Repository path: workshop/experiments/E-20260810d-carriage-naming/design.md · rendered 2026-09-09

E-20260810d — can a reader holding the original say what a foreignizing departure is carrying?

Frozen before any generation or scoring call. The R06 translation this run's materials are built out of was frozen first, at commit e077fef, before this file existed.

1. The question, and why it is about translation and not about the instrument

Foreignizing translation is defended, by Schleiermacher, by Berman and by Venuti, on the ground that it makes the foreign text's difference visible to a reader who could not otherwise see it. The claim is about a reader: at the place where the English goes strange, the reader is supposed to be able to see the original showing through.

Nothing in this project has ever asked a reader to point. Three runs have asked readers to rate — to score perceived-source-carriage, the sense created at S093 to record when a marked departure reads as a deliberate attempt to carry over a feature of the source — and all three found the same thing:

The third is the sharpest, because CLUNKY was a strict word-multiset permutation of a clean translation — proved so by the verifier, not one word added or removed — and it out-scored five of six foreignizing translations and the lead's own. RS-20260809h's limits section concluded that the sense "does not separate foreignizing translation from scrambled word order" and barred it as an uptake measure without a matched oddity control.

A prohibition is not an explanation, and the prohibition is all ARM-rule-execution step 2 was scoped to write down. Three rating runs agreeing tell us the rating does not separate carriage from non-carriage. They do not tell us whether the separation is unavailable to the reader or only unavailable to the rating. Those are different findings about literary translation:

• INVISIBLE — a reader holding the original cannot tell a departure that carries the German from a departure that carries nothing. Then the visibility claim that motivates the whole foreignizing programme fails at the level of the individual departure, and framework/v0.2's recommendations that reach for source-orientation are recommending something no reader can locate.
• UNRATED — the reader can tell, and the seven-point scale cannot. Then the three results above are about an instrument, the visibility claim is untouched, and what the framework needs is a different question put to the reader, not a different recommendation.

The unit's sentence (subject rule, continue-prompt.md §4.5): this unit asks whether a reader with the original in front of them can say what a foreignizing English departure is carrying, and whether they say it just as readily when it is carrying nothing. That is a claim about reading translations. The write-up into wiki/goodness-senses.md that ARM-rule-execution §Done when clause 2 requires is downstream of the answer, and is the reason the step was extended from writing to measurement-and-writing (arm log, S151).

What this run changes about the question put to the reader. Every prior run asked for a rating on a scale. This one asks the reader to point at the German — to quote the exact words whose feature the English departure carries, or to answer NONE. A rating is a permission; a quotation is a commitment, and it can be wrong in a way a rating cannot.

What the run can and cannot decide — narrowed after the pre-run critic (amendment A11). The scored output is whether the reader attributes the departure to the source and points at German, not whether the feature they name is the right one. So the estimand is attribution: does a reader holding the original attribute a marked English departure to the German more readily when the departure really is carrying something than when it is carrying nothing? A null on that question is enough to establish INVISIBLE-at-the-site, because a reader who attributes equally to licensed and unlicensed markedness is not seeing the source through the English. A positive is not enough to establish that the reader saw the right feature; that would need the FEATURE strings coded, and this design codes nothing by hand. The distinction is kept in every sentence of §5 and in the result page.

2. Materials

Source. Arnim, «Die Kronenwächter», Erster Band, Zweites Buch, Erste Geschichte: Die wunderbare Heilung (1817) — the chapter's opening paragraph entire, 604 German words in ten segments. Copy-text Projekt Gutenberg's Reclam digitisation, collated against the 1857 archive.org Fraktur scan (collate.py): 10 of 10 sentences present, character ratio min 0.8896 / median 0.9358 / max 0.9622. Two copy-text errors were corrected against the witness, one of them a sentence boundary that makes the span ten segments rather than nine (../../translations/kronenwaechter-heilung/R06-v1/translation.md).

