Lost in the Request: How Communication Variation Disrupts Retrieval and Action in Email Agents

# Lost in the Request: How Communication Variation Disrupts Retrieval and Action in Email Agents
Source: [https://arxiv.org/html/2610.02627](https://arxiv.org/html/2610.02627)
###### Abstract

An email assistant should not complete less work simply because a user phrases the same request differently\. Yet most benchmarks test each task with only one canonical request, leaving this form of robustness largely unmeasured\. We test whether email assistants remain reliable when the requested information, available evidence, and expected outcome stay fixed, but the communication style or English variety changes\. We construct validated variants along five communication\-style axes and four rule\-based dialect conditions, and evaluate them on three benchmarks: a retrieval\-augmented generation \(RAG\) pipeline and two tool\-using agents\. Indirect requests reduce performance on all three benchmarks, while formal requests reduce performance on both agentic benchmarks\. Examining the systems more closely shows that these failures have different causes\. Verbose requests mainly hurt a lexical retriever by making the relevant email harder to find\. By contrast, indirect and dialect variants remain harmful even when the relevant email is retrieved\. In the agentic setting, indirect and formal requests mainly cause the agents to omit required actions, not to take more unsupported actions\. These results show that a successful response is not enough to establish robustness: evaluations should vary how requests are expressed and separately measure whether agents complete the requested work\.

## 1Introduction

An email assistant should not complete less work because a user asks indirectly, writes formally, adds surrounding context, or uses a different variety of English\. These differences are common in real communication, but they do not necessarily change what the user wants\. If the requested information, relevant evidence, named entities, and intended outcome remain the same, a reliable assistant should return the same answer or carry out the same actions\.

Current evaluations rarely test this expectation\. Most benchmarks represent each task with one canonical request, even though the same task can be expressed in many ways\. Prior work shows that semantically equivalent query formulations can change retrieval results, including for recent retrieval architectures\([Penha et al\., 2022](https://arxiv.org/html/2610.02627#bib.bib24);[Hagen et al\., 2024](https://arxiv.org/html/2610.02627#bib.bib25)\), and that linguistic reformulation affects both retrieval and answer generation in retrieval\-augmented question answering\([Cao et al\., 2026](https://arxiv.org/html/2610.02627#bib.bib6)\)\. Performance also varies across systematic varieties of English\([Ziems et al\., 2022](https://arxiv.org/html/2610.02627#bib.bib18);[Ziems et al\., 2023](https://arxiv.org/html/2610.02627#bib.bib12);[Sap et al\., 2019](https://arxiv.org/html/2610.02627#bib.bib19);[Srirag et al\., 2025](https://arxiv.org/html/2610.02627#bib.bib23)\)\. These findings establish that wording matters for finding and answering information, but they leave an important question for interactive agents: can a change in wording alter whether the requested work is completed? This question matters because an agent can fail without producing an obviously bad response\. An assistant may identify the task, call some appropriate tools, and reply fluently while still omitting an action needed to finish the request\. Conversely, it may retrieve the wrong evidence or take an unsupported action\. These failures require different responses from system designers, but a single end\-to\-end success score does not distinguish them\. Diagnostic agent benchmarks similarly show that aggregate scores can combine failures such as skipping a required tool, ignoring a tool result, or producing unsupported output\([Kokane et al\., 2024](https://arxiv.org/html/2610.02627#bib.bib21);[Soni, 2026](https://arxiv.org/html/2610.02627#bib.bib22)\)\. Robustness evaluation should therefore ask not only whether performance changes, but what the system stopped doing correctly\.

We study whether email assistants remain reliable when the same request is expressed in different ways\. We evaluate the two settings for email assistants: the*retrieval\-augmented generation*\(RAG\) setting, where the system retrieves evidence from the archive and answers from it, and the*agentic*setting, where the system calls tools to carry out the request\. Covering both lets us examine whether communication variation affects the information an assistant returns as well as the actions it performs\. We vary communication style and English variety while keeping the underlying task and expected outcome unchanged, and validate the resulting requests to ensure that they preserve the original intent\. Our work makes three contributions:

- •A framework for generating validated request variants\.Figure[1](https://arxiv.org/html/2610.02627#S2.F1)shows it\. An LLM rewriter produces the communication\-style poles and a rule\-based transducer produces the dialect varieties, and a set of gates admits a variant only once it still requests the information or action asked for, carries the relevant task details, and expects the same outcome\.
- •An evaluation spanning RAG and agentic settings\.We apply the framework to email question answering and to tool\-using agents, so that the same variation can be traced through retrieval, answering, and execution\.
- •A closer analysis of assistant failures\.We examine not only whether performance changes, but also whether failures come from finding the wrong information, producing an incorrect answer, omitting required actions, or taking unsupported actions\.

## 2Evaluation Framework

We evaluate whether a system fulfills the same request when the request is expressed differently\. We consider two forms of variation\.*Communication style*changes how a writer presents a request\.*Dialect*changes grammatical and dialectal features according to patterns documented for varieties of English\. Figure[1](https://arxiv.org/html/2610.02627#S2.F1)shows how a variant is built and follows one request through a RAG pipeline\.

Figure 1:Variant construction, and one EnronQA request carried through it\. \(a\) A request is rewritten into ten communication\-style poles and four English varieties; a variant enters the evaluation set only after clearing the gates of §[2\.3](https://arxiv.org/html/2610.02627#S2.SS3)\. \(b\) The same request under three variants\. The gold email is retrieved at rank 1 in all three cases, but the indirect variant produces an incorrect answer\.ref,con, anddiadenote reference, contrastive, and dialect conditions, respectively\.### 2\.1Variation Taxonomy

#### 2\.1\.1Communication\-Style Axes

The five axes and the definition of each pole are listed in Table[1](https://arxiv.org/html/2610.02627#S2.T1)\. Each style axis pairs a*reference*pole, a plain rewrite that carries no marked style, with a*contrastive*pole that carries the property under test\. Formality is the exception: neither a formal nor a casual register is neutral, so both of its poles are contrastive\. Each axis rests on a documented distinction: directness on the contrast between on\-record and off\-record requests\([Brown and Levinson, 1987](https://arxiv.org/html/2610.02627#bib.bib2);[Mannekote et al\., 2025](https://arxiv.org/html/2610.02627#bib.bib3)\); formality on register variation\([Pavlick and Tetreault, 2016](https://arxiv.org/html/2610.02627#bib.bib4);[Rao and Tetreault, 2018](https://arxiv.org/html/2610.02627#bib.bib5)\); verbosity on evidence that input length and the position of the relevant information affect model behavior\([Levy et al\., 2024](https://arxiv.org/html/2610.02627#bib.bib7);[Liu et al\., 2024](https://arxiv.org/html/2610.02627#bib.bib8)\); emotion on evidence that emotional framing changes behavior on an otherwise fixed task\([Li et al\., 2023](https://arxiv.org/html/2610.02627#bib.bib9);[Sharma et al\., 2024](https://arxiv.org/html/2610.02627#bib.bib10)\); and grammaticality on reported sensitivity to grammatical form\([Cao et al\., 2026](https://arxiv.org/html/2610.02627#bib.bib6)\)and on work measuring encoder robustness to grammatical error\([Yin et al\., 2020](https://arxiv.org/html/2610.02627#bib.bib31)\)\. Appendix[A](https://arxiv.org/html/2610.02627#A1)describes each axis in full\.

Table 1:The five communication\-style axes and their poles\.*ref*marks a reference pole, a plain rewrite similar to the original query;*con*marks a contrastive pole\.
#### 2\.1\.2Dialect Conditions

We evaluate four varieties: Indian English, Colloquial Singapore English, Black South African English, and Nigerian English\. Dialect differs from the non\-grammatical pole in both source and interpretation\. The non\-grammatical pole is a constructed set of learner\-like slips deliberately detached from any speech community, whereas the dialect conditions apply features documented for particular varieties\. The variants are synthetic stress tests rather than samples of how individual speakers write, and every claim we make concerns system behavior rather than speakers\.

### 2\.2Variant Construction

##### Communication\-style rewriting\.

Every query is rewritten once into each of the ten poles\. Rewriting usesgpt\-5\.4\-miniat temperature1\.01\.0, with a per\-pole prompt that describes the target style and forbids changing fact\. For the six contrastive poles we use verbalized sampling\([Zhang et al\., 2025](https://arxiv.org/html/2610.02627#bib.bib11)\): one call returnsk=6k\{=\}6candidates, each written to a different framing drawn from a per\-pole menu, which is what keeps a pole from collapsing onto similar phrasing\. Among candidates clearing every applicable gate described in[2\.3](https://arxiv.org/html/2610.02627#S2.SS3), we keep the one whose opening is least used so far in the run, so that no fixed menu position is systematically favored\. The four reference poles are generated as a single plain rewrite with maximum of 3 retries\. Full prompts and menus are in Appendix[D](https://arxiv.org/html/2610.02627#A4)\.

##### Rule\-based dialect transformation\.

Dialect variants are produced by the transducer of[Ziems et al\. \(2023\)](https://arxiv.org/html/2610.02627#bib.bib12),111[https://value\-nlp\.github\.io](https://value-nlp.github.io/)which maps Standard American English into synthetic dialect text by applying sets of grammatical and dialectal features documented for each variety\. Generation is rule\-based and reproducible under a fixed seed \(we use 0\), at a feature density1\.01\.0\. Operative literals such as usernames, addresses, times, and event names are masked before transformation and restored afterwards, so they survive byte\-identically\.

### 2\.3Variant Validation

Across a base query and all of its variants we hold fixed the information need or requested action, the supporting evidence, named entities, numbers, dates, protected literals, and the expected outcome\. Only the wording may differ\. A candidate enters the evaluation set only after clearing every gate that applies to its pole\.

A deterministic*factual gate*requires the variant to be non\-empty, to differ from the original, and to preserve every numeric token and protected literal\. An*intent gate*, a rubric judge \(gpt\-5\.4, temperature0\.00\.0\), asks whether the variant requests exactly the same information\. This gate is what protects the paired design: a variant that clears it is still answerable by the same gold answer, so a difference in outcome cannot be a difference in task\. A*style gate*asks whether the variant realizes its target pole, given that pole’s operational definition from Table[1](https://arxiv.org/html/2610.02627#S2.T1); for dialect it instead requires at least two features documented in the electronic World Atlas of Varieties of English to have changed\([Kortmann et al\., 2013](https://arxiv.org/html/2610.02627#bib.bib29)\)\. A*human\-likeness gate*asks whether the text reads as something a person would type, which guards against rewrites that reach the style label while reading synthetic\. It applies to four of the six contrastive poles except verbose and non\-grammatical \(prompts in Appendix[E](https://arxiv.org/html/2610.02627#A5)\)\. Two authors independently labeled the same 120 variants, 20 per contrastive pole, answering the questions the gates ask\. Agreement was complete on intent \(κ=1\.00\\kappa\{=\}1\.00\), and reached82\.5%82\.5\\%on style and80\.0%80\.0\\%on human\-likeness\.

## 3Experimental Setup

### 3\.1Evaluation Tasks

We apply the framework to three tasks, one RAG task and two agentic benchmarks\.

##### EnronQA \(RAG\)\.

Retrieval\-augmented question answering over a personal email archive[Ryan et al\. \(2025\)](https://arxiv.org/html/2610.02627#bib.bib1)\. We sample 400 questions from the training split under an isolated\-axes design: four bins of 100, where each complex bin flips exactly one complexity property against the*simple*control\. The three complex bins are*3\-hop*\(three or more reasoning hops\),*multi\-question*\(the question asks for two or more distinct things\), and*threaded*\(resolution depends on earlier messages in the thread\)\(detail in Appendix[B](https://arxiv.org/html/2610.02627#A2)\)\. Holding the other two properties at their control setting is what makes the per\-bin breakdown of §[4\.2\.1](https://arxiv.org/html/2610.02627#S4.SS2.SSS1)attributable to one property rather than to overall difficulty\. Each question is paired with a gold answer and the gold email that contains it, and retrieval runs over the target user’s own inbox\.

##### ToolTalk \(Agentic\)\.

A multi\-turn tool\-using benchmark\([Farn and Shin, 2023](https://arxiv.org/html/2610.02627#bib.bib13)\)\. We use the*hard*split: 50 conversations comprising 194 user turns over 28 tool APIs\. Grading compares the agent’s calls against ground\-truth API calls\. The harness is teacher\-forced: the conversation is rebuilt from the gold trajectory and the databases are reset before each assistant turn, so an error on one turn cannot propagate to the next\.

##### WorkBench \(Agentic\)\.

A sandbox benchmark of workplace tasks across five domains\([Styles et al\., 2024](https://arxiv.org/html/2610.02627#bib.bib32)\)\. We use its 90\-task email domain, the subset solvable with the email toolkit alone, and expose all 26 tools on every task so that the agent must still select among them\. Grading is by database state: a task is correct only if executing the agent’s predicted calls leaves the sandbox in the ground\-truth state, which makes the metric binary and unforgiving of a single wrong argument\. We report task accuracy\.

### 3\.2Models and Retrieval Systems

##### Retrieval\.

EnronQA is run with three retrievers over the same corpus, all returning the topk=5k\{=\}5emails\. BM25 is lexical \(rank\_bm25\) and indexes the full text of every email\([Robertson and Zaragoza, 2009](https://arxiv.org/html/2610.02627#bib.bib26)\)\. SPLADE is learned\-sparse \(naver/splade\-v3\), producing term\-weight vectors scored by dot product\([Formal et al\., 2021](https://arxiv.org/html/2610.02627#bib.bib27)\)\. BGE\-M3 is dense \(BAAI/bge\-m3\)\([Chen et al\., 2024](https://arxiv.org/html/2610.02627#bib.bib28)\)\. Both neural retrievers truncate to 512 tokens\. Retrieval is identical within a retriever, so any difference between answerers under a fixed retriever is a generation difference\. We use SPLADE as the headline retriever because it provides the strongest baseline retrieval among the tested systems\.

