Cross Lingual Transfer in Tulu Legal Comprehension: Script-Dependent Improvement and RAG-Induced Knowledge Conflict

# Cross Lingual Transfer in Tulu Legal Comprehension: Script-Dependent Improvement and RAG-Induced Knowledge Conflict
Source: [https://arxiv.org/html/2608.28645](https://arxiv.org/html/2608.28645)
\[ BoldFont = NimbusRoman\-Bold\-renamed\.otf, ItalicFont = NimbusRoman\-Italic\-renamed\.otf, BoldItalicFont = NimbusRoman\-BoldItalic\-renamed\.otf, \]

###### Abstract

Low\-resource languages without an adequate training corpus often use a related, higher\-resource language as a scaffold for comprehension\. Still, there is a need to develop rigorous evaluation methods to identify when models fail in cross lingual low\-resource environments\. Using the legal domain as a backdrop, three models \(Llama3, Hex\-1, Sarvam\) were tested on the ability to classify legal complaints written in a low resource Dravidian language \(Tulu\)\. Transliterating queries across Dravidian scripts allowed models to gain a preliminary understanding of speakers’ complaints without the use of wide scale training, though the level of comprehension was heavily script dependent \(with Kannada \- another relatively low\-resource language \- producing the strongest positive trend\)\. Retrieving from a corpus of Kannada legal papers across a RAG framework caused mixed results\. Some models had a weak positive trend in comprehension under certain conditions, but when models failed, it was often across two axes: fact substitution \(fixating on specific passage excerpts that skewed reasoning\) and confabulation \(hallucination that had no basis in either query or corpus\)\. Within low resource domains, results identify the model’s parsing of information and subsequent reasoning as the source of reasoning failure, rather than corpus contents\. Script\-dependent comprehension and RAG robustness also seem to travel together\. This is further supported by the reasoning\-trace analysis and a statistical\-honesty framework deployed \- techniques that are more broadly applicable to low\-resource multilingual RAG evaluation\.

## 1Introduction

Despite the various capabilities of large language models, one of the most pressing pitfalls is how models assess information in low\-resource domains\. In many regions of the world, the common language spoken \(often a highly low\-resource language\) differs from the administrative language\. To exacerbate the issue, the administrative language is often low\-resource itself, requiring novel techniques for models to bridge this cross\-lingual low\-resource domain\. A relevant application of this dilemma is the legal field, where members of such communities often need to represent themselves in a non\-native language to access resources\.

This study uses the Dravidian language Tulu, a small language with a few million speakers off the coast of Karnataka\. The official language of Karnataka is Kannada, meaning all legal proceedings take place with Kannada or English mediums\. Kannada is itself classified a low resource language, as despite having an extensive literary history, training LLM on it has only picked up recently\. Tulu is extremely low resource; it does not contain the extensive literature used to traditionally train models\.

With a low resource language, Tulu, how would it be possible to file speakers’ complaints into a system that does not administratively use their language \(and instead uses Kannada\), without a translator? In this study, we explore whether minimal priming and legal papers written in Kannada are enough for a model to accurately interpret queries written in a language they do not have previous exposure to\.

While high\-resource languages are often used as a starting point when working with low\-resource languages, the intuition points in the other direction\. When Tulu is written, it has traditionally used Kannada abugidas rather than the Latin alphabet\. Tulu speakers are also more likely to borrow vocabulary and structure sentences similarly to the more common language around them \- Kannada\. Thus, we propose that use of the Kannada \(and other Dravidian\) scripts would trigger stronger comprehension in models\. Furthermore, legal papers written in Kannada would provide a stronger foothold for an LLM to understand Tulu complaints spoken in an informal register due to the aforementioned relation between the two languages\.

The purpose of this paper is not to create a full, working LLM for low resource languages, but to evaluate model comprehension in a cross\-lingual low\-resource domain, using the legal field as an appropriate, grounded testing ground\. To do so, we ask three LLM models to categorize legal complaints written in Tulu using the Legal Issues Taxonomy \(LIST\) from the Stanford Legal Design Lab\([Hagan 2026](https://arxiv.org/html/2608.28645#bib.bib5)\), without translation or heavy training\. We explore whether minimal prompting and RAG from a corpus in a higher resource, related language, can serve as a scaffold for models to pinpoint the gaps in their understanding of the complaint\.

Most Tulu speakers are multilingual and can communicate in Kannada, English, Hindi, etc, though doing so can still result in the loss of nuance that occurs when speaking in a second language\.

## 2Related Work

Knowledge Conflict\- When training large language models, they often run into knowledge conflict, as shown by\([Longpre et al\. 2021](https://arxiv.org/html/2608.28645#bib.bib9)\)\. When models were forced to decide between a newly given passage or their own memory, they often fall back to their own understanding, even when prompted otherwise\. However, later findings diverge from this understanding\.\([Xu et al\. 2024](https://arxiv.org/html/2608.28645#bib.bib19)\)establishes the ”context\-memory conflict”, where models struggle to properly weigh retrieved information that conflicts with their own internal understanding, and can inconsistently defer to context that is coherent and confidently stated, regardless of accuracy\. This is further explored in\([Wu et al\. 2024](https://arxiv.org/html/2608.28645#bib.bib17)\), which found that models consistently abandon their correct prior knowledge in favor of retrieved context \- including Llama3 \- one of the models used in this study\.\([Xie et al\. 2024](https://arxiv.org/html/2608.28645#bib.bib18)\)lends credence to the claim that models are receptive to confidently presented external evidence, even if a knowledge conflict is present, though notes another separate confirmation bias \- models tend to stick with their own prior understanding when the evidence set only partially agrees with it\. Given that model comprehension in a low resource, legal context is still not widely understood, our unique angle isolates how these mechanisms can be further utilized by comparing multiple models’ innate ability, retrieval analysis, and discernment between the two in a multilingual setting\.

Effect of RAG Inclusion\- RAG’s inclusion can be detrimental to knowledge synthesis\.\([Hsia et al\. 2025](https://arxiv.org/html/2608.28645#bib.bib6)\)explains how various model/retriever combinations can affect downstream performance, sometimes performing worse compared to no retrieval at all\. Witnessed as well in\([Zeng et al\. 2025](https://arxiv.org/html/2608.28645#bib.bib20)\), where models performed worse when presented with misleading retrieved evidence compared to a zero shot baseline and could not recognize when to discard incorrect information\. Our work furthers this discussion, displaying how models are able to use minimal priming to make their predictions more accurate, but the inclusion of RAG prevents most models from effectively discerning when to disregard irrelevant passages\.

Hallucination in Legal NLP\- It’s especially vital to understand this issue in legal work, where small knowledge conflicts can snowball into major hallucinations\.\([Magesh et al\. 2025](https://arxiv.org/html/2608.28645#bib.bib10)\)establishes how commercial legal AI built around RAG still meaningfully hallucinates, with literature like\([Reuter et al\. 2025](https://arxiv.org/html/2608.28645#bib.bib14)\)working towards mitigation\. Our findings add to established literature, demonstrating that even relevant legal papers can cause hallucination within a model, which may be due to cross\-lingual transfer or uncertainty in a low resource environment\.

Retrieval and Low Resource Languages\- This problem is magnified among low resource languages:\([Sun et al\. 2026](https://arxiv.org/html/2608.28645#bib.bib16)\)illustrate how low\-resource language pairs rely more heavily on retrieved context and degrade more severely when that context is noisy compared to high\-resource pairs\.\([Hu et al\. 2025](https://arxiv.org/html/2608.28645#bib.bib7)\)extends this, demonstrating that while knowledge\-free reasoning easily transfers across language despite the secondary resource effect, knowledge retrieval does not\. This may explain why RAG frequently failed when the query was not written in a high resource language, which did not allow the model to discriminate between irrelevant passages and its own reasoning\. Our methods also draw from\([Qi et al\. 2025](https://arxiv.org/html/2608.28645#bib.bib13)\), which separates context utilization from retrieval quality, though ours differ by maintaining consistent retrieval over queries written in multiple languages\. Our findings are also supported by\([Mohan et al\. 2026](https://arxiv.org/html/2608.28645#bib.bib11)\), which explains how a failure occurs not from retrieved content itself, but how the model uses it \- rationale for why introducing retrieved passages often failed, as the issue may lie with the model’s ingestion of the passages as well as confusion with multilingual transfer across a low resource domain\. This intersection of low resource multilingual transfer across the legal domain is relatively unexplored\.

Tulu\-Kannada Relation\- Within the scope of the Tulu\-Kannada bridge,\([Narayanan and Aepli 2024](https://arxiv.org/html/2608.28645#bib.bib12)\)establishes that because of the extensive contact between the two languages, using Kannada as a scaffold for Tulu seems like a viable option\. This is supported by\([Devadiga and Chopra 2026](https://arxiv.org/html/2608.28645#bib.bib4)\), which discovers how Tulu is ”nearly indistinguishable” from Kannada to Kannada trained LLM\. Our results align, given the drop in comprehension when Tulu is written in other Dravidian scripts compared to Kannada\.

## 3Methodology

### 3\.1Task Formulation

To discover to what extent a model could understand queries in Tulu, we authored 60 sentences describing common legal situations, with each sentence ascribed a primary category, inspired by the Legal Issues Taxonomy \(LIST\) from the Stanford Legal Design Lab\([Hagan 2026](https://arxiv.org/html/2608.28645#bib.bib5)\)\(where some categories were replaced\), as well as any relevant secondary categories\. The model had to use the sentence given \(as well as any additional context\) to predict what category the Tulu complaint fell under, with performance evaluated across the 60 sentences\. The hope was to build a pipeline that could classify complaints in a low resource language with minimal prompting\.

### 3\.2Data and Models

The 60 Tulu sentences were written in an informal, common register describing a situation requiring legal analysis\. The sentences were written in Latin script as well as three Indic scripts, primarily Kannada, but also Malayalam and Tamil, and translated into English as a control\.

The legal corpus consisted of 18 legal documents from Karnataka courts \(taken from ecourtsindia\.com\), written in Kannada\. Documents were orders of various sizes, ranging from a paragraph and a table describing common proceedings to 25\-page hearings detailing witness testimonies\. Many corpus documents consisted of images or used legacy non\-Unicode fonts that produced mojibake when extracted \(garbled Latin text rather than standard Kannada\), requiring automated text extraction\. To clean up extraction based errors, AI assisted correction was utilized, an established technique\([Kanerva et al\. 2025](https://arxiv.org/html/2608.28645#bib.bib8)\)\. Using Gemini, original documents were given alongside the scraped text for the model to identify and correct text based errors, ensuring the model had a reference to source material rather than open\-ended inference\.

The cleaned legal papers were each ascribed to at least one major LIST category, although secondary categories could be ascribed as well\. The documents exist solely for the model to gain an understanding of similar problems to the query, to aid diagnosis\.