Why this source. It is the only German prose this project has established as having no reachable published English translation at all (S141 §2, S146 §2, re-searched this session), which removes recall entirely: no seat can be scoring a remembered rendering. It holds the language pair, the author, the work and the prose type constant against S141 and S146, and the spans do not overlap by a word. Contamination is declared none on the artifact with its basis stated — the absence of a comparator, not a measured overlap, since dependence_check.py takes a pair and there is no second member.

Why this paragraph. It was chosen, before any English was written, for the density of German constructions a foreignizing translation would have something to carry: three verb-first conditionals, a fronted dative object with its subject held to the end, correlative so … wie, participles held behind a long middle field, a prepositional nominalization, and an impersonal passive with no subject at all.

The baseline text, P. T-kronenwaechter-heilung-R06-v1 — the lead's plain rendering of the whole paragraph under R06, one pass, source only, no rule set, frozen at e077fef with its translator's log before this design existed. Cost $0; lead translation is never ledgered.

The ten sites and the four conditions

One site per segment, ten sites. At each site a span of the frozen P translation is marked with ⟦ ⟧, and four English versions of the segment are built which are byte-identical outside the marked span:

L and F are both syntactic, deliberately. The obvious way to build unlicensed strangeness is to reach for archaism — hath, thereof, soever — and that would have confounded licence with kind of markedness: the reader could separate the arms by register alone without ever consulting the German. Every F here is instead a construction English can produce and the German at that point does not have: absolute participle, postposed adjective, comparative inversion, predicate fronting, left-dislocation with a resumptive pronoun, cleft, emphatic do-support, agentive of, existential there with heavy-NP shift, postposed quantifier, polysyndeton, periphrastic comparative. The residual mismatch is named as a limit in §6 and not claimed away: L and F are matched on being syntactic, not on being the same construction.

The pre-run critic found three sites where this had not been achieved and all three were repaired before dispatch (§11): site 1's F was an absolute participial clause, which is a near-calque of the German prepositional nominalization it was supposed to be a control for; site 6's L was a lexical choice (dwelt, confidingly) rather than a syntactic carry; and site 2's L was a bare preposition swap nameable from the lexicon without reading the German's syntax at all.

The registered site table is materials/sites.json, written before dispatch, and carries for each site: the P span verbatim, the L, F and S spans, the exact German locus the L span claims to carry, and a one-clause statement of the feature. It is reproduced in full in §7 of this file so that the freeze is legible without reading the JSON.

S is generated, not written — build_items.py tokenises the P span, permutes with a fixed seed, and verify.py asserts the case-folded word multiset is unchanged. This is the site-level analogue of RS-20260809h's CLUNKY, and the same property is proved the same way.

3. The jury and the two passes

Three seats, the same three as E-20260808c, E-20260809b and E-20260809h, so that this run is read beside them: J1 openai/gpt-5.6-terra, J2 google/gemini-3.6-flash, J3 deepseek/deepseek-v4-pro. All non-Anthropic (charter §5; the lead may not judge its own translation, and this whole run is the lead's translation).

Pass A — the naming task. 10 sites × 4 conditions × 3 seats = 120 calls, one item per call, no conversation history, temperature 0. Each call shows one German segment and one English version with one span marked, and asks for QUOTE:/FEATURE: or NONE. The prompt string is frozen in materials/prompts.json and is reproduced in §8.

Pass B — the markedness check. 10 sites × 3 seats = 30 calls. Each call shows the four English spans of one site and no German at all, labelled A–D in a per-seat registered order, and asks for a 0–6 rating of how far each departs from ordinary present-day English prose. The German is withheld because this is a target-side markedness judgment and showing the source would be exactly the contamination the check exists to rule out.