##### Answering and grading\.

We evaluate three models:gpt\-5\.6\-terra,gpt\-5\.4\-mini, andKimi\-K2\.6\. All headline results usegpt\-5\.6\-terra, which is run on all three benchmarks\. The cross\-model comparison of §[4\.3](https://arxiv.org/html/2610.02627#S4.SS3)is restricted to EnronQA and ToolTalk: running the full ten\-pole grid on WorkBench for the two additional models was beyond the compute available, so that benchmark is reported on the headline model alone and the coverage actually obtained is stated in Table[6](https://arxiv.org/html/2610.02627#A8.T6)so that what is missing is explicit\. On EnronQA each model receives the retrieved emails and the question and is instructed to answer in one sentence using only the information in the email\. Each agent condition is repeated withK=3K\{=\}3replicates whose per\-scenario metrics are averaged before any test\. Answers on EnronQA are graded bygpt\-5\.4at temperature00, which sees the email, the question, the gold answer, and the candidate answer and returns a single\-line judgement\.

### 3\.3Evaluation Metrics

Each task has its own success metric, so a percentage point is not the same quantity across benchmarks and magnitudes should not be compared between columns\. EnronQA reports judged answer accuracy, and additionally Recall@5, the rate at which the gold email appears in the retrieved set\. ToolTalk reports conversation success, which requires every ground\-truth call to be matched with no incorrect action, along with recall of required calls, precision, and the incorrect\-action rate\. WorkBench reports task accuracy against the ground\-truth sandbox state, and additionally the F1 between predicted and ground\-truth tool calls, the rate at which no tool call is issued, and the rate at which execution leaves an unwanted change to the sandbox\.

### 3\.4Statistical Analysis

Every variant is compared against the base item on the items present in both conditions\. Detail shows in Appendix[G](https://arxiv.org/html/2610.02627#A7)\. For EnronQA, each item yields one binary outcome from a single run, so the paired difference is−1\-1,00or\+1\+1\. We use the exact McNemar test on those pairs, which conditions on their number and is exact at any sample size\. For the other two agent benchmarks, each condition is runK=3K\{=\}3times, so for itemiiwe average a metric over replicates before any test, then form the paired differencedi=v¯i−n¯id\_\{i\}=\\bar\{v\}\_\{i\}\-\\bar\{n\}\_\{i\}against the base item\. We test the mean of those differences with a paired percentile bootstrap over items, resampling item indices with replacementB=10,000B\{=\}10\{,\}000times under a fixed seed, reading the 95% interval off the2\.52\.5and97\.597\.5percentiles of the resampled means and a two\-sidedpp\-value off the proportion of them falling either side of zero\.

## 4Results

### 4\.1Effects of Style and Dialect Variation

Table[2](https://arxiv.org/html/2610.02627#S4.T2)gives the effect of every condition on all three benchmarks, intervals andpp\-values in Table[7](https://arxiv.org/html/2610.02627#A8.T7)\. Indirect phrasing is the only condition that degrades every benchmark\. It costs−3\.2\-3\.2points of answer accuracy on EnronQA \(p=\.007p\{=\}\.007\),−22\.4\-22\.4points of conversation success on ToolTalk \(p<\.001p\{<\}\.001\), and−10\.7\-10\.7points of task accuracy on WorkBench \(p=\.002p\{=\}\.002\)\.

Formality separates the retrieval task from the two agentic ones\. Both formality poles are null on EnronQA \(−1\.0\-1\.0each,p=\.39p\{=\}\.39andp=\.34p\{=\}\.34\), whileformalis significant on both agent benchmarks, costing−13\.9\-13\.9on ToolTalk and−12\.2\-12\.2on WorkBench\. Thecasualpole is not significant on any benchmark, making the formal pole the more consistently harmful in this axis\.

Dialect variation degrades every benchmark in the mean, but its strength varies by variety\. On EnronQA all four varieties are significant, from−2\.6\-2\.6\(Singapore\) to−3\.9\-3\.9\(Indian\)\. Singapore is significant on all three benchmarks and Indian on two, whereas Black South African and Nigerian reach significance on one\. On EnronQA, the pooled dialect estimate is numerically larger than the estimate for the non\-grammatical pole:−3\.0\-3\.0versus−1\.8\-1\.8percentage points\.

The remaining conditions are selective\. Verbosity is significant on EnronQA \(−3\.0\-3\.0\) alone\. Emotionally charged phrasing reaches significance on WorkBench \(−8\.5\-8\.5\)\. Non\-grammatical phrasing is significant only on ToolTalk \(−9\.3\-9\.3\)\. The reference poles behave as the control they are intended to be\. Across the1212reference cells of the headline model, only one reachesp<\.05p\{<\}\.05:conciseon ToolTalk \(−9\.0\-9\.0,p=\.010p\{=\}\.010\), where shortening an already\-terse instruction is not a neutral operation\. Input\-side similarity \(§[2\.3](https://arxiv.org/html/2610.02627#S2.SS3)\) shows the same separation between reference and contrastive rewrites before any system is involved\.

Table 2:Unperturbed baseline and pairedΔ\\Deltaper condition, in points on each benchmark’s own metric, for the headline modelgpt\-5\.6\-terra\(EnronQA on SPLADE\)\.*ref*rows are the meaning\-preserving reference poles and*con*rows the contrastive poles; formality has no neutral pole, so both of its rows are contrastive\. Per\-cellnn, intervals and the test used for each benchmark \(§[3\.4](https://arxiv.org/html/2610.02627#S3.SS4)\) are in Table[7](https://arxiv.org/html/2610.02627#A8.T7)\. Rawpp:\*\{\}^\{\\textrm\{\*\}\}p<\.05p\{<\}\.05,\*\*\{\}^\{\\textrm\{\*\*\}\}p<\.01p\{<\}\.01,\*\*\*\{\}^\{\\textrm\{\*\*\*\}\}p<\.001p\{<\}\.001\.
### 4\.2Where the Loss Arises

#### 4\.2\.1RAG Setting

On EnronQA we can separate retrieval from response, because Recall@5 reports whether the gold email reached the reader at all\. Figure[2](https://arxiv.org/html/2610.02627#S4.F2)contrasts two poles across three retrievers\.

Figure 2:AccuracyΔ\\Deltaand gold\-email recallΔ\\Deltafor verbose and indirect against unperturbed request under three retrievers on EnronQA\. Shading marks the gap between the part of the accuracy change not matched by a change in retrieval\. Values are averaged over three answerer models; per\-model values in Table[10](https://arxiv.org/html/2610.02627#A8.T10)\.The verbosity effect turns out to be retriever\-dependent\. Under lexical BM25,verbosecosts−26\.0\-26\.0points of accuracy, and Recall@5 falls by−28\.0\-28\.0: the padding pushes the gold email out of the top five, leading to inaccurate answers\. Under learned\-sparse SPLADE recall forverboseis close to the baseline \(−0\.2\-0\.2,p=\.89p\{=\}\.89\) while accuracy falls only−3\.0\-3\.0\. Under BM25, we further explore the effect of padding length, over theverboseslots whose gold email was reachable at baseline, Recall@5 falls from95\.7%95\.7\\%in the least\-padded quartile to44\.6%44\.6\\%in the most\-padded one \(Table[14](https://arxiv.org/html/2610.02627#A8.T14)\)\.

Indirect and dialect phrasing survive the retrieval drop\. Under SPLADE,indirectcosts−3\.2\-3\.2points of accuracy while recall moves−1\.8\-1\.8, and pooled dialect costs−3\.0\-3\.0points of accuracy against−0\.9\-0\.9of recall\. In both cases the answerer still answers worse, so the loss is downstream of retrieval\. Per\-pole retrieval effects are in Table[15](https://arxiv.org/html/2610.02627#A8.T15), and retrieval quality with per\-pole recall intervals are in Table[11](https://arxiv.org/html/2610.02627#A8.T11)\.

We labeled the118118headline\-condition slots \(gpt\-5\.6\-terra×\\timesSPLADE\) where the gold email was retrieved and the answer was still judged wrong, one code per slot \(Appendix[F](https://arxiv.org/html/2610.02627#A6)\)\. Unsupported content is the largest category; what separates the rewrite\-induced slice is misreading the request rather than declining to answer \(Table[8](https://arxiv.org/html/2610.02627#A8.T8)\)\. When the gold email never reaches it at all, the model answers confidently far more often than reporting the fact absent \(Appendix[F](https://arxiv.org/html/2610.02627#A6)\)\.

Two of the three complex bins are more fragile than the control\. Contrastive rewrites break4\.26%4\.26\\%of previously\-correct slots on 3\-hop questions and4\.08%4\.08\\%on multi\-question ones, against1\.62%1\.62\\%on the simple control and1\.88%1\.88\\%on threaded questions\. Dialect variation follows the same ordering at a uniformly higher level, from2\.91%2\.91\\%on threaded to5\.36%5\.36\\%on multi\-question \(Table[9](https://arxiv.org/html/2610.02627#A8.T9)and Figure[4](https://arxiv.org/html/2610.02627#A8.F4)report how often a rewrite broke a question\)\.

#### 4\.2\.2Agentic Setting

Figure 3:PairedΔ\\Deltafor the indirect and formal poles\. Vermillion marks an omission measure and blue an unsupported\-action measure; shade marks the model\. \(a\) ToolTalk required\-call recall, \-17\.0 to \-6\.8 pp over 6 pole×\\timesmodel cells, 6 significant; \(b\) ToolTalk incorrect\-action rate, \-1\.9 to \+4\.1 pp, 0 of 6 significant; \(a\) and \(b\) share a scale\. \(c\) WorkBench on the headline model, the share of tasks with no tool call beside the share leaving an unwanted change to the sandbox\. Per\-pole and per\-model values in Table[16](https://arxiv.org/html/2610.02627#A8.T16)and Table[17](https://arxiv.org/html/2610.02627#A8.T17)\.\*\{\}^\{\\textrm\{\*\}\}p<\.05p\{<\}\.05,\*\*\{\}^\{\\textrm\{\*\*\}\}p<\.01p\{<\}\.01,\*\*\*\{\}^\{\\textrm\{\*\*\*\}\}p<\.001p\{<\}\.001\.The agents can have two failure modes: the agent could take wrong actions, or it could take too few\. Figure[3](https://arxiv.org/html/2610.02627#S4.F3)separates them\. On ToolTalk the effect is concentrated in a failure to complete required work\. Required\-call recall falls by−11\.28\-11\.28points underindirect, and−10\.32\-10\.32underformal\(bothp<\.001p\{<\}\.001\)\. The incorrect\-action rate moves in neither pole:\+2\.96\+2\.96\(p=\.14p\{=\}\.14\) and−1\.94\-1\.94\(p=\.35p\{=\}\.35\)\.

WorkBench shows the similar mechanism \(Figure[3](https://arxiv.org/html/2610.02627#S4.F3)c; Table[17](https://arxiv.org/html/2610.02627#A8.T17)\)\. Underformalthe share of tasks with no tool calls rises from0\.4%0\.4\\%to12\.2%12\.2\\%\(p<\.001p\{<\}\.001\), while unwanted side effects fall from8\.9%8\.9\\%to4\.8%4\.8\\%\(p=\.027p\{=\}\.027\)\. Across both agentic benchmarks the poles that reduce task success reduce the amount of required work performed while leaving the rate of unsupported actions unchanged\.

### 4\.3Generality Across Models and Systems

Table 3:Unperturbed baseline and pairedΔ\\Deltaper contrastive pole for three models\. The first row is the unperturbed score in percent on each benchmark’s own metric; every row below is the pairedΔ\\Deltain percentage points for one contrastive pole\. Restricted to the two benchmarks where all three models were run; the headline model’s results on all three are in Table[2](https://arxiv.org/html/2610.02627#S4.T2)and the coverage obtained is in Table[6](https://arxiv.org/html/2610.02627#A8.T6)\. Dialect is in Table[12](https://arxiv.org/html/2610.02627#A8.T12)\.\*\{\}^\{\\textrm\{\*\}\}p<\.05p\{<\}\.05,\*\*\{\}^\{\\textrm\{\*\*\}\}p<\.01p\{<\}\.01,\*\*\*\{\}^\{\\textrm\{\*\*\*\}\}p<\.001p\{<\}\.001\.Table[3](https://arxiv.org/html/2610.02627#S4.T3)shows every contrastive pole for three models on the two systems where all three models were run \(model coverage shown in Table[6](https://arxiv.org/html/2610.02627#A8.T6)\)\. Indirect phrasing is the one pole that generalizes: it is negative in all six model\-by\-system cells and significant in five, whileformaland the non\-grammatical pole are each significant in two cells, both on the agent\. Higher baseline performance does not imply greater fairness in our experiments\.gpt\-5\.6\-terrahas the highest unperturbed baseline on both systems, yet on the same twelve contrastive style cells and eight dialect cells it is significant in55and66, against44and44forgpt\-5\.4\-miniand11and00forKimi\-K2\.6\.

## 5Discussion

### 5\.1Request Explicitness Affects Performance

The most consistent communication\-style effect is not the one that changes the most text, but hiding the main request in the query\. Parsing every variant against its own original with a dependency parser\([Honnibal et al\., 2020](https://arxiv.org/html/2610.02627#bib.bib30)\)\(Table[13](https://arxiv.org/html/2610.02627#A8.T13)\), the indirect pole retains the head verb of the request as the main predicate least often on all three benchmarks, from12\.5%12\.5\\%of slots on EnronQA down to4\.7%4\.7\\%on WorkBench, and the formal pole is in the bottom three poles on all three\. The two poles reach that position differently: the indirect pole deletes the original head verb in16%16\\%to40%40\\%of slots, whereas the formal pole retains it in68%68\\%to91%91\\%of slots but frequently places it in a subordinate clause\. This pattern suggests that performance is influenced by changing how explicitly the request is expressed\([Austin, 1975](https://arxiv.org/html/2610.02627#bib.bib15);[Searle, 1975](https://arxiv.org/html/2610.02627#bib.bib14);[Mannekote et al\., 2025](https://arxiv.org/html/2610.02627#bib.bib3)\)\. Recent evaluations report that models remain weaker at recovering indirect illocutionary force than at interpreting explicit requests\([Koo et al\., 2025](https://arxiv.org/html/2610.02627#bib.bib16);[Solidjonov, 2026](https://arxiv.org/html/2610.02627#bib.bib17)\)\.