Three models were used to understand cross\-lingual legal transfer\. Llama3\([AI@Meta 2024](https://arxiv.org/html/2608.28645#bib.bib1)\)had no official support for Kannada or Tulu, though a small portion \(over 5%\) of its training data covered more than 30 additional languages not officially documented\. Sarvam\([Sarvam AI 2024](https://arxiv.org/html/2608.28645#bib.bib15)\)and Hex\-1\([Bud Ecosystem 2025](https://arxiv.org/html/2608.28645#bib.bib2)\)were both trained with a focus on Indic languages including Kannada, but not Tulu, testing whether a Kannada basis helps comprehension\. All three models had temperature 0\.1 for reproducibility; Llama3 and Hex\-1 had a repeat penalty of 1\.1\. Hex\-1’s context window \(num\_ctx\) was raised to 8192 to prevent truncation\. Enabling thinking mode on Hex\-1 caused inner reasoning with no response, while Sarvam’s reasoning consumed all tokens in its ”starter tier” \(4000 tokens\), also yielding no output, so thinking/reasoning was disabled for both\.

For embedding the legal texts, we originally used nomic\-embed\-text, but switched to multilingual\-e5\-large\-instruct, as the latter produced embeddings spanning a wider range rather than deferring to a default median\.

### 3\.3Experimental Conditions

Each condition was composed of four parts: \(1\) the script and language the query was written in, \(2\) whether the model received linguistic priming, \(3\) whether the model received RAG passages from the Kannada legal corpus, and \(4\) if primed, whether the priming was sparse or extensive\.

Priming had two variants\.Sparsepriming contained a short paragraph of common Tulu vocabulary \(most not present in the test sentences\) and basic grammatical structure\.Extensivepriming was a continuation of sparse priming that additionally included words related to describing legal situations \- e\.g\. family, land, permission \- some of which were used directly in the 60 test sentences\.

Tulu words in the priming block matched the query’s script \(e\.g\. a Romanized\-Tulu query received Romanized\-Tulu priming vocabulary\), while instructions remained in English throughout\. The English control never received priming\.

Retrieved passages shown to the model included the source case name, retrieval similarity score, and the passage’s own LIST category label\. Baselines included English\-gloss queries with RAG \(Condition A\), to test how RAG affected comprehension with a high\-resource query language, and Romanized/Kannada\-script queries without priming or RAG \(Condition G and Condition H, respectively\), to test latent understanding without modification\.

Table 1:Experimental conditions, defined by query script, presence of linguistic priming, and use of retrieval\-augmented generation \(RAG\)\. Conditions with priming \(C, D, E, K, L, M\) were each run at both sparse and extensive tiers\.#### 3\.3\.1Prompting

Each model was first given the query, after which retrieved RAG passages were provided if applicable\. If RAG was used, the model was instructed not to classify the passages, but to use them to understand the situation described in the query\. The model was again given the query, then presented a third time with explicit instructions not to classify retrieved passages\. This repeated instruction was necessary, as otherwise the model often resolved to classify its interpretation of retrieved passages\.

### 3\.4Retrieval Pipeline

All legal passages were first broken into 800 character chunks, which were embedded using nomic\-embed\-text, then switched to multilingual\-e5\-large\-instruct \(the query was embedded similarly\)\. The embedded query was compared against chunks and the top 3 chunks were retrieved using cosine similarity\. Mean centering was briefly deployed to test whether it increased the number of relevant passages retrieved, but provided no recognizable optimization\.

### 3\.5Diagnostic Methodology

To assess whether the model was able to comprehend the query, its response was analyzed across four separate axes\.

Tertiary Classification\- Since queries could often relate to multiple categories, each query was given two other related LIST labels, alongside its primary LIST classification\. The model, after being provided the query and possible subsequent material, had to provide a primary, secondary, and tertiary LIST category, and could not repeat categories\. If the primary category provided by the model matched the assigned primary classification, it received a point to its Hits@1 score\. If the model’s primary classification did not match, but the subsequent guesses corresponded with any of the three categories assigned to the query, it received a point to the Hits@3 score\.

A high Hits@1 and Hits@3 score indicated the model was able to successfully assign the proper LIST category, as well as any tangentially related categories\. A high Hits@3 but low Hits@1 score was seen as the model partially understanding the query, but not enough to capture the nuance described\. A low Hits@1 and Hits@3 score meant the model did not have strong comprehension of the queries\.

Reasoning Trace Verification\- Models were instructed to provide reasoning for their primary guess, to ensure whether the model had actually understood the situation, or produced an unrelated reason that coincidentally aligned with the query\. A model with a high number of such situations was seen as partially parsing the query, but then hallucinating the gaps in its comprehension\.

Statistical Honesty\- To ensure that discrepancies between conditions were significant and not due to random error, Fisher’s Exact Test was deployed \(e\.g\. comparing Condition E and Condition D for a given model\)\. Binomial tests were also used, comparing a model’s D\-condition performance against random guessing \(12\.5% accuracy\) and majority class guessing \(25% accuracy, assigning the most common category across queries\)\. Multiple runs for a specific model/condition pair were pooled for stronger statistical analysis\. The p value must be less than the standard threshold of 0\.05 to be statistically significant\. Findings that qualified as significant were described with confident language \(e\.g\. ”demonstrates”, ”shows”\), while those that did not were viewed as directional or indicative of a trend, even if raw percentage gaps seemed large\.

### 3\.6Preliminary Mitigation

After the discovery of RAG hurting a model’s discernment, a tentative prototype was tested\.

CRAG\-Style Relevance Verification\- A secondary judge layer was added before the prompt, in which the model discerned whether each retrieved passage was relevant with a binary judgment \(for 20 complaints only\)\. Each call had two forms: one required the model to use queries written in Kannada script, the other used the English \(Condition A\) queries\. Each passage was reviewed via independent calls\. Non\-relevant passages were excluded, and if all passages were deemed irrelevant, the prompt fell back to the No\-RAG format\. This was tested on the Llama3 D conditions \(RAG inclusive & priming\)\.

## 4Findings

Figure 1:Hits@1 across five conditions per model \(n=60n=60each\): bare comprehension \(H\), no\-RAG full\-priming ceiling \(E\_full\), RAG without priming \(F\), RAG with full priming \(D\_full\), English\-gloss query with RAG \(A\)\. Only Llama3 drops below its own no\-RAG ceiling once RAG is added\. Full 19\-condition breakdown: Appendix[D](https://arxiv.org/html/2608.28645#A4)\.Figure 2:Query language under RAG: English\-gloss queries \(Condition A\) substantially outperform Kannada\-script Tulu queries under identical retrieval \(Condition D\), for all three models \(Fisher’s exact,p<\.001p<\.001in every comparison\)\. This is the one RAG\-related effect that generalizes across models\.Across all three models, Kannada\-script bare comprehension \(Condition H\) is significantly above both chance and majority\-class baselines \(Llama3 46\.7%, Hex\-1 41\.7%, Sarvam 48\.3%;p<\.001p<\.001in every case\), and full\-tier priming reveals a trend with comprehension being increased further for every model, when using Kannada \(Figure[1](https://arxiv.org/html/2608.28645#S4.F1)\)\. The picture changes sharply once RAG is introduced\. For Llama3, RAG collapses performance well below its own no\-RAG ceiling \(D\_full 23\.3% vs\. E\_full 60\.0%,p<\.001p<\.001\); Hex\-1 and Sarvam show no significant degradation on the same comparison \(p=\.20p=\.20andp=\.57p=\.57, respectively\)\. The full 19\-condition, four\-script, three\-model breakdown and complete significance tests are reported in Appendix[D](https://arxiv.org/html/2608.28645#A4)\.

The one RAG\-related effect that generalizes across all three models is query language: English\-gloss queries under RAG \(Condition A\) significantly outperform Kannada\-script Tulu queries under otherwise identical retrieval \(Condition D/Full\), for every model \(p ¡ \.001 in all three comparisons; Figure[2](https://arxiv.org/html/2608.28645#S4.F2)\)\. A full comparison of Kannada and Romanized script performance across all priming and RAG conditions is reported in Table[10](https://arxiv.org/html/2608.28645#A4.T10)\. This suggests that whatever degrades performance under RAG interacts more consistently with query language than with model identity\.

## 5Discussion

### 5\.1Comprehension Transfers Without RAG

One pressing finding was the repeatable success of Condition E\. This trend emerges with Condition H, where models were able to classify given sentences without any scaffolding \(at 46\.7%, 41\.7%, and 48\.3% for Llama3, Hex\-1, and Sarvam respectively\)\. In other cases, the model was able to partially understand the sentence, but not enough to accurately diagnose it\. To ensure this was not random chance, we also checked the reasoning given\.

Table 2:Sarvam, Condition H, ids 14 and 17\. Row 1: human chosen and predicted category both Courts/Legal\-System, reasoning names the specific recognized term rather than a generic paraphrase\. Row 2: human chosen \- Land/Property, predicted \- Housing\. The model parses the core action but assigns an adjacent category, indicating partial rather than absent comprehension\.Still, this was a preliminary indication that models were somehow able to occasionally parse queries correctly, despite no priming or previous training with the language\.

Once the model received priming \(Condition E\), Llama3’s and Hex\-1’s understanding and subsequent diagnosis improved slightly \(46\.7%→\\to51\.7% and 41\.7%→\\to45\.0%, respectively, though neither reaches significance:p=\.715p=\.715andp=\.854p=\.854\), and this is not reflected with Sarvam \(48\.3%→\\to46\.7%,p=1p=1\)\.

We had two theories as to how the model understood Tulu with brief exposure to the language’s structure and no previous interactions with complete sentences in the target language\.

The first proposed that the model was parsing Tulu as a form of Kannada, since the shared history between the two languages facilitated heavy word borrowing\.111See[Devadiga and Chopra 2026](https://arxiv.org/html/2608.28645#bib.bib4), who find Kannada\-script Tulu “nearly indistinguishable from Kannada” to models trained predominantly on the latter\.The second theory asserts that the model is reading Tulu as a unique Dravidian language, using features shared across the Dravidian family to make sense of the language\. Given that Tulu exists in the same language family as Tamil and Malayalam \(as well as Kannada\), this also seemed viable\.

Given the results of Condition I\-L, script matters, though not substantially at the sparse\-priming tier, and this holds across all three models\. Sparse structural priming, grammar and word order only with no category\-revealing vocabulary, produced only modest, non\-significant changes in Kannada accuracy for every model \(Llama3 46\.7%→\\to51\.7%,p=\.715p=\.715; Hex\-1 41\.7%→\\to45\.0%,p=\.854p=\.854; Sarvam 48\.3%→\\to46\.7%,p=1p=1\), and left Tamil \(Llama3 16\.7%→\\to15\.0%; Hex\-1 28\.3%→\\to33\.3%; Sarvam 28\.3%→\\to33\.3%; allp\>\.26p\>\.26\) and Malayalam \(Llama3 21\.7%→\\to20\.0%; Hex\-1 25\.0%→\\to28\.3%; Sarvam 21\.7%→\\to38\.3%; allp\>\.07p\>\.07\) similarly flat or only weakly positive\. If priming were teaching models portable Tulu grammar, this improvement should not be script\-dependent\. Since the difference does not reach significance for any script or model, this is only suggestive of script\-specific assimilation over broad cross\-Dravidian transfer\. This effect was comparable with Romanized Tulu \(Llama3 16\.7%→\\to13\.3%,p=1p=1; Hex\-1 16\.7%→\\to18\.3%,p=1p=1; Sarvam 33\.3%→\\to30\.0%,p=\.845p=\.845\)\.