Blinding. Condition labels never appear in any prompt. Pass A items are shuffled per seat with a registered seed, under the constraint that two items of the same site are at least six positions apart (materials/orders.json, written before dispatch). Pass B's A–D order is shuffled per site per seat. What this does not achieve is stated as a limit: a seat sees all four versions of a site across its 40-item stream, and although the calls are independent and stateless, the four versions of one site are near-identical strings and a stateful provider could in principle relate them.

4. Scoring — mechanical, and no lead judgment enters any number

A Pass A body is parsed as:

• NONE iff its first non-empty line, upper-cased and stripped of punctuation, is exactly NONE;
• NAME iff it contains a line beginning QUOTE:;
• anything else is MALFORMED and goes to F5 below.

NAME(c) = the count of NAME cells in condition c, over 30 cells (10 sites × 3 seats).

HIT is defined only at L and is mechanical: normalise the quoted German and the registered locus (case-fold, fold umlauts, ß→ss, drop non-letters, split to word tokens), and score a HIT iff the quote's token sequence is contained in the locus's, or the locus's is contained in the quote's, or they share at least half of the locus's content tokens — and the quote is no longer than 60% of the segment's word tokens, so that a whole-segment dump is not scored as a hit. Both thresholds are registered here and are not tuned after seeing bodies. HIT is a secondary measure (amendment A6): the primaries are NAME contrasts, and HIT bears only on C1.

FEATURE: text is collected and is not scored. Coding whether a stated feature is the right feature would be lead judgment on panel output, and this design admits none: the FEATURE strings are reproduced verbatim in the result page as illustration and no statistic is computed from them.

The positional objection, and why it is the hypothesis rather than a confound. At every site the same German words stand at the same place in all four conditions, so a seat that simply quotes the German opposite the marked span will score HITs at L without consulting the feature at all. That same strategy would also produce quotations at F, at S and at P, where the true answer is NONE — which is exactly what NAME(F), NAME(S) and NAME(P) count. The design does not need to exclude the strategy; it needs to measure how far it is used, and that is what it measures.

5. Gates, primaries, and what each outcome licenses

Everything below is registered before dispatch. Bars are not moved after bodies are seen.

Gates — restructured after the pre-run critic (amendments A5, A6, A14)

The critic's MAJOR 5 was accepted: the original gate stack withheld the primaries on failures that are not failures of the primaries. P1 and P2 are contrasts between NAME rates. They do not depend on the HIT matcher and they do not depend on NAME(P). Only P3, which is defined against P, depends on P. The gates below withhold exactly what they bear on and nothing else.

Primaries

• P1 — does attribution track licence? NAME(L) − NAME(F), computed per site over its three seats and tested across the ten sites by an exact two-sided sign-flip permutation test (2¹⁰ = 1,024 assignments; minimum attainable two-sided P = 0.00195).
• ≥ +0.40 with P ≤ 0.05 ⇒ ATTRIBUTION TRACKS LICENCE. Readers holding the source attribute carried markedness to it more often than free markedness; the three rating results are then about the rating, and the sense needs a better question rather than retirement.
• anything else ⇒ NOT SHOWN TO TRACK LICENCE, with the point estimate and 90% interval printed.
• No equivalence claim is registered and none may be made — amendment A4, accepting the critic's BLOCKING 4. Ten sites of three binary cells cannot certify a ±0.20 band, and the design originally registered one. A null here is failure to demonstrate, and the result page must say so in those words.
• P2 — does mangling draw attributions? NAME(S). ≥ 0.50 ⇒ a site-span word-scramble draws source attributions from a majority of cells. Amendment A8, accepting the critic's MAJOR 8: this is consistent with RS-20260809h's CLUNKY result and does not establish its mechanism — that run scrambled a whole segment, this one scrambles a span of five to fifteen words, and the two manipulations are not the same size.
• P3 — does strangeness as such raise attribution? NAME(F) − NAME(P), same paired test. A positive P3 with a null P1 is the INVISIBLE reading in its strongest form: strangeness buys attribution and licence buys nothing on top of it.