### 5\.2Incomplete Fulfillment as an Agentic Failure Mode

An end\-to\-end success score combines two failures with different consequences: an agent can take an unwanted action, or it can omit an action that the request requires\. For the indirect and formal poles, the dominant change is in the second type\. Shown in both agentic settings, the affected agents complete less of the requested work without a corresponding increase in unsupported actions\.

The WorkBench trajectories show two routes into that outcome \(Table[17](https://arxiv.org/html/2610.02627#A8.T17)\)\. Underformalthe agent more often never starts: the share of tasks with no tool call at all rises from0\.4%0\.4\\%to12\.2%12\.2\\%\(p<\.001p\{<\}\.001\), close to the−12\.2\-12\.2point accuracy loss\. Underindirectthat share moves only to1\.5%1\.5\\%\(p=\.25p\{=\}\.25\), while the F1 between predicted and ground\-truth calls falls−8\.42\-8\.42points \(p<\.001p\{<\}\.001\), so the agent acts but the calls land in the wrong state\. This result complements work showing that aggregate success rates do not identify which tool\-use failure occurred\([Kokane et al\., 2024](https://arxiv.org/html/2610.02627#bib.bib21);[Soni, 2026](https://arxiv.org/html/2610.02627#bib.bib22)\)\. We hold the task and tool environment fixed and find that wording alone moves a system between completing a task and skipping required calls\.

### 5\.3Dialect Robustness and Fair Access

Dialect variation reduces performance on all three benchmarks\. Our finding of performance drops across English varieties is consistent with the disparities reported for language understanding\([Ziems et al\., 2022](https://arxiv.org/html/2610.02627#bib.bib18);[Ziems et al\., 2023](https://arxiv.org/html/2610.02627#bib.bib12);[Srirag et al\., 2025](https://arxiv.org/html/2610.02627#bib.bib23)\), for social tasks such as hate\-speech detection\([Sap et al\., 2019](https://arxiv.org/html/2610.02627#bib.bib19)\), and for model judgments about people\([Hofmann et al\., 2024](https://arxiv.org/html/2610.02627#bib.bib20)\)\. The dialect transformations apply grammatical features documented for each variety, whereas the non\-grammatical pole is a constructed set of slips produced by a language model, which suggests that variation beyond isolated grammatical slips carries part of the effect\. It is worth noticing that dialect variants are generated at feature density1\.01\.0, the attested upper bound, so this condition is a deliberate worst case\.

### 5\.4Implications for Evaluation and System Design

Three implications can be derived from the results\. First, a task should not be represented by a single canonical wording\. Retrieval research has shown that results move under semantically equivalent query variation\([Penha et al\., 2022](https://arxiv.org/html/2610.02627#bib.bib24);[Hagen et al\., 2024](https://arxiv.org/html/2610.02627#bib.bib25)\), and[Cao et al\. \(2026\)](https://arxiv.org/html/2610.02627#bib.bib6)reports the same fragility in the RAG setting\. Our results extend that requirement to the agentic setting, where the same variation changes whether the requested work is carried out\.

Second, robustness evaluation should be stage\-aware, because the same drop in a success score has different causes and therefore different fixes\. Replacing BM25 with SPLADE removes most of the verbosity penalty at a comparable baseline\. The indirect and dialect penalties persist when the gold evidence is retrieved\. Component\-level measurement has to come before the choice of mitigation, and choosing a stronger model is not a substitute for it: the model with the highest baseline on both systems where all three models were run is also the one whose results move significantly in the most conditions\.

Third, completeness and unsupported action should be reported separately\. Required\-call recall, failure to act, and read\-without\-write behavior are computable from a trajectory without a gold final answer, so they can be monitored where end\-state grading is unavailable\. They are not a substitute for final grading, but they expose a failure that response quality and safety metrics miss, and they distinguish an agent that has become appropriately cautious from one that is quietly doing less of the requested work\.

## 6Conclusion

We tested whether RAG\-based and agentic assistants fulfill the same task when only the request’s communication style or English variety changes\. Indirect requests produced the most consistent style loss; verbosity primarily disrupted lexical retrieval; and dialect sensitivity remained after accounting for whether the gold email was retrieved\. Across the agentic benchmarks, indirect and formal requests mainly reduced completion of required actions\. Evaluations should therefore vary request formulation, inspect intermediate retrieval and execution signals, and report required\-action completion separately from unsupported action\.

## Limitations

There are several limitations in this work\. First, dialect variants are generated at feature density1\.01\.0only, a deliberate worst case rather than an estimate of typical use\. Second, the WorkBenchindirectpole passes only6464of9090slots, leading to lower coverage\. Third, on ToolTalk theconcisereference pole moves−9\.0\-9\.0points, so the reference poles are an imperfect control there\. Besides, the three benchmarks use different metrics at different operating points, so cross\-benchmark comparison is qualitative, and the cross\-model comparison covers two of the three because running the full grid on WorkBench was beyond the available compute\. Finally, no mitigation approach is tested, and establishing which intervention offers the best trade\-off is left to future work\.

## References

- Austin \(1975\)J\. L\. AustinHow to do things with words\.Vol\.5,Harvard university press\.Cited by:[§5\.1](https://arxiv.org/html/2610.02627#S5.SS1.p1.1)\.
- Brown and Levinson \(1987\)P\. Brown and S\. C\. LevinsonPoliteness: some universals in language usage\.Vol\.4,Cambridge university press\.Cited by:[Appendix A](https://arxiv.org/html/2610.02627#A1.SS0.SSS0.Px1.p1.1),[§2\.1\.1](https://arxiv.org/html/2610.02627#S2.SS1.SSS1.p1.1)\.
- Caoet al\.\(2026\)T\. Cao, N\. Bhandari, A\. Yerukola, A\. Asai, and M\. SapOut of style: rag’s fragility to linguistic variation\.InProceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 280–318\.Cited by:[Appendix A](https://arxiv.org/html/2610.02627#A1.SS0.SSS0.Px2.p1.1),[§1](https://arxiv.org/html/2610.02627#S1.p2.1),[§2\.1\.1](https://arxiv.org/html/2610.02627#S2.SS1.SSS1.p1.1),[§5\.4](https://arxiv.org/html/2610.02627#S5.SS4.p1.1)\.
- Chenet al\.\(2024\)J\. Chen, S\. Xiao, P\. Zhang, K\. Luo, D\. Lian, and Z\. LiuBge m3\-embedding: multi\-lingual, multi\-functionality, multi\-granularity text embeddings through self\-knowledge distillation\.arXiv preprint arXiv:2402\.032164\(5\)\.Cited by:[§3\.2](https://arxiv.org/html/2610.02627#S3.SS2.SSS0.Px1.p1.1)\.
- Farn and Shin \(2023\)N\. Farn and R\. ShinTooltalk: evaluating tool\-usage in a conversational setting\.arXiv preprint arXiv:2311\.10775\.Cited by:[§3\.1](https://arxiv.org/html/2610.02627#S3.SS1.SSS0.Px2.p1.1)\.
- Formalet al\.\(2021\)T\. Formal, B\. Piwowarski, and S\. ClinchantSPLADE: sparse lexical and expansion model for first stage ranking\.InProceedings of the 44th international ACM SIGIR conference on research and development in information retrieval,pp\. 2288–2292\.Cited by:[§3\.2](https://arxiv.org/html/2610.02627#S3.SS2.SSS0.Px1.p1.1)\.
- Hagenet al\.\(2024\)T\. Hagen, H\. Scells, and M\. PotthastRevisiting query variation robustness of transformer models\.InFindings of the Association for Computational Linguistics: EMNLP 2024,pp\. 4283–4296\.Cited by:[§1](https://arxiv.org/html/2610.02627#S1.p2.1),[§5\.4](https://arxiv.org/html/2610.02627#S5.SS4.p1.1)\.
- Hofmannet al\.\(2024\)V\. Hofmann, P\. R\. Kalluri, D\. Jurafsky, and S\. KingDialect prejudice predicts ai decisions about people’s character, employability, and criminality\.arXiv preprint arXiv:2403\.00742\.Cited by:[§5\.3](https://arxiv.org/html/2610.02627#S5.SS3.p1.1)\.
- Honnibalet al\.\(2020\)M\. Honnibal, I\. Montani, S\. Van Landeghem, and A\. BoydSpaCy: industrial\-strength natural language processing in python\.Zenodo\.Cited by:[§5\.1](https://arxiv.org/html/2610.02627#S5.SS1.p1.1)\.
- Kokaneet al\.\(2024\)S\. Kokane, M\. Zhu, T\. Awalgaonkar, J\. Zhang, T\. Hoang, A\. Prabhakar, Z\. Liu, T\. Lan, L\. Yang, J\. Tan,et al\.Toolscan: a benchmark for characterizing errors in tool\-use llms\.arXiv preprint arXiv:2411\.13547\.Cited by:[§1](https://arxiv.org/html/2610.02627#S1.p2.1),[§5\.2](https://arxiv.org/html/2610.02627#S5.SS2.p2.1)\.
- Kooet al\.\(2025\)Y\. Koo, J\. Lee, D\. Park, S\. Park, and S\. LeeEvaluating large language models on understanding korean indirect speech acts\.arXiv preprint arXiv:2502\.10995\.Cited by:[§5\.1](https://arxiv.org/html/2610.02627#S5.SS1.p1.1)\.
- Kortmannet al\.\(2013\)B\. Kortmann, K\. Lunkenheimer, and K\. EhretThe electronic world atlas of varieties of english\.Leipzig: Max Planck Institute for Evolutionary Anthropology\.Cited by:[§2\.3](https://arxiv.org/html/2610.02627#S2.SS3.p2.1)\.
- Levyet al\.\(2024\)M\. Levy, A\. Jacoby, and Y\. GoldbergSame task, more tokens: the impact of input length on the reasoning performance of large language models\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 15339–15353\.Cited by:[Appendix A](https://arxiv.org/html/2610.02627#A1.SS0.SSS0.Px3.p1.1),[§2\.1\.1](https://arxiv.org/html/2610.02627#S2.SS1.SSS1.p1.1)\.
- Liet al\.\(2023\)C\. Li, J\. Wang, Y\. Zhang, K\. Zhu, W\. Hou, J\. Lian, F\. Luo, Q\. Yang, and X\. XieLarge language models understand and can be enhanced by emotional stimuli\.arXiv preprint arXiv:2307\.11760\.Cited by:[Appendix A](https://arxiv.org/html/2610.02627#A1.SS0.SSS0.Px4.p1.1),[§2\.1\.1](https://arxiv.org/html/2610.02627#S2.SS1.SSS1.p1.1)\.
- Liuet al\.\(2024\)N\. F\. Liu, K\. Lin, J\. Hewitt, A\. Paranjape, M\. Bevilacqua, F\. Petroni, and P\. LiangLost in the middle: how language models use long contexts\.Transactions of the association for computational linguistics12,pp\. 157–173\.Cited by:[Appendix A](https://arxiv.org/html/2610.02627#A1.SS0.SSS0.Px3.p1.1),[§2\.1\.1](https://arxiv.org/html/2610.02627#S2.SS1.SSS1.p1.1)\.
- Mannekoteet al\.\(2025\)A\. Mannekote, J\. Nam, Z\. Li, K\. Boyer, and B\. DorrMaking task\-oriented dialogue datasets more natural by synthetically generating indirect user requests\.InProceedings of the 31st International Conference on Computational Linguistics,pp\. 10449–10459\.Cited by:[Appendix A](https://arxiv.org/html/2610.02627#A1.SS0.SSS0.Px1.p1.1),[§2\.1\.1](https://arxiv.org/html/2610.02627#S2.SS1.SSS1.p1.1),[§5\.1](https://arxiv.org/html/2610.02627#S5.SS1.p1.1)\.
- Pavlick and Tetreault \(2016\)E\. Pavlick and J\. TetreaultAn empirical analysis of formality in online communication\.Transactions of the association for computational linguistics4,pp\. 61–74\.Cited by:[Appendix A](https://arxiv.org/html/2610.02627#A1.SS0.SSS0.Px2.p1.1),[§2\.1\.1](https://arxiv.org/html/2610.02627#S2.SS1.SSS1.p1.1)\.
- Penhaet al\.\(2022\)G\. Penha, A\. Câmara, and C\. HauffEvaluating the robustness of retrieval pipelines with query variation generators\.InEuropean conference on information retrieval,pp\. 397–412\.Cited by:[§1](https://arxiv.org/html/2610.02627#S1.p2.1),[§5\.4](https://arxiv.org/html/2610.02627#S5.SS4.p1.1)\.
- Rao and Tetreault \(2018\)S\. Rao and J\. TetreaultDear sir or madam, may i introduce the gyafc dataset: corpus, benchmarks and metrics for formality style transfer\.InProceedings of the 2018 conference of the north american chapter of the association for computational linguistics: human language technologies, volume 1 \(long papers\),pp\. 129–140\.Cited by:[Appendix A](https://arxiv.org/html/2610.02627#A1.SS0.SSS0.Px2.p1.1),[§2\.1\.1](https://arxiv.org/html/2610.02627#S2.SS1.SSS1.p1.1)\.
- Robertson and Zaragoza \(2009\)S\. Robertson and H\. ZaragozaThe probabilistic relevance framework: bm25 and beyond\.Foundations and trends® in information retrieval4\(1\-2\),pp\. 1–174\.Cited by:[§3\.2](https://arxiv.org/html/2610.02627#S3.SS2.SSS0.Px1.p1.1)\.
- Ryanet al\.\(2025\)M\. J\. Ryan, D\. Xu, C\. Nivera, and D\. CamposEnronqa: towards personalized rag over private documents\.arXiv preprint arXiv:2505\.00263\.Cited by:[§3\.1](https://arxiv.org/html/2610.02627#S3.SS1.SSS0.Px1.p1.1)\.
- Sapet al\.\(2019\)M\. Sap, D\. Card, S\. Gabriel, Y\. Choi, and N\. A\. SmithThe risk of racial bias in hate speech detection\.InProceedings of the 57th annual meeting of the association for computational linguistics,pp\. 1668–1678\.Cited by:[§1](https://arxiv.org/html/2610.02627#S1.p2.1),[§5\.3](https://arxiv.org/html/2610.02627#S5.SS3.p1.1)\.
- Searle \(1975\)J\. R\. SearleIndirect speech acts\.Syntax and semantics3,pp\. 59–82\.Cited by:[§5\.1](https://arxiv.org/html/2610.02627#S5.SS1.p1.1)\.
- Sharmaet al\.\(2024\)M\. Sharma, M\. Tong, T\. Korbak, D\. Duvenaud, A\. Askell, S\. Bowman, E\. Durmus, Z\. Hatfield\-Dodds, S\. Johnston, S\. Kravec,et al\.Towards understanding sycophancy in language models\.InInternational Conference on Learning Representations,Vol\.2024,pp\. 110–144\.Cited by:[Appendix A](https://arxiv.org/html/2610.02627#A1.SS0.SSS0.Px4.p1.1),[§2\.1\.1](https://arxiv.org/html/2610.02627#S2.SS1.SSS1.p1.1)\.
- Solidjonov \(2026\)D\. SolidjonovPragmatic competence without embodiment? evaluating llm performance on implicature, presupposition, and speech acts\.Journal of Cultural Cognitive Science,pp\. 1–13\.Cited by:[§5\.1](https://arxiv.org/html/2610.02627#S5.SS1.p1.1)\.
- Soni \(2026\)H\. SoniToolFailBench: diagnosing tool\-use failures in llm agents\.arXiv preprint arXiv:2607\.04686\.Cited by:[§1](https://arxiv.org/html/2610.02627#S1.p2.1),[§5\.2](https://arxiv.org/html/2610.02627#S5.SS2.p2.1)\.
- Sriraget al\.\(2025\)D\. Srirag, N\. R\. Sahoo, and A\. JoshiEvaluating dialect robustness of language models via conversation understanding\.InProceedings of the Second Workshop on Scaling Up Multilingual & Multi\-Cultural Evaluation,pp\. 24–38\.Cited by:[§1](https://arxiv.org/html/2610.02627#S1.p2.1),[§5\.3](https://arxiv.org/html/2610.02627#S5.SS3.p1.1)\.
- Styleset al\.\(2024\)O\. Styles, S\. Miller, P\. Cerda\-Mardini, T\. Guha, V\. Sanchez, and B\. VidgenWorkbench: a benchmark dataset for agents in a realistic workplace setting\.arXiv preprint arXiv:2405\.00823\.Cited by:[§3\.1](https://arxiv.org/html/2610.02627#S3.SS1.SSS0.Px3.p1.1)\.
- Yinet al\.\(2020\)F\. Yin, Q\. Long, T\. Meng, and K\. ChangOn the robustness of language encoders against grammatical errors\.InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics,pp\. 3386–3403\.Cited by:[§2\.1\.1](https://arxiv.org/html/2610.02627#S2.SS1.SSS1.p1.1)\.
- Zhanget al\.\(2025\)J\. Zhang, S\. Yu, D\. Chong, A\. Sicilia, M\. R\. Tomz, C\. D\. Manning, and W\. ShiVerbalized sampling: how to mitigate mode collapse and unlock llm diversity\.arXiv preprint arXiv:2510\.01171\.Cited by:[§2\.2](https://arxiv.org/html/2610.02627#S2.SS2.SSS0.Px1.p1.1)\.
- Ziemset al\.\(2022\)C\. Ziems, J\. Chen, C\. Harris, J\. Anderson, and D\. YangVALUE: understanding dialect disparity in nlu\.InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 3701–3720\.Cited by:[§1](https://arxiv.org/html/2610.02627#S1.p2.1),[§5\.3](https://arxiv.org/html/2610.02627#S5.SS3.p1.1)\.
- Ziemset al\.\(2023\)C\. Ziems, W\. Held, J\. Yang, J\. Dhamala, R\. Gupta, and D\. YangMulti\-value: a framework for cross\-dialectal english nlp\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 744–768\.Cited by:[§1](https://arxiv.org/html/2610.02627#S1.p2.1),[§2\.2](https://arxiv.org/html/2610.02627#S2.SS2.SSS0.Px2.p1.1),[§5\.3](https://arxiv.org/html/2610.02627#S5.SS3.p1.1)\.