It is undeniable that models do have latent capability for understanding queries in this low\-resource language, although the workings are still unclear\. Parsing through a related, higher\-resource language that the target language is lexically \(rather than grammatically\) similar to, and specifically through matching script, seems likely, although the relatively strong comprehension among other Dravidian scripts cannot be disputed\.

### 5\.2RAG Causes a Large, Reproducible Collapse or No Significant Improvement

Llama3’s comprehension dropped when presented with RAG, illustrated by comparing Condition D to F \(F 25\.0% vs\. D\_sparse 26\.7%,p=1p=1; vs\. D\_full 23\.3%,p=1p=1\), where priming did not help under RAG as it did in Condition E \(H 46\.7%→\\toE\_full 60\.0%\), though this did not reach significance\. The issue also arises in the control \(Condition A\): even when the query is written in a high\-resource language, a comparatively lower\-resource retrieval corpus still degraded the model’s own understanding of the query: Llama3 reaches only 61\.7% in this condition, well short of Hex\-1’s 86\.7% and Sarvam’s 91\.7% under the identical retrieval setup\. Meanwhile, both Sarvam and Hex\-1 were able to maintain similar scores with the inclusion of RAG compared to just priming \(Hex\-1: F 38\.3% vs\. E\_sparse 45\.0%,p=\.579p=\.579\), and priming seemed to hurt Sarvam \(Conditions E\_sparse vs\. F: 46\.7% vs\. 53\.3%,p=\.584p=\.584\), though not significantly\. This may be due to Sarvam and Hex\-1 being trained on Kannada, allowing stronger discernment between passages and their own reasoning, which may also explain Llama3’s heavier drop in Condition A\.

A disagreement\-level diagnostic further confirms this, which determined the ratio of Condition E \(non\-RAG/priming\) vs Condition D \(RAG/priming\) producing the right answer when disagreement occurred\. D failed more often E on almost all comparisons when disagreement occurred \- as high as 6\.5:1 with Llama3 at full priming \- except Sarvam and sparse priming in which this trend reverses \(0\.4:1\), with the full Table[11](https://arxiv.org/html/2608.28645#A4.T11)reflecting model robustness under RAG\.

When failure due to RAG did occur, the idea that the cause was knowledge conflict, not retrieval noise, is supported across several dimensions\.

A model’s garbled post\-RAG responses fell into two categories: fact substitution and confabulation\. Withfact substitution, the model takes a specific but wrong detail present in the retrieved passage and incorporates it directly into its reasoning, treating the snippet as part of the query rather than supporting evidence\. Withconfabulation, the model hallucinates information present in neither the query nor the retrieved passage, using it as a basis for reasoning regardless\. This occurs even when a genuinely relevant passage is retrieved: for two debt\-related queries, the retrieved case \(Sridhar P\. v\. Ramesh, see Appendix[E](https://arxiv.org/html/2608.28645#A5)\) was a real, on\-topic cheque\-bounce dispute, yet the model still produced unrelated categories for both\.

Methodological note: retrieved passages were shown to the model together with their own source category label\. A follow\-up ablation stripping this label from the retrieved context produced no consistent change in accuracy \(deltas within±\\pm5pp across all five RAG conditions, no consistent direction\), indicating label\-copying was not the primary driver of the collapse; the failure is substantially content\-level, not a result of label exposure\.

### 5\.3Relation of RAG Inclusion and Script Across Model Performance

While the independent effects of RAG inclusion and script have been stated, the intersection between the two highlights a possible relation \- primarily evident between Kannada script and Latin script queries\. With non\-RAG conditions, Kannada script significantly outperformed Latin script across majority of model and condition combinations Table[10](https://arxiv.org/html/2608.28645#A4.T10), lending more credence to the discussion present in §[5\.1](https://arxiv.org/html/2608.28645#S5.SS1)\. Models can and do exhibit comprehension of low resource languages that they had previously never encountered, but the inclusion of Latin script adds a complicating factor \- that models struggle to parse queries in a low\-resource language if the script the language is written in is far removed from related languages in the model’s training base \(though comprehension can still exist\)\.

However, this difference in comprehension between script is not universally present with RAG inclusion\. Sarvam and Hex\-1 \- both models that upheld performance with RAG inclusion \(with comparison between D and E conditions\)\- were also models that could maintain the Kannada script advantage with RAG conditions \(across Conditions C and D in Table[10](https://arxiv.org/html/2608.28645#A4.T10)\)\. Llama3 \- the model in which accuracy decreased \(Condition D vs E\), was also the only model that could not significantly beat the Latin query performance present in Condition C \(though the raw numbers of Condition D were still greater\)\. Whether a model is able to successfully overcome knowledge conflict and utilize the script advantage seem to be correlated\. Furthermore, whether a model is able to uphold performance with RAG inclusion is also an indicator of whether script advantage will remain under the same RAG conditions\.

It should be noted that while Llama3’s performance dropped \- resulting in insignificance \- Sarvam’s performance was remarkably high in Latin script queries \(once again causing insignificance\)\. So while an insignificant increase in three conditions exists with Sarvam \(Condition E/\(no RAG & sparse priming\), Condition H \(no RAG & no priming\), and Condition D/\(RAG & full priming\), the underlying cause differed from Llama3\. It’s possible that Sarvam’s explicit training in Kannada allowed it to understand some of the Latin script queries\.

Why and when a model is able to understand a script with non\-RAG conditions but not under RAG conditions is still unclear\. Both Sarvam and Hex\-1 are trained in Kannada, which may allow them to maintain script advantage even alongside complex legal passages\. Since parsing Tulu queries as Kannada produces more accurate results, the ability to uphold the script advantage could also cause the model to overcome the knowledge conflict present in RAG conditions, hence why neither model was significantly affected by the inclusion of RAG\. Llama3 \- the one model not explicitly trained in Kannada \- may not have a strong enough basis to recognize the query language and corpus language to be similar even with the same script, hence the drop in accuracy\. However, this is entirely speculative, and more work must be done to understand this phenomenon\.

### 5\.4Mitigation of RAG Failure

Using Corrective Retrieval\-Augmented Generation \(CRAG\), models tested with Kannada\-script queries performed similarly to the unmitigated RAG condition \(25\.0% Hits@1, vs\. 23\.3% for D/full\)\. There was an improvement \(to 60\.0%\) when the judge operated on English\-translated complaints instead of Kannada\-script queries, matching Llama3’s own no\-RAG ceiling \(E/full, 60\.0%\), even though RAG was still used in 17 of 20 cases; this difference does not clear significance at the standard threshold \(p = \.054, Fisher’s exact\), though it sits at the boundary\. The English\-gloss judge also discerned more passages to be irrelevant on average \(1\.65 of 3 retained, vs\. 2\.85 of 3 for the Kannada judge with full results in Table[12](https://arxiv.org/html/2608.28645#A4.T12)\), suggesting the judge layer’s ability to discriminate relevant from irrelevant retrieved content depends on the language it operates in, not solely on the underlying passage relevance\.

## 6Conclusion

In communities where the spoken language differs from the administrative language, training legal models on the latter does not address the issue of communities needing to use these tools in a cross lingual manner in courts, testimonies, etc\. Common languages spoken in the affected area \(Tulu, in this case\) may not even have a literary history, making large scale training of models difficult\. With the language pair of Kannada \(administrative\) and Tulu \(spoken\), the two share a long history of word borrowing and a similar syntactical structure\. Considering this basis, is it possible to bridge the gap between two low resource languages, and alleviate the difficulty of switching between vernacular and formal legal speak with RAG and minimal priming alone? The strategy seems viable, although it heavily depends on the model used\.

Given the task of classifying Tulu queries written in an informal register into specific LIST categories, models could be given two forms of priming, RAG retrieved passages, both priming and the inclusion of RAG, or neither technique at all\. Surprisingly, all three models \(Llama\-3, Hex\-1, and Sarvam\) displayed relatively high levels of comprehension of the queries even with no priming or RAG at all, and both extensive and sparse priming produced an improvement over bare comprehension within conditions utilizing Kannada script for most conditions and models, though this effect did not reach significance for any model in Kannada script\. However, RAG had the opposite effect on Llama3\. Despite preliminary evidence that Tulu is parsed as Kannada by LLMs, the inclusion of actual Kannada legal papers only served to confuse the model on two fronts: fact substitution and confabulation\. This may be because the model, treating Tulu as a form of Kannada, overly trusted retrieved passages over its own reasoning\. Meanwhile, Sarvam and Hex\-1 were not negatively or positively affected by RAG inclusion, though the same two mechanisms still appeared\. Why Sarvam and Hex\-1 did not produce the same RAG induced collapse as Llama3 is still unclear, though comparison of performance between Romanized and Kannada script queries signals that latent comprehension between scripts and the effect of RAG inclusion are interrelated rather than independent phenomena\. While Llama\-3 did exhibit comprehension with Kannada script, this effect did not significantly manifest once RAG was included\. Hex\-1 was able to maintain the advantage of Kannada script under both RAG and non\-RAG conditions as well as Sarvam at most tiers, beating performance under Romanized queries\. It is possible that due to training on Kannada, both Sarvam and Hex\-1 were able to use internal knowledge and treat external sources as confirmation rather than a definitive answer\. Further work must be done to understand this trend\.

Llama3 \(the only model hallucination reduction attempts were studied on\) was somewhat responsive to the CRAG\-style mitigation approach\. CRAG Relevance Verification indicated the model could more easily discern irrelevant passages when the query was written in English, though this effect did not transfer to Kannada script\.

Research in multi\-lingual legal transfer that does not require heavy training is necessary, as affected communities often do not have the means to do so\. In future work, we hope to continue bridging this gap, exploring cross\-Dravidian transfer, larger datasets, more extensive RAG hallucination mitigation methods, and different language pairs to isolate the mechanisms and make such techniques generalizable to the greatest number of communities affected by this issue\.

## Limitations

Due to the time intensive need to carefully review queries to ensure native\-like production, accurate grammar, and correct transliteration, it was considered unrealistic to generate an extensive number of queries for a pilot study\. This study originally employed only 20 queries to test comprehension, but this number was deemed to be too small to be able to confidently assess results\. Therefore, the final number of 60 queries was used to draw preliminary conclusions, although many of the findings were indicative of trends rather than statistically significant behavior, and would require a number of queries that enters the hundreds or thousands range to be considered significant\. However, for large effects, such as the drop in Llama3 comprehension in condition D, these reach statistical significance even with a small dataset\. Even so, if this study was to be expanded, having a larger number of queries would be a top priority\.

