was added: "counting letters only, and not counting spaces, hyphens or accents". Recorded here rather than silently rewritten; build_prompts.py prints both clause lengths on every run.

Matched to C15′ on nothing else, deliberately. Unmatched, deliberately: the criterion's connection to any evidence. Nothing in this project's anchors, claims, results or theory pages connects handling to the relative orthographic length of an item and its equivalent.

It must never enter framework/traceability-inventory.md, and any session that finds it there should strike it. The same standing rule candidates.md sets for C16 and C17.

What is deliberately NOT changed

• The reader instruction block is byte-identical to S056's. Not one word. The C15 condition in this experiment is a byte-identical repeat of S056's prescribe-C15.prompt.md, so that the difference between S056's κ and this session's κ on that condition is stochasticity and nothing else. Every temptation to improve the instrument was declined for this reason, including one worth having: the readers are not asked to name the dictionary equivalent they had in mind, which would have made the disagreements auditable, because adding a field would have broken the repeat.
• The 23 Gogol sites and their glosses, which are E-20260729d's frozen census.json used unmodified.
• The eight handling labels and their definitions.