## Appendix ACommunication\-Style Axes in Detail

The five axes summarized in Table[1](https://arxiv.org/html/2610.02627#S2.T1)are described here in full\.

##### Directness\.

This axis separates an explicit request from one whose information need must be inferred\. At the direct pole the request is stated as a question or command; at the indirect pole it is conveyed through a need, hint, observation, or expression of uncertainty\. The contrast follows the distinction between on\-record and off\-record requests in[Brown and Levinson \[1987\]](https://arxiv.org/html/2610.02627#bib.bib2)and[Mannekote et al\. \[2025\]](https://arxiv.org/html/2610.02627#bib.bib3)\.

##### Formality\.

This axis separates a polished professional register from a casual, chat\-like one, following the register variation studied in[Pavlick and Tetreault \[2016\]](https://arxiv.org/html/2610.02627#bib.bib4)and[Rao and Tetreault \[2018\]](https://arxiv.org/html/2610.02627#bib.bib5)\. Formality has been reported as a source of variation in retrieval\-augmented question answering in[Cao et al\. \[2026\]](https://arxiv.org/html/2610.02627#bib.bib6)\.

##### Verbosity\.

This axis separates a concise request from the same request embedded in additional text\. The added material is generic chatter, hedging, or digression, and introduces no new fact about the task, the evidence, or the expected answer\. Prior work suggests that input length and the position of the relevant information affect model behavior\[[Levy et al\., 2024](https://arxiv.org/html/2610.02627#bib.bib7),[Liu et al\., 2024](https://arxiv.org/html/2610.02627#bib.bib8)\]\.

##### Emotion\.

This axis separates neutral wording from wording that expresses frustration, stress, anxiety, or urgency\. Emotional content in a prompt is known to change model behavior on an otherwise fixed task\[[Li et al\., 2023](https://arxiv.org/html/2610.02627#bib.bib9)\], and models have been shown to revise responses under user pressure\[[Sharma et al\., 2024](https://arxiv.org/html/2610.02627#bib.bib10)\]\. The axis tests whether that occurs when the task, evidence, and expected outcome are held constant\.

##### Grammaticality\.

This axis separates standard written English from understandable English carrying systematic grammatical deviation, including the topic or object fronted, the question word moved with the auxiliary dropped, the verb flattened with an adverb carrying the time, or clauses split and reordered\. Systematic slips sit on top of that restructuring, including dropped or wrong articles, missing plural marking, and tense or agreement errors\. The pole is defined by properties of the text alone and is not conditioned on any writer attribute\.

## Appendix BEnronQA Sampling Design

EnronQA is the only benchmark in the study for which we choose the items, so the choice is described in full\. The goal was a sample on which question complexity could be read as an effect of one property at a time rather than of overall difficulty\.

##### Classification\.

Three complexity properties are assigned by a rule\-based classifier over the question text alone: reasoning depth \(1\-hop,2\-hop,3\+\-hop\), scored from the presence of a context prefix, a relative clause, a nested reference to something mentioned elsewhere, and two or more clause connectives; question multiplicity \(single\-question,multi\-question\), triggered by coordinating*and*,*as well as*or*along with*that introduce a new wh\-clause; and context dependency \(self\-contained,thread\-dependent\), triggered by explicit reference to an earlier or prior message\. Broad threading markers such as*forwarded*or*re:*were deliberately excluded, since a message in a thread can still be self\-answerable\. A fourth label, reasoning type, is assigned from the question and gold answer under a fixed priority order: causal, comparative, aggregation, temporal, bridging, and otherwise extraction\.

##### Isolation\.

Each bin of100100requires exactly one of the three properties at its hard setting and the other two at the easy setting\. The*simple*control is at most two hops, single\-question and self\-contained;*3\-hop*raises depth only;*multi\-question*raises multiplicity only;*threaded*raises context dependency only\. The final selection asserts that the number of hard properties is00for every simple item and exactly11for every other item\. A difference between the control and one complex bin is therefore a difference in that one property\.

##### Selection\.

The pool is the EnronQA train split restricted to users with at least2525emails, deduplicated on question text and thinned to at most one question per email\. Within each bin, items are drawn stratified by reasoning type so that every type reaches at least3030items across the400400, with the residual filled to100100per bin\. Two filters run during selection\. A near\-duplicate guard rejects a candidate whose embedding has cosine similarity≥0\.80\\geq 0\.80to any question already kept\. A content screen runs the question through an oracle answerer and drops it if the provider blocks or the model refuses\.

##### Composition\.

The result is400400questions over400400distinct emails and9999users,100100per bin\. Reasoning types are distributed as extraction126126, bridging7474, causal6161, temporal5757, aggregation4545and comparative3737\. The distribution tries to replicate the distribution in the whole dataset\. Retrieval runs over the target user’s own inbox, which ranges from4141to10231023emails with a median of273273\.

## Appendix CVerbalized Sampling

A single rewrite request per item produces low\-diversity output: the model converges on one stock opener and reuses it across items, so the resulting variant set measures one phrasing rather than a style\. We therefore generate each contrastive pole with verbalized sampling, in which one call returnskkcandidates and each candidate is assigned a different framing drawn from a per\-pole menu\.

##### Procedure\.

For an item with identifierqqand a pole with menuMM, we requestk=6k=6candidates in a single call\. The candidates are returned as a JSON array\. The six rewrites restructure the sentence six different ways rather than repeating one edit\. The rotation is a deterministic function ofqq, seeded by an MD5 digest, so runs are reproducible and order\-independent\.

##### Selection\.

Candidates are first filtered by the deterministic checks of Appendix[E](https://arxiv.org/html/2610.02627#A5)\. The surviving candidates are then scored by the novelty of their opening words, measured as the number of times that opener has already been kept during the run, and the least\-used opener wins\. Ties are broken by a per\-item seeded permutation of the candidate order, which prevents a fixed menu position from being systematically preferred whenever several candidates score equally\. The winning candidate is then passed to the judges\. If none passes, the call is repeated, twice for most poles and three times for the charged pole\.

##### Prompt template\.

The user message is constructed as shown in Appendix[D\.19](https://arxiv.org/html/2610.02627#A4.SS19), wherekis the candidate count\.

## Appendix DCommunication\-Style Prompts

Each contrastive pole has one system prompt and one menu of framings listed below\. The four reference poles use the plain rewrite prompt of Appendix[D\.13](https://arxiv.org/html/2610.02627#A4.SS13)with the per\-pole instruction of Appendix[D\.14](https://arxiv.org/html/2610.02627#A4.SS14)\.The rewrite instruction is given to the generator, and the fidelity criterion is supplied to Gate B, the style\-realization gate\.

### D\.1System Prompt:indirect

Yourewriteauser'sQUESTIONaboutanemailintoanINDIRECT,off\-recordrequest,changingONLYitsdirectness\-\-\-neveritsmeaning\.

Hardrules:

\-PreservetheinformationneedEXACTLY:areadermustrecovertheSAMEsingleanswer,anditmustbeanswerablebythatsameanswer\.Keepeveryspecificdetail\(names,dates,numbers,quotedphrases,thesubject\)\.AddnonewfactanddoNOTanswerit\.

\-Begenuinelyindirect:implytherequestviathegivenmechanisminsteadofablunton\-record'what/who/when'questionorcommand\.Theoutputneednotbeaquestion\.

\-Stayspecific:thereadermuststillknowexactlywhichdetailisaskedabout\.

MakeitreadlikeaREALpersonactuallytypingthis,notanassistant:

\-Firstperson,naturalcontractions\(I'm,can't,don't,it's,there's\)\.

\-Plain,concrete,everydaywords\-\-\-NOTformal,elaborate,orcorporatevocabulary,andnonominalisations\('clarificationissought'\)\.

\-Keepitshortandalittleuneven;atrailingthoughtorasentencefragmentisfine\.

\-Dropassistantscaffolding:no'Ihopethishelps','kindly','Iwouldbegrateful',noexhaustivehedging,nostiffregister\.

\-Oneclearask;don'tover\-explainorpileonqualifiers\.

OutputONLYtherewrittentext:noquotes,labels,orcommentary\.

### D\.2Indirectness Mechanisms

1\.PREPARATORY\(hearer'sknowledge\):implytheaskbywonderingwhetherthereaderhappenstoknoworcouldtellyouthespecificthing\-\-notabare'what/who/when'question\.

2\.DESIRE\(speaker'swant\):statethatyouwouldlikeorwanttoknowthespecificthing,asawishratherthanaquestion\.

3\.REASON/MOTIVATION:givethereasonorgoalthatmakesthespecificanswernecessary,andlettherequestfollowfromitratherthanaskingoutright\.

4\.UNMETPRECONDITION:notethatyoustilllackthespecificpieceofinformation\(youdon'thaveit,orcan'tfindwhereitisstated\),sotheneedforitisimplied\.

5\.EPISTEMICUNCERTAINTY:sayyoucan'ttellorworkthespecificansweroutfromtheemail,leavingthequestionimplied\.

6\.ASSOCIATIVEHINT:nametheexactdetailintheemailtheanswerhangsonandflagthatitsmeaningorreferenceisunclear,sothereaderinferstheask\.

7\.UNDERSTATEMENT:betentativeandunder\-specify\-\-sayyou'realittlefuzzyorhazyonthespecificpoint,withoutnamingitasaquestion\.

8\.PRESUPPOSITION:presupposethespecificanswerexistsintheemailandcommentarounditratherthanrequestingitoutright\.

9\.HYPOTHETICAL/CONDITIONAL:embeddingtherequestinsidethecondition\.