The hallucination findings of this study do raise the question of whether similar risk was present during corpus creation, as final documents were scanned but not thoroughly reviewed by a native speaker\. Kanerva et al\.\([Kanerva et al\. 2025](https://arxiv.org/html/2608.28645#bib.bib8)\)similarly found LLM\-based OCR correction to be highly effective for a well\-supported language \(English\) while not reaching such levels in a lower\-resource language \(Finnish\)\. It should be noted that Gemini officially supports Kannada\([Deccan Herald 2024](https://arxiv.org/html/2608.28645#bib.bib3)\)\. Additionally, the task required character verification rather than semantic deduction, distinguishing this corpus\-creation context from that failure case\. We still note this does not fully eliminate the residual risk discussed above and maintaining accuracy in future studies requiring a similar corpus should be a priority\.

As for the legal documents used themselves, their public availability raises some concerns\. Specifically, there is no way to verify if the models used in this study had been previously trained on said documents\. Performance under RAG\-inclusive conditions could be confounded if models were previously exposed to documents present in the corpus, and the non\-RAG/RAG divide discussed in this paper should be revisited if this is found to be the case\.

With regards to models, Sarvam and Hex\-1 had their reasoning/thinking modes disabled, as compared to Llama3 which was allowed unconstrained free\-form generation\. Therefore, Sarvam’s and Hex\-1’s responses should be analyzed knowing that they were not produced under equivalent conditions\.

## Ethics Statement

While all 60 queries are authored to represent realistic hypothetical scenarios, they are not real individuals’ complaints and do not draw from real case intake data\. We believe this is the best way to maintain accurate analysis without the need for private individuals’ legal situations\. Meanwhile, corpus documents were directly retrieved from ecourtsindia\.com, where court orders are publicly accessible\. Retrieved passages were not anonymized in this study\. Court orders were only used to extract document content and LIST category relevant information, and personal information was never used as a basis for classification\. We acknowledge this as a limitation and recommend that named\-entity redaction be utilized before using retrieved passages in further studies\.

It should be heavily noted that this work is not deployment ready, given our findings of model hallucination present throughout experimentation\. Deploying such pipelines has high chances of directing clients to LIST categories irrelevant to their query\. Furthermore, due to the identified phenomena of fact substitution and confabulation, models can confidently provide incorrect reasoning, which can further mislead clients who do not have extensive experience in the legal field\.

While prompt construction and priming methodology are detailed in Appendix[A\.2](https://arxiv.org/html/2608.28645#A1.SS2)and Appendix[B](https://arxiv.org/html/2608.28645#A2), all scripts used during experimentation will be accessible in their entirety upon publication, as well as the 60 authored sentences\. As for the legal corpus, the redistribution rights of the papers remain unclear, although case identifiers for examples discussed in Appendix[E](https://arxiv.org/html/2608.28645#A5)are available upon request\.

## References

- AI@Meta \(2024\)AI@Meta\. 2024\.[The llama 3 herd of models](https://arxiv.org/abs/2407.21783)\.*Preprint*, arXiv:2407\.21783\.
- Bud Ecosystem \(2025\)Bud Ecosystem\. 2025\.Introducing hex\-1: A fully open\-source LLM for indic languages\.[https://blog\.budecosystem\.com/hex1\-open\-source\-llm\-for\-indic\-languages/](https://blog.budecosystem.com/hex1-open-source-llm-for-indic-languages/)\.
- Deccan Herald \(2024\)Deccan Herald\. 2024\.Google’s gemini AI app now available in india, supports 9 languages\.[https://www\.deccanherald\.com/technology/googles\-gemini\-ai\-app\-now\-available\-in\-india\-supports\-9\-languages\-3070891](https://www.deccanherald.com/technology/googles-gemini-ai-app-now-available-in-india-supports-9-languages-3070891)\.Confirms official Kannada support; cited as a news source, not a Google primary document\.
- Devadiga and Chopra \(2026\)Prathamesh Devadiga and Paras Chopra\. 2026\.[Making large language models speak tulu: Structured prompting for an extremely low\-resource language](https://doi.org/10.18653/v1/2026.loreslm-1.5)\.In*Proceedings of the Second Workshop on Language Models for Low\-Resource Languages \(LoResLM\)*, pages 50–61, Rabat, Morocco\.
- Hagan \(2026\)Margaret Hagan\. 2026\.[You say potato, we say legal issue: Developing the legal issues taxonomy \(list\) for smarter issue\-spotting and access to justice infrastructure](https://doi.org/10.2139/ssrn.6388659)\.SSRN\.Available at SSRN: https://ssrn\.com/abstract=6388659\.
- Hsia et al\. \(2025\)Jennifer Hsia, Afreen Shaikh, Zhiruo Wang, and Graham Neubig\. 2025\.[RAGGED: Towards informed design of scalable and stable RAG systems](https://proceedings.mlr.press/v267/hsia25a.html)\.In*Proceedings of the 42nd International Conference on Machine Learning*, volume 267 of*Proceedings of Machine Learning Research*, pages 24139–24155\.
- Hu et al\. \(2025\)Peng Hu, Sizhe Liu, Changjiang Gao, Xin Huang, Xue Han, Junlan Feng, Chao Deng, and Shujian Huang\. 2025\.[Large language models are cross\-lingual knowledge\-free reasoners](https://doi.org/10.18653/v1/2025.naacl-long.72)\.In*Proceedings of the North American Chapter of the Association for Computational Linguistics \(NAACL\)*, pages 1525–1542, Albuquerque, New Mexico\.
- Kanerva et al\. \(2025\)Jenna Kanerva, Cassandra Ledins, Siiri Käpyaho, and Filip Ginter\. 2025\.[OCR error post\-correction with LLMs in historical documents: No free lunches](https://aclanthology.org/2025.resourceful-1.8/)\.In*Proceedings of the Third Workshop on Resources and Representations for Under\-Resourced Languages and Domains \(RESOURCEFUL\)*, pages 38–47, Tallinn, Estonia\. University of Tartu Library\.
- Longpre et al\. \(2021\)Shayne Longpre, Kartik Perisetla, Anthony Chen, Nikhil Ramesh, Chris DuBois, and Sameer Singh\. 2021\.[Entity\-based knowledge conflicts in question answering](https://doi.org/10.18653/v1/2021.emnlp-main.565)\.In*Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing \(EMNLP\)*, pages 7052–7063\.
- Magesh et al\. \(2025\)Varun Magesh, Faiz Surani, Matthew Dahl, Mirac Suzgun, Christopher D\. Manning, and Daniel E\. Ho\. 2025\.[Hallucination\-free? assessing the reliability of leading AI legal research tools](https://doi.org/10.1111/jels.12413)\.*Journal of Empirical Legal Studies*, 22\(2\):216–242\.
- Mohan et al\. \(2026\)Varalekshmy M Mohan, Swathi Jayakumar, Gadha Saji Menon, Sachin Kurup, Veena G, and Vani Kanjirangat\. 2026\.[When retrieval hurts: Evidence utilization, script fidelity, and knowledge conflicts in multilingual RAG](https://aclanthology.org/2026.mellm-1.9/)\.In*Proceedings of the 1st Workshop on Multilinguality in the Era of Large Language Models \(MeLLM\)*, pages 92–107, San Diego, United States\.
- Narayanan and Aepli \(2024\)Manu Narayanan and Noëmi Aepli\. 2024\.[A tulu resource for machine translation](https://aclanthology.org/2024.lrec-main.155/)\.In*Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation \(LREC\-COLING\)*, pages 1756–1767, Torino, Italia\. ELRA and ICCL\.
- Qi et al\. \(2025\)Jirui Qi, Raquel Fernández, and Arianna Bisazza\. 2025\.[On the consistency of multilingual context utilization in RAG](https://doi.org/10.18653/v1/2025.mrl-main.15)\.In*Proceedings of the 5th Workshop on Multilingual Representation Learning \(MRL\)*, pages 199–225, Suzhou, China\.Best Paper Award\.
- Reuter et al\. \(2025\)Markus Reuter, Tobias Lingenberg, Rūta Liepiņa, Francesca Lagioia, Marco Lippi, Giovanni Sartor, Andrea Passerini, and Burcu Sayin\. 2025\.[Towards reliable retrieval in RAG systems for large legal datasets](https://doi.org/10.18653/v1/2025.nllp-1.3)\.In*Proceedings of the Natural Legal Language Processing Workshop \(NLLP\)*, pages 17–30, Suzhou, China\.
- Sarvam AI \(2024\)Sarvam AI\. 2024\.Sarvam\-1: A small yet powerful LLM for indic languages\.[https://www\.sarvam\.ai/blogs/sarvam\-1](https://www.sarvam.ai/blogs/sarvam-1)\.
- Sun et al\. \(2026\)Yanming Sun, Runzhe Zhan, Chi Seng Cheang, Han Wu, Xuebo Liu, Yuyao Niu, Fengying Ye, Kaixin Lan, Lidia S\. Chao, and Derek F\. Wong\. 2026\.[Exposing the cracks: Vulnerabilities of retrieval\-augmented LLM\-based machine translation](https://doi.org/10.1609/aaai.v40i39.40597)\.In*Proceedings of the AAAI Conference on Artificial Intelligence*, volume 40, pages 33135–33143\.
- Wu et al\. \(2024\)Kevin Wu, Eric Wu, and James Zou\. 2024\.[Clasheval: Quantifying the tug\-of\-war between an LLM’s internal prior and external evidence](https://arxiv.org/abs/2404.10198)\.In*Advances in Neural Information Processing Systems \(NeurIPS\), Datasets and Benchmarks Track*\.
- Xie et al\. \(2024\)Jian Xie, Kai Zhang, Jiangjie Chen, Renze Lou, and Yu Su\. 2024\.[Adaptive chameleon or stubborn sloth: Revealing the behavior of large language models in knowledge conflicts](https://arxiv.org/abs/2305.13300)\.In*International Conference on Learning Representations \(ICLR\)*\.Spotlight\.
- Xu et al\. \(2024\)Rongwu Xu, Zehan Qi, Zhijiang Guo, Cunxiang Wang, Hongru Wang, Yue Zhang, and Wei Xu\. 2024\.[Knowledge conflicts for llms: A survey](https://doi.org/10.18653/v1/2024.emnlp-main.486)\.In*Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing \(EMNLP\)*, pages 8541–8565, Miami, Florida\.
- Zeng et al\. \(2025\)Linda Zeng, Rithwik Gupta, Divij Motwani, Diji Yang, and Yi Zhang\. 2025\.[Worse than zero\-shot? a fact\-checking dataset for evaluating the robustness of RAG against misleading retrievals](https://arxiv.org/abs/2502.16101)\.In*Advances in Neural Information Processing Systems \(NeurIPS\)*\.