Registered secondary (amendment A9, critic MAJOR 9)

Three of the ten L spans (sites 4, 5, 7) restore a verb-first German construction. P1 is therefore recomputed on those three sites and on the other seven separately, and a positive P1 carried only by the verb-first subgroup is reported as "verb-first markedness is nameable", not as "licence is visible". Registered here, before dispatch, so that the subgroup split cannot be chosen after seeing which way it falls.

What would refute the expectation

NAME(L) high, NAME(F) and NAME(S) and NAME(P) all low is a clean UNRATED verdict: the reader can point, the scale cannot, the three prior results are instrument findings, and this design's motivating worry is wrong. It is a real possible outcome, it is written here before dispatch, and it would be reported as the headline with the same prominence as any other.

6. Known limits, written before the data exists

1. Three seats, all LLMs, and their German competence is asserted by nobody. C1 is the only evidence in the run that they can read the German at all, and it is a floor, not a certificate. Every figure here is panel-perceived and internal-judgment-only; Tier D is NOT PASSED and no jury verdict carries evidential weight (charter §2.4).
2. One work, one language pair, one translator, one paragraph, ten sites. The L spans are the lead's judgments about what the German has, and a different translator would have chosen different features. Nothing here generalises to other pairs without being run on them.
3. L and F are matched on being syntactic and not on being the same construction. If the NAME rates differ, the difference could be licence or could be construction; §7's table is printed so a later reader can see exactly what was compared, and the verb-first subgroup is pre-registered in §5 for exactly this reason. 3a. Content parity is declared, not certified. Every L and F span says what its P span says — the rewrites are syntactic and the propositional content is held — and S is word-identical to P by construction. No independent content-parity call was made, as RS-20260808d made for its sixteen devices; the parity here rests on inspection and on the minimal-pair construction, and a reader who doubts a row can read it in §7.
4. F is unlicensed at the site, not in the segment. A F construction could accidentally answer to something elsewhere in the same German sentence; the registered claim is only that the German at that point does not have it.
5. The four versions of a site are near-identical strings seen by one seat across its stream (§3).
6. P's true answer is NONE by construction and by intent — but a determined reader can say that any faithful rendering "carries" its source. The prompt is worded to ask about departure from ordinary English, and C2 is the check that the wording worked. If C2 fails, that is the finding about the wording and it is reported as one.

7. The registered site table

Frozen here and in materials/sites.json. L locus is the exact German substring the L span claims to carry.

Three of the ten L spans (4, 5, 7) restore the German verb-first construction that the frozen translator's log records losing three times over (log items 7, 13, 23) — the paragraph's signature construction, and the reason §5 pre-registers a verb-first subgroup contrast. Two more (9, 10) are the literal alternatives the log records as live and not taken (items 19, 22). The L arm is therefore not an invention for the experiment: it is the road the translation did not take, written down at the time and frozen before this design existed.

The S spans, generated and printed here so the freeze is complete:

• site 1 — height certain reached having a
• site 2 — find is harder it to
• site 3 — as as altered our completely commerce has
• site 4 — has in art one been churches part of ashamed of Christendom If
• site 5 — of everyday piece and formed and be out Sunday, house the the one must church,
• site 6 — heightened people sense a he his own looked in at
• site 7 — account, Luther though every lived, splendour and plainly without by for
• site 8 — artists adorned and kind still by more richly of visited fertilized, every
• site 9 — affairs driven Germany were of these Many of by to state foreigners
• site 10 — bestowed because had no and it God these on others

8. The two prompt strings (frozen)

Pass A, one item per call:

You are reading an English translation of a German literary text.

Below is one sentence of the German original, and one sentence of an English
translation of it. In the English, one span is marked with ⟦ ⟧.

Your task concerns the marked span only.

The marked span may be a deliberate departure from ordinary English, made in
order to carry over into English a specific feature of the German — its word
order, its word-formation, its grammatical construction, its register, or its
figure. Or it may not be: it may be ordinary English, or it may be strange for
some reason of its own that answers to nothing in the German.