10\.REPORTED/THIRD\-PARTY:frameitassomeoneelsehavingaskedyouthisandyoubeingunabletoansweritfromtheemail\.

11\.GAPOBSERVATION:pointoutthattheemailleavesthespecificpointunspecifiedoropen,implyingyouwantitresolved\.

12\.HEDGEDAPOLOGY:prefacewithahedgeorapologyforpossiblymissingit,thengestureatthespecificpoint\.

### D\.3System Prompt:formal

Yourewriteauser'sQUESTIONaboutanemailintoaformal,professionalregister\-\-\-thewayareal,competentprofessionalactuallywritesacourteousworkemail\-\-\-changingONLYtheregister,neverthemeaning\.

Hardrules:

\-Formalandpolite:completesentences,standardcapitalisationandpunctuation,nocontractions\(write'Iwould','couldyou',not'I'd','you're'\)andnoslangorabbreviations\.

\-ButsoundlikeaREALPERSON,notaninstitution\.AVOIDroboticofficialese:nonominalisationpile\-ups\('clarificationissoughtregarding'\),no'theundersigned',no'kindlyascertain',noarchaicorbureaucraticformulas,nostiffpassivethroat\-clearing\.

\-Anaturalprofessionalvoice:'Couldyoupleaseconfirm…','Iwantedtocheck…','Wouldyoubeabletoletmeknow…','Iamtryingtoconfirm…'\.

\-Oneclear,courteousask\.Keeptheexactsameaskandeveryspecificdetail\(names,dates,numbers,quotedphrases,thesubject\);addnonewfactanddoNOTanswerit\.

Varythecourteousopeningfromitemtoitemratherthanreusingonestockformula\.OutputONLYtherewrittenquestion:noquotes,labels,orcommentary\.

### D\.4Formal Framings

1\.acourteous,directrequest\-\-\-'Couldyoupleaseconfirm…'/'Iwantedtocheck…'

2\.apoliterequestthatbrieflyreferencestheemail,thenasks

3\.aprofessionalwantingtoverifyonespecificdetailbeforeproceeding

4\.ameasured,respectfulrequestforclarification

5\.awarmbutbusinesslikerequestfromacolleague

6\.aprofessionalrestatingwhattheyneed,tobesuretheyhaveitright

7\.abrief,courteousrequestthatgetstothepointpolitely

8\.aconsideraterequestthatacknowledgesthereader'stime

### D\.5System Prompt:casual

Yourewriteauser'sQUESTIONaboutanemailasacasual,chat\-likemessage\-\-\-thewayyou'dquicklytextorSlackacoworkeryou'refriendlywith\-\-\-changingONLYtheregister,neverthemeaning\.

Hardrules:

\-WriteEVERYTHINGinlowercase\-\-\-nocapitallettersanywhere,notevenatthestartofamessage,on'i',oronnames\(jim,socal,theqcomdeal\);thatall\-lowercaselookisjusthowpeopledashoffaquicktext\.

\-Makeitgenuinelyinformal,notaformalsentencewitha'hey'boltedonthefront\.Actuallyrewriteitintohowyou'dtext:contractionsandcommonshortforms\(u,ur,pls,thx,lmk,btw,tbh,rn,asap,idk,fyi\),everydaychattywordingandlightslang\(gonna,wanna,kinda,gotta,ngl,lowkey,realquick,anychance\),looseordroppedpunctuation,and\-\-\-whereitfits\-\-\-anemojiortwo\.Leanintothecolloquialvoice:useseveralofthesetouchespermessageandvarythemacrossrewrites,butkeepitreadableandhuman\-\-\-notakeyboard\-smashofslang\.

\-ByDEFAULTstripthestiff,contentlessscaffoldingatexterwouldn'tbothertyping\-\-\-'accordingtotheemail','asmentionedintheemail','intheemailchain','respectively','peryourmessage'\-\-\-insteadofcarryingitoverfromtheoriginal\.Especiallydropabare'accordingtotheemail'wrapper;keeponlyarealdisambiguatorthatridesalongwithit\(atopiclike'aboutoptionsettingsinastructurednote',aperson,oradocumentname\)\.It'sthesameask,juststrippedoftheformalpackaging\.

\-ButKEEPeveryrealdetail:names,dates,numbers,quotedtitlesorphrases,andexactlywhat'sbeingasked\.Ifanattributionnamesaspecificpersonordocumentthatpinsdownwhichanswerismeant,keepthatname\.AddnonewfactanddoNOTanswerit\.

MakeitreadlikeaREALpersonactuallytypingthis,notanassistant:

\-Firstperson,naturalcontractions\(I'm,can't,don't,it's,there's\)\.

\-Plain,concrete,everydaywords\-\-\-NOTformal,elaborate,orcorporatevocabulary,andnonominalisations\('clarificationissought'\)\.

\-Keepitshortandalittleuneven;atrailingthoughtorasentencefragmentisfine\.

\-Dropassistantscaffolding:no'Ihopethishelps','kindly','Iwouldbegrateful',noexhaustivehedging,nostiffregister\.

\-Oneclearask;don'tover\-explainorpileonqualifiers\.

Varytheopening\-\-\-sometimesaquick'hey',sometimesjusttheaskwithnogreetingatall;don'treuseonestockopener\.OutputONLYtherewrittenquestion:noquotes,labels,orcommentary\.

### D\.6Casual Framings

1\.aquickslackpingtoaworkfriend\-\-\-lowercase,acontractionortwo

2\.atextdashedoffbetweenmeetings\-\-\-nogreeting,straighttotheask,loosepunctuation

3\.achattymessageusingashortformortwo\(lmk,u,pls\)andmaybeoneemoji

4\.abreezy'srytobugu'askthattrimstheformalbits

5\.anoffhandone\-liner,righttothepoint,noopener

6\.arelaxedpingliketextingateammateyougetalongwith

7\.aslightlyrushed'realquick'askwithacontractionortwo

8\.afriendlynotethatdropsthestiff'accordingtotheemail'framingandjustasks

### D\.7System Prompt:verbose

Yourewriteauser'sQUESTIONaboutanemailasalong,ramblingmessage\-\-\-thewayareal,chattypersonover\-writes\-\-\-changingONLYthelengthandpackaging,neverthemeaning\.

Hardrules:

\-MakeitclearlyLONGandwandering,andBURYtheactualquestionintheMIDDLE\-\-\-neverstatetheaskcleanlyupfrontorasatidyfinalline\.

\-Paditlikearealpersonrambles:aconcreteaside,thinkingoutloud,aspecifictangent,anafterthought,arun\-on\.NOTwithabstractself\-commentaryaboutbeingthorough/precise/careful,andNOTbystackinghedgesandqualifiers\-\-\-thatreadslikeanAI\.

\-ThepaddingmustbeCONTENT\-FREEabouttheemail:rambleaboutyourself,yourday,yourworkload,yourprocess\-\-\-neverinventanynewfact,detail,name,number,orclaimabouttheemailitself\.

\-Keeptheexactsameaskandeveryspecificdetail\(names,dates,numbers,quotedphrases,thesubject\);doNOTanswerit\.

Writeinfirstperson,naturalcontractions,plainconcretewords\.Varytheramblingand\-\-\-importantly\-\-\-theOPENINGWORDSfromitemtoitem:doNOTkeepstartingwith'Ikeep','I'vebeen',orthesamefirstfewwords;leadsometimeswiththeaside,sometimesmid\-thought,sometimeswithacomplaintoranapology\.OutputONLYtherewrittenmessage:noquotes,labels,orcommentary\.

### D\.8Verbose Framings

1\.openwithaconcrete,unrelatedasideaboutyourdayorworkload\(ameetingyoujustleft,anoverflowinginbox,needingcoffee\)beforedriftingtowardtheask

2\.openbyapologisingfortheramblymessage,thenwanderthroughanafterthoughtortwobeforetheasklandsmid\-paragraph

3\.gooffononespecifictangentaboutwhyyouneedthisrightnow,thensliptheaskinpartwaythrough

4\.startmid\-thoughtasifcontinuinganearlierconversation,backtrackandcorrectyourself,thentucktheaskinamongtheafterthoughts

5\.mildlycomplainabouthowmuchemailorpaperworkyou'rewadingthroughtoday,thengettothequestioninthemiddleofit

6\.letseveralloosely\-relatedthoughtspileupinonelongrun\-onsotheasksitssomewhereinthemiddle

7\.openwithasmallworryaboutgettingadetailwrong,ramblearoundit,andburytheaskinsidethereassuranceyougiveyourself

8\.narrateyourselfhuntingaroundforthis\-\-\-scrolling,losingyourplace\-\-\-beforeandafteryoufinallyask

### D\.9System Prompt:charged

YouAREarealpersonundergenuinestress,dashingoffaquickquestionaboutanemail\.Yourexactsituationisthepersonayou'regiven\.WritethequestionthewayTHATperson,inTHATmoment,wouldactuallytypeit\-\-\-changingONLYtheemotionaltone,neverthemeaning\.

Hardrules:

\-Stayincharacterandletthepressureofyoursituationbleedintothewholemessage\-\-\-wordchoice,rhythm,whatyoufixateon\-\-\-sothefrustration,stress,anxiety,orurgencyisUNMISTAKABLE\.ItmustNOTreadasacalm,neutralquestion\.

\-WeavetheemotionINTOtheaskitself,andRETYPEthewholequestioninyourownrushedwords\-\-keepeveryname,dateandnumber,butthewordingaroundthemmustbeYOURS,nottheoriginalformalsentencepastedback\.DoNOTopenwithadetachedreactionandthenpastethetidyquestionafterit\(NOT'I'maskingagain\-\-whattwoprograms\.\.\.',NOT'Beforethislandsonme,what'sthecontext\.\.\.'\),anddoNOTtackanemotionalbolt\-on\('now\.','ASAP\!\!'\)ontoanotherwiseneutralsentence\-\-thestrainhastorunTHROUGHhowyouask,fromthefirstwordtothelast\.

\-Showitthewaypeopleactuallytypeunderpressure,don'tannounceit:terse,clippedphrasing,fixatingontheonedetail,anexasperatedorworriedaside,apointedjab,andATMOSToneexclamationmarkoronecapitalisedword\.DoNOTstackintensifiers\('reallysoincrediblyurgent\!\!\!'\),doNOTopenwithastocklabel\('URGENT:','I'msofrustrated:','Ugh,'\),anddoNOTsoundlikeanassistant\.

\-Yourstressisyourownstateofmind,notanewfactabouttheemail\.Keeptheexactsameaskandaddnonewfactualdetail;theemotionalwordingdoesNOTcountasanewfact\.

MakeitreadlikeaREALpersonactuallytypingthis,notanassistant:

\-Firstperson,naturalcontractions\(I'm,can't,don't,it's,there's\)\.

\-Plain,concrete,everydaywords\-\-\-NOTformal,elaborate,orcorporatevocabulary,andnonominalisations\('clarificationissought'\)\.

\-Keepitshortandalittleuneven;atrailingthoughtorasentencefragmentisfine\.

\-Dropassistantscaffolding:no'Ihopethishelps','kindly','Iwouldbegrateful',noexhaustivehedging,nostiffregister\.

\-Oneclearask;don'tover\-explainorpileonqualifiers\.

OutputONLYtherewrittenquestion:noquotes,labels,orcommentary\.

### D\.10Emotional\-Charge Situations

1\.you'vebeenheads\-downfor12\+hoursandyou'reexhaustedandoutofpatience\-\-\-youjustwantthisonedetailsoyoucanfinallylogoff

2\.you'reminutesfromaharddeadlineandstillmissingthis\-\-\-notimeforniceties,youneeditrightnow

3\.something'sblowingupandyourbossisonyourbackforthisexactdetail;you'reanxiousandscrambling

4\.you'llbetheoneblamedifthisdetailiswrong,andthat'sstressingyouout\-\-\-youneedtobesurebeforeitlandsonyou

5\.youalreadyaskedaboutthisonce\(maybetwice\)andheardnothingback;you'refedupandlosingpatience

6\.you'redrowningintoomanythingsatonceandfrazzled\-\-\-youcanbarelysparetheseconds,soitcomesoutsharpandclipped

7\.aclientoranexeciswaitingonyouandbreathingdownyourneck;thepressureismakingyoutenseandurgent

8\.it'slate,you'reworndownandabitdesperate,quietlydreadingwhathappensifyoucan'tpinthisdetaildown

### D\.11System Prompt:non\-grammatical

YouarewritingEnglishasasecondlanguage:fluentenoughtobeunderstood,butyouBUILDTHESENTENCEwithnon\-nativewordorderandmakereal,systematiclearnermistakes\.WritetherequestthewayYOUwouldactuallytypeitinahurry,fromscratch\-\-\-NOTacorrectEnglishsentencewithacoupleofmistakessprinkledontop\.

Hardrules:

\-RESTRUCTURETHESENTENCE\-\-\-thisisthemostimportantrule\.ThewordorderandclausestructureMUSTvisiblychangefromtheoriginal;doNOTleavetheoriginalsentenceintact\.Usenon\-nativestructuralmovessuchas:frontthetopicorobject\('thisreport,whenyousend?','thetwoaddresses,whattheyare?'\);putthequestionwordinthemiddleorattheENDinsteadofthefront\('theconfirmationcomebywhattwomethod?','theaddressesarewhat?'\);dropthequestionauxiliarydo/does/did/will\('whyhenottellme?','whentheyarrive?'\);flattenverbstobaseformandletanadverbcarrythetime\('yesterdayhesendit'\);orreorder/split/runtogethertheclauses\.AtleastONEclearstructuralchangeisREQUIREDineveryrewrite\.

\-ONTOPofthereorder,addafewsystematicgrammarandspellingslips:droppedorwrongarticles\('a/the'\),wrongormissingplural\-s,wrongverbtenseoragreement,amissingorextraauxiliary,adroppedcopula\('is/are'\),wrongprepositions\('dependof','discussabout'\),andoccasionalmisspellingsofORDINARYwords\('recieve','adress','seperate','there/their'\)\.ThesearetheSEASONING;therestructuredwordorderisthemaindish\-\-\-asentencethatkeepsnativewordorderandonlydropsanarticleoraddsatypodoesNOTcount\.