## Appendix AFull Query Set and Priming Glossary

### A\.1The 60\-Sentence Query Set

Table[3](https://arxiv.org/html/2608.28645#A1.T3)lists all 60 Tulu legal\-complaint sentences used in this study, with their human chosen LIST primary and secondary categories, English gloss, and Romanized\-Tulu form\. Table[4](https://arxiv.org/html/2608.28645#A1.T4)gives the corresponding Kannada script form\.

Table 3:Full 60\-sentence query set: human chosen LIST categories, English gloss, and Romanized Tulu\.IDPrimarySecondaryEnglish GlossRomanized Tulu1HousingConsumer/Money/DebtMy landlord did not give back my moneyenna dhani kasn vapas kortijer\.2Work/Employment–My employer did not pay me this monthenna dhani i tingolda sanbalan kortijer\.3Torts/Individuals–Someone broke my fence and did not fix itera enna belin hal maltd nanla sari maltijer\.4Consumer/Money/Debt–I gave someone money and they disappearedyan erega kas kori bokka akulu tappayer\.5Land/Property–My neighbor took my landenna baritakulu enna jagan detter\.6Torts/Individuals–After someone hit me with a stick my hand hurtsera enk badutu aki bokka enna kai bene avondundu\.7Work/Employment–My boss told me to leave without any reasonenna dhani karana korande yan kelasa budodu pand pander\.8Consumer/Money/Debt–The person I lent money to is not returning ityan sala kori bokka enk kas vapas korondijjer\.9Housing–Someone came into my house without my permissionenna oppige ijjande ero enna illda ulayi batter\.10Consumer/Money/Debt–They want more money than what we decidedakaleg enkulu nirnaya maltina aidd jasti kas kenonduler\.11Consumer/Money/DebtCourts/Legal\-SystemThey ended our agreement before the time we decidedakulu enkulu nirnaya maltina samayadda dunbe oppandan mugipayer\.12Courts/Legal\-SystemConsumer/Money/DebtI signed a paper but did not understand what it saidyan daskat padye, matra kakajid dada baretnd pand gottiji\.13Consumer/Money/Debt–After I signed the paper they did not do what they agreed toyan kakajid daskat padi bokka enkulu tirpu maltina vishayan akulu malpere kenondijer\.14Courts/Legal\-System–I don’t know which court I should go toenk vovu nyayalayag podu pand gottiji\.15Courts/Legal\-SystemConsumer/Money/DebtThe person who owes me money is in a different placeenadda sala detondina jana bete jagad uller\.16Torts/IndividualsHousingMy neighbor’s tree fell on my house and damaged itenna baritaklena mara enna illda mitt burda enna illan hal maltnd\.17Land/PropertyHousingSomeone is living on my land without asking and won’t leaveero enna oppige ijjande enna jagad kulondu, itte budda popujer\.18IP/Identity–Someone is selling their things using my nameero aklena samann enna pudarda maronduler\.19Work/Employment–My old employer is stopping me from starting my own workenna paratta dhani enna svanta vyaparan suru malpere budondijer\.20Estates/WillsLand/PropertyAfter my father passed away my family took the land that was meant to be given to meenna popa tir poyi bokka enk barpina jagan enna kutunbadakulu detonyer\.21HousingConsumer/Money/DebtMy landlord is trying to raise my rent before our agreement endsenklena tirpu maltina samayada dunbu enna dhani badige jasti malpere prayatna maltonduller\.22HousingTorts/IndividualsThe people next door make so much noise I cannot sleep at nightbaritakulu et bobbe padonduller panda enk ratred jeppere avondiji\.23Housing–My landlord will not fix the broken pipe in my kitchenenna adipilda hal ayina nallin enna dhani sari malpere tayar ijjer\.24Housing–I was told to leave my rented house with only two days noticeenk badigeda illan budiyere khali radd dinata vayide koryer\.25HousingConsumer/Money/DebtMy roommate moved out and left me to pay all the rentenna sahavasi ill buddu badadigen pura enna mitte padd poyer\.26Consumer/Money/Debt–The shop owner sold me a broken phone and won’t give my money backenk onji hal ayna phon\-n mard daangadida dhani kasn pira korere tayar ijjer\.27Consumer/Money/Debt–I paid for a service that was never completedyan pura kas kori bokkala belen purti maltijer\.28Consumer/Money/Debt–My brother borrowed money from me two years ago and still has not paidenna megge radd varshada dunbu ennadda sara dettonde, aven nanla pira kortije\.29Consumer/Money/Debt–The bank is charging me fees I never agreed toenna oppandina shulkag byank enna mitt kas padondundu\.30Consumer/Money/Debt–Someone convinced me to invest in something that turned out to be fakeero enan onji dongi astida mitt kas padere oppayer\.31Work/Employment–My workplace has not paid me for the extra hours I workedyan maltina jasti kelsag enk beleda jagadd kas tikkiji\.32Work/Employment–I was fired the same week I told my boss I was sickyan dhanita yan hushar ijji pandd pandina varade ar enan beledd detter\.33Work/Employment–My employer is making me work without the safety equipment I was promisedenna dhani enan suraksha dipina saman korper pand pandi bokka aven korande enada bele malpavonduller\.34Work/Employment–I was promised a promotion but someone else got it without explanationenk onji promoshan korper pandd patera koryer, matra dala karana korande bete janak koryer\.35Work/EmploymentConsumer/Money/DebtMy employer refuses to give me my final paycheck after I quityan belen budi bokka enna dhani enna akherida sanbalan korondijjer\.36Land/Property–My neighbor built a wall that crosses onto my landenna nerekaredakulu aklena belin enna jagada ullayi mutta kattader\.37Land/Property–Someone is tilling on land that belongs to my family without askingenda kenande yera enna kutunbada jagad benni dattonduller\.38Land/Property–The boundary markers on my property were moved without my knowledgeenna jagada belin yera dett budter\.39Land/PropertyEstates/WillsMy brother is claiming a part of our family’s land is his aloneenna palaye enna kutunbada jagada onji bhagan ayena matra pandd panondulle\.40Land/PropertyTorts/IndividualsSomeone cut down trees on my property without permissionenda kenande yera enna jagada maran kadter\.41Torts/Individuals–My neighbor’s dog bit my child while we were walkingenkulu nadetonduittinapaga enna baritaklena nayi enna balen cuccind\.42Torts/Individuals–Someone crashed into my parked vehicle and drove awayero enklena untondu ittina karg akd poyer\.43Torts/Individuals–A worker dropped something that injured meonji beledaye enna mitt saman budpad enk tagnd\.44Torts/Individuals–Someone spread lies about me that damaged my reputation in the villagehallid yera enna bagge sullu pandd enna pudarn hal malter\.45Torts/IndividualsLand/PropertyMy neighbor’s cattle keep destroying my cropsenna nerekareda aklena pettalu enna pura bulen tindd hal maltnd\.46Courts/Legal\-System–I received a court notice but I don’t understand what it is asking me to doenk nyayalayada onji kakaji baidnd, matra ait dada panonduller pandd gottiji\.47Courts/Legal\-System–I missed a court date because no one explained when I needed to appearenk nyayalayag popina tarikh tatnd, dayeg panda