Repository path: workshop/experiments/E-20260812c-grade-shift/design.md · rendered 2026-09-09

When the source marks contempt only in the grammar, does the translator put it back as a word?

Study limb of ARM-dakghar step 2 (T1). Second design. The first (design-v1-superseded.md) was frozen, criticised and killed before any datum — 27 blocking findings — and its finding 14 is the reason this one exists. The translation both hang on, T-dakghar-R05-v1 span B, was frozen and committed at 231e084 before either design was written.

1. The question

English has no second-person grade. Bengali has three. When a Bengali speaker drops from তুমি to তুই — the move from neutral address to intimate-or-contemptuous address — the translator into English has no grammatical slot to put it in, and the binding register of this work (V5, V6) forbids inventing one: no thou, no archaism, no compensating stiffness.

So what does the translator actually do? The hypothesis this design tests is that the rule is obeyed in the letter and broken in the substance: that where the source marks the drop only in the grammar, the English quietly acquires a lexical mark — a pejorative or diminutive vocative, an insult noun, a status noun — that the source does not have at that point.

And there is a second question folded into the first, which is the one this project keeps finding: does the translator know he has done it? The lead's translator's log sealed a claim at D33, written before any count and before the critic saw anything:

"the downward shift survives at [176] (monkey) and [188] (Here, boy), and is lost at [190], [238] and [242]."

That claim is now a registered prediction with a prior count of 0 added cues at grammar-only sites, and it is tested against the frozen artifact.

2. Materials — a census, not a selection

Every second-person grade site in span B, found by exhaustive search of the copy-text for the তুই paradigm and for the তুমি speeches of the same speakers. Not a chosen subset: the whole span. This is what the first design lacked and what its finding 29 demanded.

TUI sites — five, the whole of them in section ২:

TUMI control sites — four, speeches by the same two speakers at plain তুমি, including two that are sarcastic without any grade drop ([186], the headman's mock congratulation), so the control is not merely "polite text": [184], [186] (headman → Amal), [236], [248] (boys → Amal).

Nine speeches. build_items.py asserts every one verbatim against source-ipublishinghouse.txt and every lead rendering verbatim against the frozen translation before any dispatch.

3. Subjects

Three English renderings of each of the nine speeches:

• LEAD — the frozen T-dakghar-R05-v1, produced under a register that forbids grammatical compensation, by a translator who then sealed D33. Hypothesis-aware, and that is the point: this subject is the one whose self-report is on trial.
• H1 = moonshotai/kimi-k3, H2 = deepseek/deepseek-v4-pro — naive hands, translating the same speeches blind: no mention of pronouns, grade, contempt, footing, register or of an experiment, and no sight of any other rendering. They establish whether lexical compensation is a property of this translator or of the crossing.

4. Measure, and who codes it

For each (speech × rendering), a coder marks every lexical marker of contempt or of the addressee's inferior standing in the English — pejorative or diminutive vocative (boy, brat, lad, you there), insult noun (monkey, rascal), or dismissive status noun — and marks, for each, whether a corresponding word stands in the Bengali speech.

• added = 1 for a rendering that contains at least one such marker with no counterpart in the source speech; else 0.
• added_rate(cell) = mean of added over the renderings in that cell.
• false_add = added_rate on the four TUMI control speeches. A translator who sprinkles boy everywhere is not compensating for grade; the control is what tells the two apart, and it carries sarcastic source speeches precisely because those are where a spurious insult is most tempting.

Coding is done three ways and the lead's coding is not privileged. Two non-Anthropic seats (google/gemini-3.6-flash, x-ai/grok-4.5) code every rendering blind — blind to which rendering is the lead's, blind to D33, blind to the TUI/TUMI classification, and shown the Bengali only as a word-list to check counterparts against. The lead codes independently. The reported figure is the majority of three, and disagreements are listed.

5. Gates — withholding

• G1 (coding reliability). Three-way agreement on added must be ≥ 0.80 across the 27 codings. Below that the primary is not read and the page reports the disagreement instead.
• G2 (source-side coding). The TUI/TUMI and grammar-only/grammar+lexis classification of the nine source speeches is the lead's and is single-coded; it is a morphological fact (তুই paradigm vs তুমি paradigm) checkable from the text, and the tokens are listed in §2 so that any reader can check it. This is a declared deviation, not a control.

6. Predictions, registered before dispatch

• P1 (primary, and the test of D33). The lead's added_rate on the three grammar-only TUI sites is greater than 0 — i.e. the sealed self-report is wrong. D33 predicts 0.
• P2. The lead's added_rate on grammar-only sites exceeds its false_add on the four TUMI controls — the addition tracks the source's grade drop rather than being a tic.
• P3. At least one naive hand also shows added > 0 on at least one grammar-only site — the compensation is not peculiar to a translator working under V5.
• P4. On the two grammar + lexis sites every rendering carries a marker with a source counterpart (added = 0 there, because the cue is not added — it is translated).

Failure criteria. P1 fails if the lead's grammar-only added_rate is 0, and then D33 stands and the interesting result is that the register held. P2 fails if false_add is at or above the grammar-only rate, and then the primary is withheld: the additions are a stylistic habit, not compensation. Any G1 failure withholds everything.

7. What this cannot show, declared in advance

1. Nine speeches, five of them TUI, three of those grammar-only. That is the whole of section ২ — the census is complete and it is still small. No interval is computed and none is reported; the result is a count, stated as a count. The first design's fault was to dress a count of three in inferential clothes.
2. "Lexical marker of contempt" is a judgement, not a measurement. That is why it is coded three ways behind a reliability gate, and why disagreements are printed rather than resolved.
3. This says nothing about readers. It is a fact about what translators write, not about what anyone perceives. The reception question is what killed design v1 and it is not asked here.
4. One work, one language, one direction, three renderings. Bengali তুই is not every language's intimate pronoun and the headman is not every rude speaker.
5. The lead is contaminated by construction — it wrote the register, the translation and D33. It cannot be a naive subject and is not offered as one; it is the subject on trial. The naive hands are what carry P3, and P3 is the only sentence here that generalises beyond this translator.
6. Tier D has not passed. Every evaluative sentence downstream is provisional and internal-judgment-only.

8. Cost

Pre-flight ceiling, within the $0.50 declared for this experiment and now itemised again: six translation calls (cap 1,500 out; three speeches per call, two hands), two blind coder calls (cap 3,000 out), one pre-run critic call on this design (cap 8,000 out). Worst case from max_tokens, note (abc): $0.068 + $0.006 (hands) + $0.023 + $0.018 (coders) + $0.055 (critic) + prompts ≈ $0.19, on top of the $0.04057375 already spent criticising design v1. A body returning finish_reason == "length" is re-dispatched once at the same cap whether or not it has content, and is recorded dead and excluded if it truncates again.