\-BeUNEVEN,likeareallearner:afewerrors,somepartsfluent\.DoNOTstackeverypossibleerroratonce\-\-\-itmuststayreadableandhuman,NOTgibberishormachine\-translationnoise\.

\-DoNOTcheatbyleavingtheoriginalquestionperfectandboltinganextraawkwardsentenceontotheend\-\-\-theCOREquestionitselfmustberestructured\.

\-DoNOTuseanyforeigngreeting,discourseparticle,honorific,accent,orcode\-switchedword\(no'Bonjour','Namaste','na?','\-san','ji',andthelike\)\.ThisisplainEnglishwithnon\-nativeSTRUCTUREanderrors,nothingculturalboltedon\.

\-Keepeveryname,number,date,quotedtitle,andspecificdetailEXACTLYcorrect\(spellingerrorsareONLYforordinarywords\),keeptheexactsameask\(answerablebythesameanswer\),andaddnonewfactabouttheemail\.

Writeinfirstperson,plainandshort\.DoNOTbeginwithanylabel\(never'Non\-native:','ESL:','\[L2\]'orsimilar\)\-\-\-startdirectlywithyourmessage\.Whenseveralrewritesarerequested,VARYthestructure\-\-\-eachrewritereordersthesentenceaDIFFERENTwayandleansonadifferentmixofmistakes,sothesetdoesnotreadasonetemplate\.OutputONLYtherewrittenrequest:noquotes,labels,orcommentary\.

### D\.12Non\-Grammatical Structural Profiles

1\.Topic\-fronting\-\-\-movetheobjectortopictotheFRONTandcommentafterit,droppingarticlesand'is/are'\('thisreport,alreadysent?','thetwoaddresses,whattheyare?'\)\.Leadwiththereordering;adroppedarticleorpluralmaytagalong\.

2\.Question\-inversion\-\-\-breakthenormalquestionorder:putthewh\-wordinthemiddleorattheENDanddropthedo/does/did/willauxiliary\('LarryCampbellreceiveconfirmationbywhattwomethod?','thetimingandstartdateiswhat?','whynotyoutoldme?'\)\.Leadwiththescrambledquestionstructure\.

3\.Word\-orderandtense\-\-\-reordersubject/verb/objectandusethewrongorbase\-formverbwiththetimesetbyanadverb,nottheverb\('yesterdayhesendthereport','nextweekshecallyou,whyshenotsay?'\)\.Leadwiththeword\-ordershiftplustenseslips\.

4\.Clause\-reorder\-\-\-swaporsplitthemainandsubordinateclauses,movethe'accordingto\.\.\.'or'intheemail\.\.\.'parttoanoddspot,orfuseclausestogether\('intheemailabouttheforecast,fiveissuesAA&Coreview,thetreatmentiswhat?'\)\.Leadwithclauserestructuring;aprepositionslipmaycomealong\.

5\.Fragmentandrun\-on\-\-\-chopthesentenceintoshortchoppyfragmentsorfuseitintoonelongrun\-on,misplaceamodifier,anddroptheconnectingwords\('thepostdate,thetime,forAllShippersLetter\#01\-317\-\-\-whatis?'\)\.Leadwiththebrokenflow\.

6\.Spelling\-and\-rebuild\-\-\-rebuildthesentenceinaDIFFERENTwordorderANDmisspellordinarywordsormixhomophones\('recieve','adress','seperate','there/their'\)\-\-\-butkeepeveryname,place,date,number,andquotedtitlespelledexactlyright\.Thereorderisrequired;themisspellingssitontop\.

### D\.13Reference\-Pole Rewrite System Prompt

Yourewriteauser'sQUESTIONaboutanemailtochangeONLYitscommunicationSTYLE,neveritsmeaning\.

Hardrules:

\-PreservetheintentEXACTLY:therewrittenquestionmustrequestthesameinformationandbeanswerablebythesameanswerastheoriginal\.

\-Keepeveryspecificdetailfromtheoriginal:names,dates,numbers,quotedphrases,andthesubjectofthequestion\.

\-DoNOTaddanynewquestion,requirement,fact,orassumption,anddoNOTanswerthequestion\.

\-ChangeONLYthetargetstyledescribedbytheuser\.Neverdescribeoraltertheemailitself\.

\-OutputONLYtherewrittenquestiontext\-\-\-noquotes,nolabels,nocommentary\.

### D\.14Reference\-Pole Rewrite Instructions

Given to the generator for the four reference poles, which are produced by a single plain rewrite without verbalized sampling\. The contrastive poles are generated from the pole\-specific system prompts above instead\.

\[directness/direct\]

rewrite\_instruction:Rewriteasafullyexplicit,on\-recordrequest:stateexactlywhatisbeingasked,upfront,asacleardirectquestionorimperative\.

\[verbosity/concise\]

rewrite\_instruction:Rewriteasterselyaspossible:puttheaskfirst,inoneshortsentence,withnoextracontext\.

\[emotional\_charge/calm\]

rewrite\_instruction:Rewritewithcalm,matter\-of\-fact,neutralaffectandnoemotionallanguage\.

\[native\_likeness/native\]

rewrite\_instruction:Rewriteinfluent,idiomaticnative\-Englishphrasing:naturalwordorder,correctarticlesandprepositions,andnaturalcollocations\.Keeptheexactsameaskandaddnonewfacts\.

### D\.15Gate A: Intent Preservation

Youcomparetwoversionsofaquestionaskedaboutanemail\.

DecidewhethertheVARIANTasksforEXACTLYthesameinformationastheORIGINAL\-\-\-i\.e\.anyanswerthatcorrectlyanswersonewouldcorrectlyanswertheother\.

Ignoredifferencesintone,politeness,directness,length,orwording;judgeONLYtheinformationrequested\.

Respondwithexactlyoneline:'Verdict:yes'\(sameinformation\)or'Verdict:no'\(differentinformation\)\.

### D\.16Gate B: Style Realization

Youcheckwhetherarewrittenquestionexhibitsaspecifiedcommunicationstyle\.

Judgeonlythestyle,notthecorrectnessormeaningofthequestion\.

Respondwithexactlyoneline:'Verdict:yes'\(clearlyexhibitsthestyle\)or'Verdict:no'\(doesnot\)\.

### D\.17Fidelity Criteria for All Ten Poles

Supplied to Gate B as the definition of the target pole\.native\_likenessis the grammaticality axis, andnativeandnon\_nativeare its two poles\.

\[directness/direct\]

fidelity\_criterion:Therequestisexplicitandon\-record\-\-\-areaderimmediatelyseesexactlywhatisbeingasked,asadirectquestionorcommand\.

\[directness/indirect\]

fidelity\_criterion:Therequestisindirect\-\-\-hedged,hinted,orphrasedasawonderingorastatedneed\-\-\-ratherthanablunt,bare,on\-recordquestionorcommand\.Conventionally\-indirectforms\('Iwaswonderingwhat…','I'mtryingtofigureout…','Doyouhappentoknow…'\)andoff\-recordhintsbothcount;aplaindirectquestion\('WhatisX?'\)oranimperativedoesNOT\.

\[formality/formal\]

fidelity\_criterion:Thetextreadsasaformal/professionalregister\-\-\-fullwords\(nocontractionsorslang\)withstandardcapitalisationandpunctuation\.

\[formality/casual\]

fidelity\_criterion:Thetextreadsascasual/chat\-like/informal\(contractions,slang,orrelaxedpunctuation\),whiletheunderlyingaskisunchanged\.

\[verbosity/concise\]

fidelity\_criterion:Thetextisshortandtheaskisstatedupfront\.

\[verbosity/verbose\]

fidelity\_criterion:Thetextislongandtheaskisburiedinthemiddleratherthanstatedupfront,andnonewrequestorfactualdetailhasbeenadded\.

\[emotional\_charge/calm\]

fidelity\_criterion:Thetoneiscalmandneutral,withnostrongemotionallanguage\.

\[emotional\_charge/charged\]

fidelity\_criterion:Thetextclearlyconveysfrustration,anxiety,stress,orurgency\(notjustaplain,neutralquestion\),whiletheunderlyingaskisunchanged\.

\[native\_likeness/native\]

fidelity\_criterion:Thetextreadsasfluent,idiomaticnativeEnglish\-\-\-naturalgrammar,articles,prepositions,andwordchoice,withnolearner\-likeerrors\.

\[native\_likeness/non\_native\]

fidelity\_criterion:Thetextreadsasnon\-native/second\-language\(L2\)EnglishratherthanfluentidiomaticnativeEnglish\.Itmaycarryoneormorelearner\-likeerrors\(article,preposition,agreement,verb\-tense,orword\-orderslips\)orsimplyreadasawkwardorunidiomatic;itneednotbeheavilybrokentocount\.

### D\.18Gate C: Human\-Likeness

YoujudgewhetherashortmessagereadsasifaREALPERSONgenuinelywroteit,orreadsasARTIFICIAL\-\-\-machine\-generated,templated,orperformative\.

ThemessageisaQUESTIONaboutanemail,rephrasedinaparticularstyle\.ItmaylegitimatelybeLONGandDETAILEDandnamespecificpeople,dates,addresses,orfigures\-\-\-length,detail,andformalityareNOTsignsofartificiality,soneverpenalizeamessagejustforbeinglongorthorough\.

Itmaybeinanyregister\(formal,casual,emotional,ornon\-native/learnerEnglish\)andstillbehuman:judgeauthenticityandnaturalness,NOTfluency,correctness,orpoliteness\.Anon\-nativewriter'scareful,mostly\-fluentquestioncarryingafewunevenslips\(missingarticles,agreement,tense,orwordorder\)IShuman\-\-\-itneednotbeheavilyorevenlybroken\.

Markitartificial\('no'\)onlyifitreadsasmachine\-produced:itstacksunnaturalintensifiers,usesroboticortemplatedphrasingormechanicaleye\-dialect,orboltsaforeigngreeting,honorific,ortagontootherwise\-perfectnativeEnglishasdecorationratherthanagenuinechangeinhowthesentenceisbuilt\.

Respondwithexactlyoneline:'Verdict:yes'\(readsasgenuinelyhuman\)or'Verdict:no'\(readsasartificial/AI\-like\)\.

### D\.19Verbalized\-Sampling User Template

Produce\{k\}DIFFERENTrewritesofthequestionbelow,eachadifferent\{noun\}drawnfromthislist:

\{menu\}

Makethe\{k\}rewritesasdifferentfromoneanotheraspossible\(differentopeningsandsentenceshapes\)whileeachkeepstheSAMEsingleanswerandeveryspecificdetail\.Eachmustfollowalltherulesinthesystemmessage\.

Writeonlytherewrittenquestionitselfforeach\-\-neverprefixarewritewithits\{noun\}name,number,oranylabel\.

Originalquestion:

\{question\}

ReturnONLYaJSONarrayofexactly\{k\}strings,e\.g\.\["\.\.\.","\.\.\."\]\.Noothertext\.

## Appendix EValidation Gates

The candidate rewrite passes through a three\-layer gate before it is admitted\. Overall gate\-passing rate per dataset is in Table[4](https://arxiv.org/html/2610.02627#A5.T4)and per\-pole gate\-passing rate in Table[5](https://arxiv.org/html/2610.02627#A5.T5)\.

##### Layer 1: deterministic checks\.

A candidate is rejected if it is empty, if it is byte\-identical to the original after normalization, or if any numeric token present in the original is absent from the variant\. Missing capitalized tokens are recorded as a soft signal but do not reject, since names may legitimately be rephrased\.

##### Layer 2: rubric judges\.

Gate A \(Appendix[D\.15](https://arxiv.org/html/2610.02627#A4.SS15)\) asks whether the variant requests exactly the same information\. Gate B \(Appendix[D\.16](https://arxiv.org/html/2610.02627#A4.SS16)\) asks whether the variant realizes the target pole, and is given that pole’s fidelity criterion from Appendix[D\.17](https://arxiv.org/html/2610.02627#A4.SS17)\. Gate C \(Appendix[D\.18](https://arxiv.org/html/2610.02627#A4.SS18)\) asks whether the text reads as something a person would genuinely type\. All three answer on a single line so the verdict can be parsed deterministically\.

##### Gate C exceptions\.

Gate C is not applied to the verbose pole, where a long message wrapped around a one\-line question is inherently machine\-like and shortening it would remove the property under test; that pole is admitted on Gates A and B alone\. On the non\-grammatical pole Gate C is computed and recorded but does not filter, because blocking on it would bias selection toward the mildest, least restructured rewrites\.

##### One low\-gate\-passing cell\.

Yield is above 80% for 29 of the 30 dataset\-pole cells\. WorkBenchindirectadmits 64 of 90, with the loss concentrated at Gate A: WorkBench instructions name a specific record to modify, and an indirect rewrite that hides the operation often also loses the record identifier, at which point the variant is no longer holding the same question intent as the original query\.

Table 4:Overall gate\-passing rate per dataset\. A slot is one \(item, pole\) pair; on ToolTalk a slot is one \(user turn, pole\) pair\.Table 5:Per\-pole gate\-passing rates, in percent of slots\.Ais intent preservation,Bis style realization, andCis human\-likeness\. Slots per pole are 400 \(EnronQA\), 50 \(ToolTalk\), 90 \(WorkBench\)\.*Role*is reference \(ref\) or contrastive \(con\)\. Gate C is evaluated only on five contrastive poles: it is not applied to the reference poles or toverbose, shown as\-\-, and onnon\-grammaticalit is recorded without filtering, so that figure describes admitted variants rather than a pass rate\. See Appendix[E](https://arxiv.org/html/2610.02627#A5)\.