enk erela epa e podu pandd pandijer\.48Courts/Legal\-SystemWork/EmploymentI don’t know how to file a complaint against my employerenk enna kelsada dhanina mitt enca duru korodu pandd gottiji\.49Courts/Legal\-System–The police would not register my complaintpolis\-da akulu enna durun dettonere tayar ijjer\.50Courts/Legal\-System–I was told I need a lawyer but I don’t know how to find oneenk onji vakil bodu pandd pander, matra enk vakiln enca nadodu pandd gottiji\.51IP/IdentityConsumer/Money/DebtSomeone opened a loan in my name without my knowledgeenk gottijande yera enna pudard sala dettonyer\.52IP/Identity–My photo is being used in an advertisement without my permissionenada kenande yera enna photon jahiratud upayoga maltonduller\.53IP/Identity–Someone is pretending to be me online to trick my relativesyera ennaleka anlain\-d natane maltd enna kutunbadaklen dongi maltonduller\.54IP/Identity–A shop is using my family’s business name without askingvova onji angadi enna kutunbada angadida pudaran upayoga maltondundu55IP/Identity–Someone stole the design of the products my family makes and sellsenklena kutunba maltd maruna samanda disain\-n yera kandder\.56Estates/Wills–My father died without a will and my siblings are dividing everything without meenna popa onji vil budande tir poyi bokka enna megge\-palayenakulu enan buddu astin pura pattonduller\.57Estates/Wills–My mother’s jewelry was promised to me but my brother took it after she diedenna ammana bangar enk korper pand pandi bokka amma tir poyi bokka enna palaye aven pura dettonye\.58Estates/WillsLand/PropertyI found out my parents’ house was sold after my father’s death without our agreementenna popa tir poyi bokka enna oppige ijjande enna popa\-ammana illan maryer pandd gottand\.59Estates/WillsLand/PropertyMy uncle is refusing to share the property left by my grandfatherenna mamu, ajjer\-na jagan pattere kenondijjer\.60Estates/WillsHousingAfter my husband passed away his family is not letting me stay in our houseenna kandani tir poyi bokka enna kandanina kutunbadakulu enan enklena illad kulyere budondijjer\.Table 4:Full 60\-sentence query set in native script \(Kannada\)\.ID\\kannadafontKannada1\\kannadafontಎನ್ನ ಧನಿ ಕಾಸ್ನ್ ವಾಪಸ್ ಕೊರ್ತಿಜೇರ್\.2\\kannadafontಎನ್ನ ಧನಿ ಈ ತಿಂಗೋಲ್ದ ಸಂಬಳನ್ ಕೊರ್ತಿಜೇರ್\.3\\kannadafontಏರಾ ಎನ್ನ ಬೇಲಿನ್ ಹಾಳ್ ಮಲ್ತ್ದ್ ನನ್ಲಾ ಸರಿ ಮಲ್ತಿಜೇರ್\.4\\kannadafontಯಾನ್ ಏರೇಗಾ ಕಾಸ್ ಕೊರಿ ಬೊಕ್ಕ ಅಕುಲು ತಪ್ಪಾಯೇರ್\.5\\kannadafontಎನ್ನ ಬರೀತಕುಲು ಎನ್ನ ಜಾಗನ್ ದೆತ್ತೇರ್\.6\\kannadafontಏರಾ ಎಂಕ್ ಬಡುಟು ಆಕಿ ಬೊಕ್ಕ ಎನ್ನ ಕೈ ಬೇನೆ ಆವೊಂದುಂಡು\.7\\kannadafontಎನ್ನ ಧನಿ ಕಾರಣ ಕೊರಂದೆ ಯಾನ್ ಕೆಲಸ ಬುಡೊಡು ಪಂದ್ ಪಂಡೆರ್\.8\\kannadafontಯಾನ್ ಸಾಲ ಕೊರಿ ಬೊಕ್ಕ ಎಂಕ್ ಕಾಸ್ ವಾಪಾಸ್ ಕೊರೋಂದಿಜ್ಜೆರ್\.9\\kannadafontಎನ್ನ ಓಪ್ಪಿಗೆ ಇಜ್ಜಂದೆ ಏರೋ ಎನ್ನ ಇಲ್ಲ್ದ ಉಲಯಿ ಬತ್ತೇರ್\.10\\kannadafontಅಕಲೆಗ್ ಎಂಕುಲು ನಿರ್ಣಯ ಮಲ್ತಿನ ಐಡ್ದ್ ಜಾಸ್ತಿ ಕಾಸ್ ಕೇನೋಂದುಲೇರ್\.11\\kannadafontಅಕುಲು ಎಂಕುಲು ನಿರ್ಣಯ ಮಲ್ತಿನ ಸಮಯಡ್ದ ದುಂಬೇ ಓಪ್ಪಂದನ್ ಮುಗಿಪಾಯೇರ್\.12\\kannadafontಯಾನ್ ದಸ್ಕತ್ ಪಾಡ್ಯೇ, ಮಾತ್ರ ಕಾಕಜಿಡ್ ದಾದಾ ಬರೆತ್ಂಡ್ ಪಂದ್ ಗೊತ್ತಿಜಿ\.13\\kannadafontಯಾನ್ ಕಾಕಜಿಡ್ ದಸ್ಕತ್ ಪಾಡಿ ಬೊಕ್ಕ ಎಂಕುಲು ತಿರ್ಪು ಮಲ್ತಿನ ವಿಷಯನ್ ಅಕುಲು ಮಲ್ಪೇರೆ ಕೇನೋಂದಿಜೇರ್\.14\\kannadafontಎಂಕ್ ವೋವೂ ನ್ಯಾಯಾಲಯಗ್ ಪೋಡು ಪಂದ್ ಗೊತ್ತಿಜಿ\.15\\kannadafontಎನಡ್ದ ಸಾಲ ದೆತೋಂದಿನ ಜನ ಬೇತೆ ಜಾಗಡ್ ಉಲ್ಲೇರ್\.16\\kannadafontಎನ್ನ ಬರೀತಕ್ಲೇನ ಮರ ಎನ್ನ ಇಲ್ಲ್ದ ಮಿತ್ತ್ ಬೂರ್ದ ಎನ್ನ ಇಲ್ಲನ್ ಹಾಳ್ ಮಲ್ತ್ಂಡ್\.17\\kannadafontಏರೋ ಎನ್ನ ಓಪ್ಪಿಗೆ ಇಜ್ಜಂದೆ ಎನ್ನ ಜಾಗಡ್ ಕುಲೋಂದು, ಇತ್ತೇ ಬುಡ್ದ ಪೋಪುಜೇರ್\.18\\kannadafontಏರೋ ಅಕ್ಲೇನ ಸಾಮಾನ್ನ್ ಎನ್ನ ಪುದರ್ಡ ಮಾರೋಂದುಲೇರ್\.19\\kannadafontಎನ್ನ ಪರಾತ್ತ ಧನಿ ಎನ್ನ ಸ್ವಂತ ವ್ಯಾಪಾರನ್ ಸುರು ಮಲ್ಪೆರೆ ಬುಡೋಂದಿಜೇರ್\.20\\kannadafontಎನ್ನ ಪೋಪ ತೀರ್ ಪೋಯಿ ಬೊಕ್ಕ ಎಂಕ್ ಬರ್ಪಿನ ಜಾಗನ್ ಎನ್ನ ಕುಟುಂಬದಕುಲು ದೇತೋಣ್ಯೇರ್\.21\\kannadafontಎಂಕ್ಲೇನ ತಿರ್ಪು ಮಲ್ತಿನ ಸಮಯದ ದುಂಬು ಎನ್ನ ಧನಿ ಬಾಡಿಗೆ ಜಾಸ್ತಿ ಮಲ್ಪೆರೆ ಪ್ರಯತ್ನ ಮಲ್ತೋಂದುಲ್ಲೇರ್\.22\\kannadafontಬರೀತಕುಲು ಏತ್ ಬೊಬ್ಬೆ ಪಾಡೋಂದುಲ್ಲೇರ್ ಪನ್ಡ ಎಂಕ್ ರಾತ್ರೆಡ್ ಜೆಪ್ಪೆರೇ ಆವೊಂದಿಜಿ\.23\\kannadafontಎನ್ನ ಅಡಿಪಿಲ್ದಾ ಹಾಳ್ ಆಯಿನ ನಳ್ಳಿನ್ ಎನ್ನ ಧನಿ ಸರಿ ಮಲ್ಪೆರೆ ತಯಾರ್ ಇಜ್ಜೆರ್\.24\\kannadafontಎಂಕ್ ಬಾಡಿಗೆದ ಇಲ್ಲನ್ ಬುಡಿಯೆರೆ ಖಾಲಿ ರಡ್ಡ್ ದಿನತ ವಯಿದೇ ಕೊರ್ಯೇರ್\.25\\kannadafontಎನ್ನ ಸಹವಾಸಿ ಇಲ್ಲ್ ಬುಡ್ದು ಬಾದಡಿಗೆನ್ ಪುರಾ ಎನ್ನ ಮಿತ್ತೆ ಪಾಡ್ದ್ ಪೋಯೇರ್\.26\\kannadafontಎಂಕ್ ಒಂಜಿ ಹಾಳ್ ಆಯ್ನ ಫೋನ್\-ನ್ ಮಾರ್ದ್ ದಅಂಗಡಿದ ಧನಿ ಕಾಸ್ನ್ ಪಿರಾ ಕೊರೆರೆ ತಯಾರ್ ಇಜ್ಜೆರ್\.27\\kannadafontಯಾನ್ ಪೂರಾ ಕಾಸ್ ಕೊರಿ ಬೊಕ್ಕಲಾ ಬೇಲೆನ್ ಪೂರ್ತಿ ಮಲ್ತಿಜೆರ್\.28\\kannadafontಎನ್ನ ಮೆಗ್ಗೆ ರಡ್ಡ್ ವರ್ಷದ ದುಂಬು ಎನ್ನಡ್ದ ಸಾರ ದೆತ್ತೊಂಡೆ, ಅವೆನ್ ನನ್ಲಾ ಪಿರಾ ಕೊರ್ತಿಜೆ\.29\\kannadafontಎನ್ನ ಒಪ್ಪಂದಿನ ಶುಲ್ಕಗ್ ಬ್ಯಾಂಕ್ ಎನ್ನ ಮಿತ್ತ್ ಕಾಸ್ ಪಾಡೊಂದುಂಡು\.30\\kannadafontಏರೊ ಎನನ್ ಒಂಜಿ ಡೋಂಗಿ ಆಸ್ತಿದ ಮಿತ್ತ್ ಕಾಸ್ ಪಾಡೆರೆ ಒಪ್ಪಾಯೆರ್\.31\\kannadafontಯಾನ್ ಮಲ್ತಿನ ಜಾಸ್ತಿ ಕೆಲ್ಸಗ್ ಎಂಕ್ ಬೇಲೆದ ಜಾಗಡ್ಡ್ ಕಾಸ್ ತಿಕ್ಕಿಜಿ\.32\\kannadafontಯಾನ್ ಧನಿಟ ಯಾನ್ ಹುಷಾರ್ ಇಜ್ಜಿ ಪಂಡ್ಡ್ ಪಂಡಿನ ವಾರಡೆ ಆರ್ ಎನನ್ ಬೇಲೆಡ್ಡ್ ದೆತ್ತೆರ್\.33\\kannadafontಎನ್ನ ಧನಿ ಎನನ್ ಸುರಕ್ಷ ದೀಪಿನ ಸಾಮಾನ್ ಕೊರ್ಪೆರ್ ಪಂದ್ ಪಂಡಿ ಬೊಕ್ಕ ಅವೆನ್ ಕೊರಂದೇ ಎನಡ ಬೇಲೆ ಮಲ್ಪಾವೊಂದುಲ್ಲೆರ್\.34\\kannadafontಎಂಕ್ ಒಂಜಿ ಪ್ರೊಮೋಷನ್ ಕೊರ್ಪೆರ್ ಪಂಡ್ಡ್ ಪಾತೇರ ಕೊರ್ಯೆರ್, ಮಾತ್ರ ದಾಲ ಕಾರಣ ಕೊರಂದೇ ಬೇತೆ ಜನಕ್ ಕೊರ್ಯೆರ್\.35\\kannadafontಯಾನ್ ಬೇಲೆನ್ ಬುಡಿ ಬೊಕ್ಕ ಎನ್ನ ಧನಿ ಎನ್ನ ಅಖೇರಿದ ಸಂಬಳನ್ ಕೊರೋಂದಿಜ್ಜೆರ್\.36\\kannadafontಎನ್ನ ನೆರೆಕರೆದಕುಲು ಅಕ್ಲೆನ ಬೇಲಿನ್ ಎನ್ನ ಜಾಗದ ಉಲ್ಲಯಿ ಮುಟ್ಟ ಕಟ್ಟಾದೆರ್\.37\\kannadafontಎನ್ಡ ಕೇಣಂದೇ ಯೇರಾ ಎನ್ನ ಕುಟುಂಬದ ಜಾಗಡ್ ಬೆನ್ನಿ ದಟ್ಟೊಂದುಲ್ಲೆರ್\.38\\kannadafontಎನ್ನ ಜಾಗದ ಬೇಲಿನ್ ಯೇರಾ ದೆತ್ತ್ ಬುಡ್ತೆರ್\.39\\kannadafontಎನ್ನ ಪಲಯೆ ಎನ್ನ ಕುಟುಂಬದ ಜಾಗದ ಒಂಜಿ ಭಾಗನ್ ಆಯೇನ ಮಾತ್ರ ಪಂಡ್ಡ್ ಪಣೊಂದುಲ್ಲೆ\.40\\kannadafontಎನ್ಡ ಕೇಣಂದೇ ಯೇರಾ ಎನ್ನ ಜಾಗದ ಮರನ್ ಕಡ್ತೆರ್\.41\\kannadafontಎಂಕುಲು ನಡೆತೊಂದುಿತ್ತಿನಪಗ ಎನ್ನ ಬರಿತ್ಅಕ್ಲೆನ ನಾಯಿ ಎನ್ನ ಬಾಲೆನ್ ಚೂಚ್ಚಿಂಡ್\.42\\kannadafontಏರೊ ಎಂಕ್ಲೆನ ಉಂತೋಂದು ಇತ್ತಿನ ಕಾರ್ಗ್ ಆಕ್ದ್ ಪೋಯೆರ್\.43\\kannadafontಒಂಜಿ ಬೇಲೆದಾಯೆ ಎನ್ನ ಮಿತ್ತ್ ಸಾಮಾನ್ ಬೂಡ್ಪಾದ್ ಎಂಕ್ ತಾಗ್ಂಡ್\.44\\kannadafontಹಳ್ಳಿಡ್ ಯೇರಾ ಎನ್ನ ಬಾಗ್ಗೆ ಸುಳ್ಳು ಪಂಡ್ಡ್ ಎನ್ನ ಪುದರ್ನ್ ಹಾಳ್ ಮಲ್ತೆರ್\.45\\kannadafontಎನ್ನ ನೆರೆಕರೆದ ಅಕ್ಲೆನ ಪೆತ್ತಲು ಎನ್ನ ಪೂರಾ ಬುಳೆನ್ ತಿಂದ್ಡ್ ಹಾಳ್ ಮಲ್ತ್ಂಡ್\.46\\kannadafontಎಂಕ್ ನ್ಯಾಯಾಲಯದ ಒಂಜಿ ಕಾಕಜಿ ಬೈದ್ಂಡ್, ಮಾತ್ರ ಐಟ್ ದಾದ ಪಣೊಂದುಲ್ಲೆರ್ ಪಂಡ್ಡ್ ಗೊತ್ತಿಜಿ\.47\\kannadafontಎಂಕ್ ನ್ಯಾಯಾಲಯಗ್ ಪೋಪಿನ ತಾರೀಖ್ ತತ್ಂಡ್, ದಾಯೆಗ್ ಪಂಡ ಎಂಕ್ ಏರೇಲಾ ಏಪ ಏ ಪೋಡು ಪಂಡ್ಡ್ ಪಂಡಿಜೆರ್\.48\\kannadafontಎಂಕ್ ಎನ್ನ ಕೆಲ್ಸದ ಧನಿನ ಮಿತ್ತ್ ಎಂಚ ದೂರು ಕೊರೋಡು ಪಂಡ್ಡ್ ಗೊತ್ತಿಜಿ\.49\\kannadafontಪೊಲೀಸ್\-ದ ಅಕುಲು ಎನ್ನ ದೂರುನ್ ದೆತ್ತೊನೆರೆ ತಯಾರ್ ಇಜ್ಜೆರ್\.50\\kannadafontಎಂಕ್ ಒಂಜಿ ವಕೀಲ್ ಬೋಡು ಪಂಡ್ಡ್ ಪಂಡೆರ್, ಮಾತ್ರ ಎಂಕ್ ವಕೀಲ್ನ್ ಎಂಚ ನಾಡೋಡು ಪಂಡ್ಡ್ ಗೊತ್ತಿಜಿ\.51\\kannadafontಎಂಕ್ ಗೊತ್ತಿಜಂದೇ ಯೇರಾ ಎನ್ನ ಪುದರ್ಡ್ ಸಾಲ ದೆತ್ತೊನ್ಯೆರ್\.52\\kannadafontಎನಡ ಕೇನಂದೇ ಯೇರಾ ಎನ್ನ ಫೋಟೋನ್ ಜಾಹೀರಾತುಡ್ ಉಪಯೋಗ ಮಲ್ತೊಂದುಲ್ಲೆರ್\.53\\kannadafontಯೇರಾ ಎನ್ನಲೆಕ ಆನ್‌ಲೈನ್\-ಡ್ ನಟನೆ ಮಲ್ತ್ಡ್ ಎನ್ನ ಕುಟುಂಬದಕ್ಲೆನ್ ಡೋಂಗಿ ಮಲ್ತೊಂದುಲ್ಲೆರ್\.54\\kannadafontವೊವಾ ಒಂಜಿ ಅಂಗಡಿ ಎನ್ನ ಕುಟುಂಬದ ಅಂಗಡಿದ ಪುದರನ್ ಉಪಯೋಗ ಮಲ್ತೊಂದುಂಡು55\\kannadafontಎಂಕ್ಲೆನ ಕುಟುಂಬ ಮಲ್ತ್ಡ್ ಮಾರುನ ಸಾಮಾಂದ ಡಿಸೈನ್\-ನ್ ಯೇರಾ ಕಂಡ್ದೆರ್\.56\\kannadafontಎನ್ನ ಪೋಪ ಒಂಜಿ ವಿಲ್ ಬುಡಂದೇ ತೀರ್ ಪೋಯಿ ಬೊಕ್ಕ ಎನ್ನ ಮೆಗ್ಗೆ\-ಪಲಯೆನಕುಲು ಎನನ್ ಬುಡ್ಡು ಆಸ್ತಿನ್ ಪೂರಾ ಪಟ್ಟೋಂದುಲ್ಲೆರ್\.57\\kannadafontಎನ್ನ ಅಮ್ಮನ ಬಂಗಾರ್ ಎಂಕ್ ಕೊರ್ಪೆರ್ ಪಂಡ್ ಪಂಡಿ ಬೊಕ್ಕ ಅಮ್ಮ ತೀರ್ ಪೋಯಿ ಬೊಕ್ಕ ಎನ್ನ ಪಲಯೆ ಅವೆನ್ ಪೂರಾ ದೆತ್ತೊನ್ಯೆ\.58\\kannadafontಎನ್ನ ಪೋಪ ತೀರ್ ಪೋಯಿ ಬೊಕ್ಕ ಎನ್ನ ಒಪ್ಪಿಗೆ ಇಜ್ಜಂದೆ ಎನ್ನ ಪೋಪ\-ಅಮ್ಮನ ಇಲ್ಲನ್ ಮಾರ್ಯೆರ್ ಪಂಡ್ಡ್ ಗೊತ್ತಾಂಡ್\.59\\kannadafontಎನ್ನ ಮಾಮು, ಅಜ್ಜೆರ್\-ನ ಜಾಗನ್ ಪಟ್ಟೆರೆ ಕೇನೋಂದಿಜ್ಜೆರ್\.60\\kannadafontಎನ್ನ ಕಂಡನಿ ತೀರ್ ಪೋಯಿ ಬೊಕ್ಕ ಎನ್ನ ಕಂಡನಿನ ಕುಟುಂಬದಕುಲು ಎನನ್ ಎಂಕ್ಲೆನ ಇಲ್ಲಡ್ ಕುಲ್ಯೆರೆ ಬುಡೋಂದಿಜ್ಜೆರ್\.
### A\.2Priming Glossary