Rendering the German faithfully is not by itself such a departure. Every English
sentence here translates the German sentence beside it; the question is only
whether the marked span departs from ordinary English IN ORDER TO carry a
particular feature of the German across. If it does not, the answer is NONE.

Decide which, and answer in exactly one of these two forms, adding nothing else.

If the marked span carries over a specific feature of the German:
QUOTE: <the exact German words, copied from the German below, whose feature it carries>
FEATURE: <one clause naming the property of those German words that the marked span reproduces>

If it does not — if the marked span is ordinary English, or if its strangeness
carries over no specific feature of the German:
NONE

=== GERMAN ===
{german}

=== ENGLISH ===
{english}

Pass B, one site per call:

Below are four English versions of the same passage. They differ only inside the
span marked ⟦ ⟧.

Rate each marked span for how far it departs from ordinary present-day English
prose, on this scale:

0  ordinary English, nothing unusual
1  very slightly unusual
2  noticeably unusual but unremarkable in literary prose
3  distinctly odd
4  strongly odd
5  very strongly odd; a reader would stumble
6  extremely strange; barely English

Judge the English only. Answer with exactly four lines and nothing else:

A: <0-6>
B: <0-6>
C: <0-6>
D: <0-6>

=== A ===
{a}

=== B ===
{b}

=== C ===
{c}

=== D ===
{d}

9. Pre-flight cost estimate

Built from max_tokens, not from an expected output length — note (abc), the one estimate this project has overrun failed exactly there.

Per seat: ~39k input, ~13.2k output at the cap. At the config/models.md list prices — J1 $1.00/$6.00, J2 $1.50/$7.50, J3 $0.435/$0.87 — that is $0.114 + $0.158 + $0.029 = $0.301.

The declared worst case is $2.40, which is the list-price figure multiplied by 4× for provider routing (config/models.md's S022 caution: a slug's billed price has been 3.8× its listed one) and rounded up, plus $0.20 for one pre-run critic call at a 16,000 cap. Reasoning is suppressed where the provider accepts it (note (bkw)); a body that returns length with zero visible content is re-dispatched once at a raised cap and the dead body is preserved under runs/discarded/.

Today's ledger (UTC 2026-08-10) stands at $0.582308840991 of $5.00 across three sessions, leaving $4.417691 — so $2.40 fits with room. Actuals are read from each response's usage.cost and cross-checked against the key-usage delta.

10. Verification

verify.py recomputes every reported number by a path that imports nothing from analyse.py, and asserts: every German segment byte-identical to the frozen source file; every P span an exact substring of the frozen R06 segments; every L/F/S variant identical to P outside the marked span; every S span a strict case-folded word-multiset permutation of its P span; both prompt strings identical to materials/prompts.json; the item shuffle obeying the six-position separation constraint; body counts with runs/discarded/ excluded; a deterministic re-run reproducing results.json; and every mean and every exact P recomputed independently. Mutation testing: at least three seeded corruptions must each move a named figure, and at least one order-only permutation must move nothing.

Per the standing rule this is code and materials integrity, not scientific validity.

11. The pre-run critic, and what was changed before dispatch

One adversarial pass over this frozen design and the frozen translation, by x-ai/grok-4.5 — deliberately not one of the three judging seats. Verdict NEEDS-REDESIGN, 16 findings, 4 BLOCKING. Raw body at runs/critic.txt, prompt at materials/critic-prompt.txt, cost $0.0356864. Thirteen findings accepted, two accepted in part, one overruled in writing. Every change below was made before a single item was dispatched.

What the critic did not find, and it is worth recording: it did not challenge the source, the collation, the contamination declaration, the seat composition, or the cost estimate. Its four BLOCKING findings were all about the materials — three individual rows built wrong, and one statistical reading the material could not support — which is the failure mode a design this dependent on hand-built minimal pairs should expect.