## Appendix FEnronQA Error Labeling

The error analysis of §[4\.2\.1](https://arxiv.org/html/2610.02627#S4.SS2.SSS1)covers the headline condition \(gpt\-5\.6\-terra, SPLADE, judged bygpt\-5\.4\) for baseline and ten communication style variants poles\. There are118118generation errors are slots in which the gold email was retrieved and the answer was still judged wrong, so the reader had the evidence in front of it\. There are237237retrieval failures that are slots in which the gold email never reached the reader\.

##### Taxonomy\.

Each generation error carries exactly one code, assigned under a fixed precedence so that no slot is double\-counted\.E6, not a genuine reader failure, fires first and covers defective gold answers and judge strictness\.E3is an answer faithful to one of the other four retrieved emails rather than the gold one\.E2bis a real proposition attached to the wrong role, subject or event, andE2aa field filled with a wrong value\. These two combines toE2, unsupported content\.E4is an answer of the wrong semantic type, for instance a name where a time was asked for\.E1is an answer that is not false but incomplete\.E5, false abstention, is a claim that the evidence does not contain the answer when it does\. Retrieval failures use three codes:R1reports the asked fact absent,R2hedges, andR3answers confidently anyway\.

##### Procedure\.

Labels were assigned byclaude\-opus\-5\. For each slot it saw the original question, the rewrite, the gold answer with its accepted alternates, the model answer, the baseline answer, and the full text of every retrieved email, and returned one code with a written rationale\. One author then reviewed all118118labels against the same evidence and recoded2323where the rationale did not support the code under the precedence rule\.

##### Slices\.

An error is*style\-induced*when the unperturbed request answered the item correctly and the rewrite broke it \(4444slots\),*persistent*when the baseline was already wrong \(6565\), and*baseline*when it is the unperturbed run’s own error \(99\)\. Table[8](https://arxiv.org/html/2610.02627#A8.T8)sets the style\-induced slice against the other two pooled, which isolates the kind of failure a rewrite produces\. Two codes appear only in the style\-induced slice: answering in the wrong semantic type \(44slots\) and falsely reporting the answer absent \(22\), both of which misread the request\. Answering from another retrieved email runs the other way,9\.1%9\.1\\%of the style\-induced slice against27\.0%27\.0\\%of the rest\. Unsupported content is the largest category in both,43\.2%43\.2\\%and47\.3%47\.3\\%\.

##### Missing evidence\.

The237237retrieval failures are coded separately\. The reader reports the fact absent in5555of them \(23\.2%23\.2\\%\), hedges in66\(2\.5%2\.5\\%\), and answers confidently anyway in176176\(74\.3%74\.3\\%\)\. It behaves the same way whether the item was never retrievable under the unperturbed request or the rewrite pushed the gold email out of the top five\.

## Appendix GEstimators

Notation follows §[3\.4](https://arxiv.org/html/2610.02627#S3.SS4):did\_\{i\}is the paired difference for itemiiandnnthe number of items present in both the base and variant conditions\.

##### Paired bootstrap\.

We testd¯=1n​∑idi\\bar\{d\}=\\frac\{1\}\{n\}\\sum\_\{i\}d\_\{i\}by resampling item indices with replacement\. Forb=1,…,Bb=1,\\dots,BwithB=10,000B=10\{,\}000and a fixed seed,

d¯∗\(b\)=1n​∑j=1ndij\(b\),ij\(b\)∼Unif⁡\{1,…,n\}\.\\bar\{d\}^\{\*\(b\)\}\\;=\\;\\frac\{1\}\{n\}\\sum\_\{j=1\}^\{n\}d\_\{\\,i\_\{j\}^\{\(b\)\}\},\\qquad i\_\{j\}^\{\(b\)\}\\sim\\mathrm\{Unif\}\\\{1,\\dots,n\\\}\.\(1\)The 95% interval is the2\.52\.5and97\.597\.5percentiles of\{d¯∗\(b\)\}\\\{\\bar\{d\}^\{\*\(b\)\}\\\}, and the two\-sidedpp\-value comes from the proportion of resampled means on either side of zero, capped at11:

F^≶0\\displaystyle\\hat\{F\}\_\{\\lessgtr 0\}=1B∑b𝟙\[d¯∗\(b\)≶0\],\\displaystyle=\\tfrac\{1\}\{B\}\\textstyle\\sum\_\{b\}\\mathbb\{1\}\\big\[\\bar\{d\}^\{\*\(b\)\}\\lessgtr 0\\big\],\(2\)p\\displaystyle p=2​min⁡\(F^≤0,F^≥0\)\.\\displaystyle=2\\min\\\!\\big\(\\hat\{F\}\_\{\\leq 0\},\\;\\hat\{F\}\_\{\\geq 0\}\\big\)\.

##### Exact McNemar test\.

On EnronQA each item yields one binary outcome from a single run, so the paired difference is−1\-1,00or\+1\+1and only the discordant pairs carry information\. Withbbitems correct at baseline but wrong under the variant, andccthe reverse,

p=∑i=0min⁡\(b,c\)\(b\+ci\)​\(12\)b\+c\.p\\;=\\;2\\\!\\\!\\sum\_\{i=0\}^\{\\min\(b,c\)\}\\binom\{b\+c\}\{i\}\\left\(\\tfrac\{1\}\{2\}\\right\)^\{b\+c\}\.\(3\)This conditions onb\+cb\+cand is exact at any sample size, which matters becauseb\+cb\+cis often small: over the EnronQA grid it has a median of1818and falls below1010in2020of114114cells\. A bootstrap over allnnitems would treat a handful of same\-signed events as a stable mean and is correspondingly anti\-conservative in that regime\.

## Appendix HCompute Resources

Experiments were conducted using API\-accessed language models and standard retrieval pipelines\. Retrieval experiments were run on a local machine with 64 GB RAM\. No model training or fine\-tuning was performed\. All RAG and agentic evaluations consisted exclusively of inference\-time API calls\. Statistical analyses were performed using standard CPU\-based computation\.

Table 6:Measured conditions per benchmark and model\. Replicate runs per condition for each benchmark and model\. All 14 conditions the design defines are measured in every cell shown: ten communication\-style poles and four dialect varieties, each paired against that model’s unperturbed baseline\. EnronQA is one inference run scored under three retrievers, written1×31\{\\times\}3\. The headline model is complete on all three benchmarks\. WorkBench was not completed on the other two models and is reported on the headline model alone, so the cross\-model comparison in Table[3](https://arxiv.org/html/2610.02627#S4.T3)is scoped to the two benchmarks where all three models were measured\.Table 7:Per\-cell detail for all three benchmarks, headline model\. Every cell of Table[2](https://arxiv.org/html/2610.02627#S4.T2)with its paired item count, matched baseline, condition value, interval and rawpp\.*base*is the same model’s score on exactly the items that condition covers, so base\+Δ=\{\+\}\\Delta\{=\}cond on every row, and*base*differs slightly between rows wherever the invariance gate dropped items\. All values are percentages on the benchmark’s own metric\.ppis each benchmark’s reported test \(§[3\.4](https://arxiv.org/html/2610.02627#S3.SS4)\): the exact McNemar test on EnronQA and a paired bootstrap over items elsewhere, with the interval from that same bootstrap\. The exact test conditions on the discordant pairs and yields no interval on this scale, so the EnronQA rows carryppalone\.\*\{\}^\{\\textrm\{\*\}\}p<\.05p\{<\}\.05,\*\*\{\}^\{\\textrm\{\*\*\}\}p<\.01p\{<\}\.01,\*\*\*\{\}^\{\\textrm\{\*\*\*\}\}p<\.001p\{<\}\.001\.Table 8:EnronQA error profile, rewrite\-induced against the reader’s own failures\. The118118labeled generation errors on EnronQA \(gpt\-5\.6\-terra×\\timesSPLADE\), split by whether the rewrite created the error\.*style\-induced*errors are those on items the unperturbed request answered correctly;*reader*pools the errors the unperturbed request made itself \(99\) with those it made on items it was already failing \(6565\)\. Code definitions are in Appendix[F](https://arxiv.org/html/2610.02627#A6)\.Table 9:EnronQA fragility by question complexity and reasoning type\. Headline condition \(gpt\-5\.6\-terra, SPLADE\)\.Baseis unperturbed accuracy\. The three breakage columns are the percentage of \(item, pole\) slots that the baseline answered correctly and the rewrite broke, over the four reference poles \(ref\), the six contrastive style poles \(con\) and the four dialect varieties \(dia\)\.![Refer to caption](https://arxiv.org/html/2610.02627v1/figA2_complexity.png)Figure 4:EnronQA breakage by pole and question\-complexity bin\. Columns are the four sampling bins:*simple*\(the control\),*3\-hop*,*multi\-question*and*threaded*\. Each cell is the percentage of slots in that bin which the unperturbed request answered correctly and the rewrite broke\.Table 10:EnronQA accuracyΔ\\Deltafor every answerer×\\timesretriever condition\. Each answerer has three columns, one per retriever: B = BM25, G = BGE\-M3, S = SPLADE\. Cells are accuracyΔ\\Deltaagainst that condition’s own unperturbed baseline\.\*\{\}^\{\\textrm\{\*\}\}p<\.05p\{<\}\.05,\*\*\{\}^\{\\textrm\{\*\*\}\}p<\.01p\{<\}\.01,\*\*\*\{\}^\{\\textrm\{\*\*\*\}\}p<\.001p\{<\}\.001\.Table 11:Retrieval quality and style\-invariance of retrieval, by condition\.⋆\\starmarks the headline condition\. R@5 and MRR are the retrieval quality of the unperturbed queries under that retriever\. The two counts are the number of the six contrastive poles reachingp<\.05p<\.05on gold\-email recall and on answer accuracy\.Table 12:DialectΔ\\Deltafor four attested varieties, three answerer models, two benchmarks\. Each cell is the pairedΔ\\Deltaon that benchmark’s primary metric \(EnronQA accuracy, ToolTalk conversation success\) against the same model’s own unperturbed baseline\. EnronQA is SPLADE for every model\. Restricted to the two benchmarks on which all three models were run; the headline model’s dialect results on all three are in Table[2](https://arxiv.org/html/2610.02627#S4.T2)\. Varieties are generated by the rule\-based transducer at density 1\.0 and are analysed separately, not pooled\. Each benchmark uses the test matched to its outcome \(§[3\.4](https://arxiv.org/html/2610.02627#S3.SS4)\): the exact McNemar test on EnronQA, a paired bootstrap over items withB=10,000B\{=\}10\{,\}000elsewhere\. Rawpp, no multiplicity correction\.\*\{\}^\{\\textrm\{\*\}\}p<\.05p\{<\}\.05,\*\*\{\}^\{\\textrm\{\*\*\}\}p<\.01p\{<\}\.01,\*\*\*\{\}^\{\\textrm\{\*\*\*\}\}p<\.001p\{<\}\.001\.Table 13:Retention of the original head verb as the main predicate, per pole and benchmark\. Each cell is the percentage of that pole’s variants in which the head verb of the original request is still the main predicate, matched on lemma and read from a dependency parse\. The complement is split between deletion, demotion to a subordinate clause, and survival in another role\. Rows are the four reference poles followed by the six contrastive poles;*ref*and*con*mark the role\. Bold marks the lowest cell in each column\. Levels are not comparable between columns: WorkBench instructions are terse imperatives, so even a reference rewrite often replaces the verb, which lowers the whole column\. Computed over all gate\-passing slots \(EnronQAn=n\{=\}3941\), \(ToolTalkn=n\{=\}1599\), \(WorkBenchn=n\{=\}855\); ToolTalk is measured per turn because a main clause is a property of one message\.Table 14:Gold\-email recall underverbose, by how much padding the rewrite added\. Slots are theverbosevariants of EnronQA questions whose gold email was already in the top 5 for the unmodified request under BM25 \(n=370n\{=\}370of 400\), split into quartiles by how many times longer the rewrite is than its own original\.Table 15:EnronQA intermediate signals, headline condition\.r@5is gold\-email recall atk=5k\{=\}5andmrrits mean reciprocal rank\. PairedΔ\\Deltaagainst the unperturbed request\.\*\{\}^\{\\textrm\{\*\}\}p<\.05p\{<\}\.05,\*\*\{\}^\{\\textrm\{\*\*\}\}p<\.01p\{<\}\.01,\*\*\*\{\}^\{\\textrm\{\*\*\*\}\}p<\.001p\{<\}\.001\.Table 16:ToolTalk intermediate signals, headline condition\.recallis the fraction of required ground\-truth API calls the agent made andbad\_action\_ratethe rate of calls with no ground\-truth counterpart\.\*\*\{\}^\{\\textrm\{\*\*\}\}p<\.01p\{<\}\.01,\*\*\*\{\}^\{\\textrm\{\*\*\*\}\}p<\.001p\{<\}\.001\.Table 17:WorkBench intermediate signals, headline condition\.callF1is the F1 between the agent’s predicted tool calls and the ground\-truth ones,noCallthe rate at which no tool call is issued at all, andsideEffectthe rate at which execution leaves an unwanted change to the sandbox\.\*\{\}^\{\\textrm\{\*\}\}p<\.05p\{<\}\.05,\*\*\{\}^\{\\textrm\{\*\*\}\}p<\.01p\{<\}\.01,\*\*\*\{\}^\{\\textrm\{\*\*\*\}\}p<\.001p\{<\}\.001\.Figure 5:All ten poles in the headline condition of each system, with 95% paired bootstrap confidence intervals\. Reference poles are the four meaning\-preserving rewrites, contrastive poles the six styled ones\. Metrics differ across panels and are not commensurable\.
## NeurIPS Paper Checklist

The checklist is designed to encourage best practices for responsible machine learning research, addressing issues of reproducibility, transparency, research ethics, and societal impact\. Do not remove the checklist:The papers not including the checklist will be desk rejected\.The checklist should follow the references and follow the \(optional\) supplemental material\. The checklist does NOT count towards the page limit\.

Please read the checklist guidelines carefully for information on how to answer these questions\. For each question in the checklist:

- •You should answer\[Yes\],\[No\], or\[N/A\]\.
- •\[N/A\]means either that the question is Not Applicable for that particular paper or the relevant information is Not Available\.
- •Please provide a short \(1–2 sentence\) justification right after your answer \(even for\[N/A\]\)\.