The priming glossary \(Table[5](https://arxiv.org/html/2608.28645#A1.T5)and Table[6](https://arxiv.org/html/2608.28645#A1.T6)\) lists the vocabulary items used to construct the sparse and extensive priming blocks \(§[3\.3](https://arxiv.org/html/2608.28645#S3.SS3)\)\.Safe\-tieritems are grammatical/functional vocabulary with no category\-revealing content;leaky\-tieritems are content words \(e\.g\. “landlord,” “money,” “land”\) that overlap with terms likely to appear in specific LIST categories and are the vocabulary most responsible for the jump in accuracy under extensive priming \(§[5\.1](https://arxiv.org/html/2608.28645#S5.SS1)\)\.

Table 5:Safe\-tier glossary items \(grammatical/function vocabulary, no category\-revealing content\)\.\\kannadafont\\kannadafontKannadaRomanizedMeaning\\kannadafontಆಂಜೊaanjoman\\kannadafontಆಲ್alshe\\kannadafontಬಾಲೆbaalechild\\kannadafontಮೀನ್meenfish\\kannadafontಪೋಪಿನಿpopinito go\\kannadafontಮಲ್ಪುನಿmalpunito do/make\\kannadafontಕೊರ್ಪುನಿkorpunito give\\kannadafontಗೊತ್ತುgothuto know\\kannadafontಬೋಡುboduto want\\kannadafontವಾಪಸ್wapasreturn/back\\kannadafontಮುಗ್ಪಾವುನಿmugpavunito end/finish\\kannadafontಇಜ್ಜಿijjinegation\(not\)\\kannadafontಅತ್ತ್attnegative\\kannadafontಬೊಕ್ಕbocaafter\\kannadafontದುಂಬುdumbubefore\\kannadafontಮಾತ್ರmatrabut/contrast\\kannadafontಪಂಡpandaquotative\(THAT\)\\kannadafontಉಂಡುunduexists/is\\kannadafontಅಮ್ಮammamotherTable 6:Leaky\-tier glossary items \(content vocabulary overlapping with LIST category terms\)\.\\kannadafont\\kannadafontKannadaRomanizedMeaning\\kannadafontಯಜಮಾನyejamaanalandlord/owner\\kannadafontಧನಿdhanilandlord/employer \(authority figure over you\)\\kannadafontಕಾಸುkasmoney\\kannadafontಜಾಗjaggaland\\kannadafontಇಲ್ಲ್illehouse\\kannadafontಬೇಲಿbeleefence\\kannadafontನೆರೆಕರೆnerakareneighbor\\kannadafontಪುದರ್pudarname\\kannadafontಕಾಕಜಿkakajidocument\\kannadafontಕುಟುಂಬkotumbafamily\\kannadafontಸಾಲsalaloan/debt\\kannadafontಓಪ್ಪಿಗೆoppigaypermission

## Appendix BPrompt Template

Every model call in this study, across all 19 conditions and all three models, was assembled by a single unified prompt\-builder function shared across the whole pipeline \(build\_unified\_prompt\), so that script, priming tier, and RAG inclusion could vary independently without a separate, drifting prompt implementation for each condition\. Each call is fully stateless: no conversation history or KV\-cache continuity carries over between queries \(Section[5\.4](https://arxiv.org/html/2608.28645#S5.SS4)\)\. The full prompt sent to the model is assembled from up to five parts, always in this order:

1. 1\.Astatic prefix, identical across all 19 conditions: system role, LIST category definitions, an anti\-collapse instruction, and two few\-shot calibration examples\.
2. 2\.An optionalpriming block\(sparse\- or full\-tier conditions only\), giving a script\-specific Tulu vocabulary guide\.
3. 3\.Thesituation block: the query itself, with a script\-specific framing sentence\.
4. 4\.For RAG conditions only, up to threeretrieved passagesfrom the Kannada court corpus\.
5. 5\.Output instructions: a three\-step reasoning scaffold ending in a top\-3 JSON prediction\.