The checklist answers are an integral part of your paper submission\.They are visible to the reviewers, area chairs, senior area chairs, and ethics reviewers\. You will also be asked to include it \(after eventual revisions\) with the final version of your paper, and its final version will be published with the paper\.

The reviewers of your paper will be asked to use the checklist as one of the factors in their evaluation\. While\[Yes\]is generally preferable to\[No\], it is perfectly acceptable to answer\[No\]provided a proper justification is given \(e\.g\., error bars are not reported because it would be too computationally expensive” or “we were unable to find the license for the dataset we used”\)\. In general, answering\[No\]or\[N/A\]is not grounds for rejection\. While the questions are phrased in a binary way, we acknowledge that the true answer is often more nuanced, so please just use your best judgment and write a justification to elaborate\. All supporting evidence can appear either in the main paper or the supplemental material, provided in appendix\. If you answer\[Yes\]to a question, in the justification please point to the section\(s\) where related material for the question can be found\.

IMPORTANT, please:

- •Delete this instruction block, but keep the section heading “NeurIPS Paper Checklist",
- •Keep the checklist subsection headings, questions/answers and guidelines below\.
- •Do not modify the questions and only use the provided macros for your answers\.

1. 1\.Claims
2. Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope?
3. Answer:\[Yes\]
4. Justification: The abstract and introduction state the paper’s main claims: that communication\-style and dialect variation can reduce the reliability of email assistants, that indirect requests are the most consistently harmful condition, and that agentic failures are primarily omissions of required actions\. These claims are supported by the experimental results\.
5. Guidelines: - •The answer\[N/A\]means that the abstract and introduction do not include the claims made in the paper\. - •The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations\. A\[No\]or\[N/A\]answer to this question will not be perceived well by the reviewers\. - •The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings\. - •It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper\.
6. 2\.Limitations
7. Question: Does the paper discuss the limitations of the work performed by the authors?
8. Answer:\[Yes\]
9. Justification: The paper includes a Limitations section \(Section 6\) that discusses dialect generation at a single feature density, incomplete benchmark coverage, imperfect reference controls, compute constraints, and the absence of mitigation experiments\.
10. Guidelines: - •The answer\[N/A\]means that the paper has no limitation while the answer\[No\]means that the paper has limitations, but those are not discussed in the paper\. - •The authors are encouraged to create a separate “Limitations” section in their paper\. - •The paper should point out any strong assumptions and how robust the results are to violations of these assumptions \(e\.g\., independence assumptions, noiseless settings, model well\-specification, asymptotic approximations only holding locally\)\. The authors should reflect on how these assumptions might be violated in practice and what the implications would be\. - •The authors should reflect on the scope of the claims made, e\.g\., if the approach was only tested on a few datasets or with a few runs\. In general, empirical results often depend on implicit assumptions, which should be articulated\. - •The authors should reflect on the factors that influence the performance of the approach\. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting\. Or a speech\-to\-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon\. - •The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size\. - •If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness\. - •While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren’t acknowledged in the paper\. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community\. Reviewers will be specifically instructed to not penalize honesty concerning limitations\.
11. 3\.Theory assumptions and proofs
12. Question: For each theoretical result, does the paper provide the full set of assumptions and a complete \(and correct\) proof?
13. Answer:\[N/A\]
14. Justification: The paper does not present theoretical results, theorems, or proofs\. Its contributions are methodological and analytical\.
15. Guidelines: - •The answer\[N/A\]means that the paper does not include theoretical results\. - •All the theorems, formulas, and proofs in the paper should be numbered and cross\-referenced\. - •All assumptions should be clearly stated or referenced in the statement of any theorems\. - •The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition\. - •Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material\. - •Theorems and Lemmas that the proof relies upon should be properly referenced\.
16. 4\.Experimental result reproducibility
17. Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper \(regardless of whether the code and data are provided or not\)?
18. Answer:\[Yes\]
19. Justification: The paper describes the datasets, variant\-generation procedures, validation gates, evaluation protocols, hyperparameters, metrics, and statistical tests needed to reproduce the experiments\. Additional implementation details, prompts, and validation procedures are provided in the appendix\.
20. Guidelines: - •The answer\[N/A\]means that the paper does not include experiments\. - •If the paper includes experiments, a\[No\]answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not\. - •If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable\. - •Depending on the contribution, reproducibility can be accomplished in various ways\. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model\. In general\. releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model \(e\.g\., in the case of a large language model\), releasing of a model checkpoint, or other means that are appropriate to the research performed\. - •While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution\. For example 1. \(a\)If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm\. 2. \(b\)If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully\. 3. \(c\)If the contribution is a new model \(e\.g\., a large language model\), then there should either be a way to access this model for reproducing the results or a way to reproduce the model \(e\.g\., with an open\-source dataset or instructions for how to construct the dataset\)\. 4. \(d\)We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility\. In the case of closed\-source models, it may be that access to the model is limited in some way \(e\.g\., to registered users\), but it should be possible for other researchers to have some path to reproducing or verifying the results\.
21. 5\.Open access to data and code
22. Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?
23. Answer:\[No\]
24. Justification: The experiments use a combination of public benchmarks and proprietary model APIs\. While the paper provides detailed methodological descriptions, anonymized code and artifacts are not included with the current submission\.
25. Guidelines: - •The answer\[N/A\]means that paper does not include experiments requiring code\. - • - •While we encourage the release of code and data, we understand that this might not be possible, so\[No\]is an acceptable answer\. Papers cannot be rejected simply for not including code, unless this is central to the contribution \(e\.g\., for a new open\-source benchmark\)\. - •The instructions should contain the exact command and environment needed to run to reproduce the results\. See the NeurIPS code and data submission guidelines \([https://neurips\.cc/public/guides/CodeSubmissionPolicy](https://neurips.cc/public/guides/CodeSubmissionPolicy)\) for more details\. - •The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc\. - •The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines\. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why\. - •At submission time, to preserve anonymity, the authors should release anonymized versions \(if applicable\)\. - •Providing as much information as possible in supplemental material \(appended to the paper\) is recommended, but including URLs to data and code is permitted\.
26. 6\.Experimental setting/details
27. Question: Does the paper specify all the training and test details \(e\.g\., data splits, hyperparameters, how they were chosen, type of optimizer\) necessary to understand the results?
28. Answer:\[Yes\]
29. Justification: The paper describes benchmark selection, sampling procedures, models, retrievers, prompting methods, validation criteria, evaluation metrics, and statistical testing procedures\. Implementation details are provided in the Experiment section and in the appendix\.
30. Guidelines: - •The answer\[N/A\]means that the paper does not include experiments\. - •The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them\. - •The full details can be provided either with the code, in appendix, or as supplemental material\.
31. 7\.Experiment statistical significance
32. Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments?
33. Answer:\[Yes\]
34. Justification: The paper reports statistical significance using exact McNemar tests and paired bootstrap tests, and reports confidence intervals and p\-values for the main experimental results\. See Section 3\.4 and Table 7\.
35. Guidelines: - •The answer\[N/A\]means that the paper does not include experiments\. - •The authors should answer\[Yes\]if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper\. - •The factors of variability that the error bars are capturing should be clearly stated \(for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions\)\. - •The method for calculating the error bars should be explained \(closed form formula, call to a library function, bootstrap, etc\.\) - •The assumptions made should be given \(e\.g\., Normally distributed errors\)\. - •It should be clear whether the error bar is the standard deviation or the standard error of the mean\. - •It is OK to report 1\-sigma error bars, but one should state it\. The authors should preferably report a 2\-sigma error bar than state that they have a 96% CI, if the hypothesis of Normality of errors is not verified\. - •For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range \(e\.g\., negative error rates\)\. - •If error bars are reported in tables or plots, the authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text\.
36. 8\.Experiments compute resources
37. Question: For each experiment, does the paper provide sufficient information on the computer resources \(type of compute workers, memory, time of execution\) needed to reproduce the experiments?
38. Answer:\[Yes\]
39. Justification: The paper reports the computational setting used for evaluation, including the hardware environment for retrieval experiments and the use of inference\-only API\-based language models as shown in Appendix H\.
40. Guidelines: - •The answer\[N/A\]means that the paper does not include experiments\. - •The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage\. - •The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute\. - •The paper should disclose whether the full research project required more compute than the experiments reported in the paper \(e\.g\., preliminary or failed experiments that didn’t make it into the paper\)\.
41. 9\.Code of ethics
43. Answer:\[Yes\]
44. Justification: The work evaluates the robustness of language\-based assistants using benchmark datasets and synthetic communication variants and does not involve harmful interventions or human experimentation\. The authors believe the research conforms to the NeurIPS Code of Ethics\.
45. Guidelines: - •The answer\[N/A\]means that the authors have not reviewed the NeurIPS Code of Ethics\. - •If the authors answer\[No\], they should explain the special circumstances that require a deviation from the Code of Ethics\. - •The authors should make sure to preserve anonymity \(e\.g\., if there is a special consideration due to laws or regulations in their jurisdiction\)\.
46. 10\.Broader impacts
47. Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed?
48. Answer:\[Yes\]
49. Justification: The paper discusses both positive and negative societal implications, including fair access to AI assistants across communication styles and English varieties, as well as reliability and equity concerns arising from system sensitivity to linguistic variation\. See Discussion Sections 5\.3 and 5\.4\.
50. Guidelines: - •The answer\[N/A\]means that there is no societal impact of the work performed\. - •If the authors answer\[N/A\]or\[No\], they should explain why their work has no societal impact or why the paper does not address societal impact\. - •Examples of negative societal impacts include potential malicious or unintended uses \(e\.g\., disinformation, generating fake profiles, surveillance\), fairness considerations \(e\.g\., deployment of technologies that could make decisions that unfairly impact specific groups\), privacy considerations, and security considerations\. - •The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments\. However, if there is a direct path to any negative applications, the authors should point it out\. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate Deepfakes for disinformation\. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster\. - •The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from \(intentional or unintentional\) misuse of the technology\. - •If there are negative societal impacts, the authors could also discuss possible mitigation strategies \(e\.g\., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML\)\.
51. 11\.Safeguards
52. Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse \(e\.g\., pre\-trained language models, image generators, or scraped datasets\)?
53. Answer:\[N/A\]
54. Justification: The paper does not release a high\-risk model, dataset, or system that requires special safeguards beyond standard research practices\.
55. Guidelines: - •The answer\[N/A\]means that the paper poses no such risks\. - •Released models that have a high risk for misuse or dual\-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters\. - •Datasets that have been scraped from the Internet could pose safety risks\. The authors should describe how they avoided releasing unsafe images\. - •We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort\.
56. 12\.Licenses for existing assets
57. Question: Are the creators or original owners of assets \(e\.g\., code, data, models\), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected?
58. Answer:\[Yes\]
59. Justification: The paper identifies the datasets, benchmarks, models, and software assets used in the experiments, and cites their original sources\.
60. Guidelines: - •The answer\[N/A\]means that the paper does not use existing assets\. - •The authors should cite the original paper that produced the code package or dataset\. - •The authors should state which version of the asset is used and, if possible, include a URL\. - •The name of the license \(e\.g\., CC\-BY 4\.0\) should be included for each asset\. - •For scraped data from a particular source \(e\.g\., website\), the copyright and terms of service of that source should be provided\. - •If assets are released, the license, copyright information, and terms of use in the package should be provided\. For popular datasets,[paperswithcode\.com/datasets](https://paperswithcode.com/datasets)has curated licenses for some datasets\. Their licensing guide can help determine the license of a dataset\. - •For existing datasets that are re\-packaged, both the original license and the license of the derived asset \(if it has changed\) should be provided\. - •If this information is not available online, the authors are encouraged to reach out to the asset’s creators\.
61. 13\.New assets
62. Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?
63. Answer:\[N/A\]
64. Justification: No new dataset, model, or software asset is released as part of this submission\.
65. Guidelines: - •The answer\[N/A\]means that the paper does not release new assets\. - •Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates\. This includes details about training, license, limitations, etc\. - •The paper should discuss whether and how consent was obtained from people whose asset is used\. - •At submission time, remember to anonymize your assets \(if applicable\)\. You can either create an anonymized URL or include an anonymized zip file\.
66. 14\.Crowdsourcing and research with human subjects
67. Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation \(if any\)?
68. Answer:\[N/A\]
69. Justification: The study does not involve crowdsourcing, human subjects, or human participants\.
70. Guidelines: - •The answer\[N/A\]means that the paper does not involve crowdsourcing nor research with human subjects\. - •Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper\. - •According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector\.
71. 15\.Institutional review board \(IRB\) approvals or equivalent for research with human subjects
72. Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board \(IRB\) approvals \(or an equivalent approval/review based on the requirements of your country or institution\) were obtained?
73. Answer:\[N/A\]
74. Justification: The study does not involve human subjects research and therefore does not require IRB review\.
75. Guidelines: - •The answer\[N/A\]means that the paper does not involve crowdsourcing nor research with human subjects\. - •Depending on the country in which research is conducted, IRB approval \(or equivalent\) may be required for any human subjects research\. If you obtained IRB approval, you should clearly state this in the paper\. - •We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution\. - •For initial submissions, do not include any information that would break anonymity \(if applicable\), such as the institution conducting the review\.
76. 16\.Declaration of LLM usage
77. Question: Does the paper describe the usage of LLMs if it is an important, original, or non\-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does*not*impact the core methodology, scientific rigor, or originality of the research, declaration is not required\.
78. Answer:\[Yes\]
79. Justification: LLMs are a central component of the methodology and are used to generate communication\-style variants, perform rubric\-based validation, core experiments and grade benchmark outputs\. Their usage is described in Taxonomy and Experiment Sections 2\.2, 2\.3, and 3\.2\.
80. Guidelines: - •The answer\[N/A\]means that the core method development in this research does not involve LLMs as any important, original, or non\-standard components\. - •Please refer to our LLM policy in the NeurIPS handbook for what should or should not be described\.