### B\.1Static Prefix \(all 19 conditions\)

> You are a legal AI assistant helping classify everyday legal situations described by lay speakers into legal issue categories\. The legal categories are: \- Housing: landlord\-tenant disputes, security deposits not returned, trespass into a home or dwelling, damage to rented property, eviction, someone entering your house without permission\. Examples: ‘My landlord did not give back my deposit’, ‘Someone came into my house without asking’\. \- Consumer/Money/Debt: money owed but not paid, loans not repaid, overcharging beyond what was agreed, contract breaches involving payment, fraud where money was taken and the person disappeared\. Examples: ‘I lent money and they won’t return it’, ‘They charged more than we agreed’, ‘I gave someone money and they disappeared’, ‘They ended our agreement early’\. \- Work/Employment: wages not paid, wrongful dismissal without reason, employer stopping you from starting your own business, non\-compete situations\. Examples: ‘My employer did not pay me this month’, ‘My boss told me to leave without any reason’, ‘My old employer is stopping me from starting my own work’\. \- Land/Property: neighbor taking your land, someone occupying your land without permission and refusing to leave, boundary disputes, land stolen after inheritance\. Examples: ‘My neighbor took my land’, ‘Someone is living on my land without asking and won’t leave’\. \- Torts/Individuals: personal injury caused by another person, property damage caused by another person’s negligence, someone breaking your property and not fixing it, a neighbor’s tree falling on your house\. Examples: ‘Someone hit me with a stick and my hand hurts’, ‘Someone broke my fence and did not fix it’, ‘My neighbor’s tree fell on my house’\. \- Courts/Legal\-System: confusion about which court to go to, not knowing where to file a complaint, the other party being in a different location or jurisdiction, signing a document without understanding what it said\. Examples: ‘I don’t know which court I should go to’, ‘The person who owes me money is in a different place’, ‘I signed a paper but did not understand what it said’\. \- IP/Identity: someone using your name without permission, someone selling things under your name or brand, misuse of personal or business identity\. Examples: ‘Someone is selling their things using my name’\. \- Estates/Wills: inheritance disputes after a family member’s death, family members taking property that was meant to be given to you, land or assets disputed after someone passes away\. Examples: ‘After my father passed away my family took the land that was meant for me’\. IMPORTANT: These categories describe EVERYDAY PROBLEMS, not commercial contract clauses\. Classify based on what the PERSON’S SITUATION is about, NOT based on legal terminology in retrieved passages\. Do not default to the same category repeatedly\. Weigh each situation against ALL categories before deciding\. Here are two example classifications to calibrate your reasoning: Example 1: Situation: "My landlord kept my deposit money after I moved out\." Classification reasoning: The person had a rental relationship \(Housing\) and the landlord failed to return money \(Consumer/Money/Debt\)\. The primary issue is the landlord\-tenant relationship and deposit \-\- Housing\. \{"category": "Housing", "confidence": 0\.85, "reasoning": "Landlord\-tenant deposit dispute"\} Example 2: Situation: "Someone broke my fence and refused to fix it\." Classification reasoning: This is damage caused by another individual to my property\. It is not a housing/landlord issue \(no tenancy\)\. It is not a debt \(no money owed from an agreement\)\. It is physical harm/damage caused by another person \-\- Torts/Individuals\. \{"category": "Torts/Individuals", "confidence": 0\.80, "reasoning": "Property damage caused by another person’s action"\}

### B\.2Priming Block \(sparse/full\-tier conditions only\)

For conditions with a priming tier, this block is appended immediately after the static prefix\. Sparse tier draws only from the “safe” \(grammatical function vocabulary\) items in the glossary \(Appendix[A\.2](https://arxiv.org/html/2608.28645#A1.SS2)\); full tier draws from the entire glossary, safe and leaky tiers combined\. The template \(script\-agnostic:\{SCRIPT\}and eachword=meaningpair change per script and tier\):

> Vocabulary guide for Tulu in \{SCRIPT\} script: Nouns: \{word=meaning, word=meaning, \.\.\.\} Legal nouns: \{word=meaning, \.\.\.\}\(full tier only\) Verbs: \{word=meaning, \.\.\.\} Grammar: \{word=meaning, \.\.\.\} SOV word order \-\- verb always at end\. Suffix \-g/\-k = ‘to/towards’\. Suffix \-n/\-d = ‘of/belonging to’\. Quotative marks embedded speech\.

### B\.3Situation Block

The framing sentence changes by script; the closing instruction changes by whether a priming block was included\.

> English:SITUATION \(in English\): "\{display\_text\}" Romanized:SITUATION \(Tulu, romanized\): "\{display\_text\}" Kannada:SITUATION \(Tulu written in Kannada script\): "\{display\_text\}" \-\- closing instruction, with priming: No translation available\. Use the vocabulary guide above and any words you recognize in the query to understand the situation\. \-\- closing instruction, without priming: No translation available\. Use any patterns you can detect in the text\.

### B\.4Retrieved\-Passage Block \(RAG conditions only\)

> The passages below are BACKGROUND LAW ONLY\. Do NOT classify what the passage is about \-\- classify what the PERSON’S SITUATION is about\. Retrieved passages from the Kannada court corpus: Passage 1 \(similarity: \{score\}, source: \{case\_name\}, LIST categories for this case: \{labels\}\): \{passage text\} Passage 2 \.\.\. Passage 3 \.\.\.

Passage text and scores shown here are template placeholders, not a fabricated example; see the methodological note in Appendix[C](https://arxiv.org/html/2608.28645#A3)regarding retrieved\-passage preservation for this batch\.

### B\.5Output Instructions \(all conditions\)

> Step 1: Re\-read the SITUATION above\. In one sentence, describe in plain English what problem this person is facing \-\- ignore the passages for now\. Step 2: Based on YOUR description of the situation \(not the passages\), identify the best\-fitting LIST category and explain why\. Rule out at least two other categories explicitly\. Step 3: Provide your TOP 3 category predictions in order, as JSON: ``` ‘‘‘json { "top1": { "category": "...", "confidence": 0.0-1.0, "reasoning": "..." }, "top2": { "category": "...", "confidence": 0.0-1.0, "reasoning": "..." }, "top3": { "category": "...", "confidence": 0.0-1.0, "reasoning": "..." } } ‘‘‘ ``` CRITICAL: Every category value MUST be exactly one of: \- Housing \- Consumer/Money/Debt \- Work/Employment \- Land/Property \- Torts/Individuals \- Courts/Legal\-System \- IP/Identity \- Estates/Wills Do NOT invent new category names or use terms from the passages\.

### B\.6Complete Worked Example

To ground the template above in an actual assembled call rather than placeholders, this is the complete, real prompt for Condition E\_full \(Kannada script, full priming, no RAG\) on query 1 \(human chosen label: Housing\) – static prefix \(Appendix[B\.1](https://arxiv.org/html/2608.28645#A2.SS1)\) followed by the full\-tier Kannada priming block \(built from the real glossary in Appendix[A\.2](https://arxiv.org/html/2608.28645#A1.SS2)\) and situation block, exactly as sent to the model:

> \[ \.\.\.static prefix as in Appendix[B\.1](https://arxiv.org/html/2608.28645#A2.SS1)\.\.\. \] Vocabulary guide for Tulu in Kannada script: Nouns:\\kannadafontಆಂಜೊ=man,\\kannadafontಆಲ್=she,\\kannadafontಬಾಲೆ=child,\\kannadafontಮೀನ್=fish,\\kannadafontಅಮ್ಮ=mother Legal nouns:\\kannadafontಯಜಮಾನ=landlord/owner,\\kannadafontಧನಿ=landlord/employer \(authority figure over you\),\\kannadafontಕಾಸು=money,\\kannadafontಜಾಗ=land,\\kannadafontಇಲ್ಲ್=house,\\kannadafontಬೇಲಿ=fence,\\kannadafontನೆರೆಕರೆ=neighbor,\\kannadafontಪುದರ್=name,\\kannadafontಕಾಕಜಿ=document,\\kannadafontಕುಟುಂಬ=family,\\kannadafontಸಾಲ=loan/debt,\\kannadafontಓಪ್ಪಿಗೆ=permission,\\kannadafontದಸ್ಕತ್=signature Verbs:\\kannadafontಪೋಪಿನಿ=to go,\\kannadafontಮಲ್ಪುನಿ=to do/make,\\kannadafontಕೊರ್ಪುನಿ=to give,\\kannadafontಗೊತ್ತು=to know,\\kannadafontಬೋಡು=to want,\\kannadafontವಾಪಸ್=return/back,\\kannadafontಮುಗ್ಪಾವುನಿ=to end/finish Grammar:\\kannadafontಇಜ್ಜಿ=negation\(not\),\\kannadafontಅತ್ತ್=negative,\\kannadafontಬೊಕ್ಕ=after,\\kannadafontದುಂಬು=before,\\kannadafontಮಾತ್ರ=but/contrast,\\kannadafontಪಂಡ=quotative\(THAT\),\\kannadafontಉಂಡು=exists/is SOV word order \-\- verb always at end\. Suffix \-g/\-k = ‘to/towards’\. Suffix \-n/\-d = ‘of/belonging to’\. Quotative marks embedded speech\. \-\-\- CLASSIFY THIS SITUATION \-\-\- SITUATION \(Tulu written in Kannada script\): "\\kannadafontಎನ್ನ ಧನಿ ಕಾಸ್ನ್ ವಾಪಸ್ ಕೊರ್ತಿಜೇರ್\." No translation available\. Use the vocabulary guide above and any words you recognize in the query to understand the situation\. No retrieved passages are provided for this classification\. Classify based solely on the situation description and the category definitions above\. REMINDER \-\- classify this situation: SITUATION \(Tulu written in Kannada script\): "\\kannadafontಎನ್ನ ಧನಿ ಕಾಸ್ನ್ ವಾಪಸ್ ಕೊರ್ತಿಜೇರ್\." No translation available\. Use the vocabulary guide above and any words you recognize in the query to understand the situation\. \[ \.\.\.output instructions as in Appendix[B\.5](https://arxiv.org/html/2608.28645#A2.SS5)\.\.\. \]

## Appendix CRepresentative RAG Reasoning\-Trace Examples

### C\.1Fact Substitution

> Model:Llama3Condition:B \(Romanized Tulu, no priming, RAG\) Query \(gloss\):“Someone cut down trees on my property without permission” Human chosen label:Land/Property Predicted:Consumer/Money/Debt Reasoning trace:“The situation appears to be related to a financial issue, specifically a debt or payment problem, given the mention of insufficient funds and a check being returned\.” Verification:The third retrieved passage was drawn fromSridhar P\. v\. Ramesh, a cheque\-bounce dispute containing Section 138 Negotiable Instruments Act language and explicit references to a returned cheque \(\\kannadafontಚೆಕ್\)\. Neither cheques nor money are mentioned anywhere in the query, which concerns unauthorized tree\-felling on the speaker’s land; the retrieved passage’s content was substituted directly into the reasoning in place of the query’s actual content\.

> Model:SarvamCondition:C\-sparse \(Romanized Tulu, sparse priming, RAG\) Query \(gloss\):“My roommate moved out and left me to pay all the rent” Human chosen label:Housing Predicted:Courts/Legal\-System Reasoning trace:“The person is actively involved in a legal process, mentioning filing a document, producing evidence \(WhatsApp screenshots\), and undergoing a court procedure \(fingerprinting\)\.” Verification:The first retrieved passage \(Amith Ramachandra v\. Shantha\) contains both details verbatim:\\kannadafontವಾಟ್ಸಾಪ್ ಸಂದೇಶಗಳ ಸ್ಕ್ರೀನ್ ಶಾಟ್ \(WhatsApp message screenshots\), and ಬೆರಳಚ್ಚು \(typing, which the model misunderstood as fingerprinting\)\. Nothing resembling either detail appears in the query \(a roommate/rent dispute\)\.

### C\.2Confabulation

> Model:Llama3Condition:D\-full \(Kannada script, full priming, RAG\) Query \(gloss\):“The people next door make so much noise I cannot sleep at night” Human chosen label:Housing Predicted:Courts/Legal\-System Reasoning trace:“The situation involves threats and demands, which suggests that legal action may be involved or necessary to resolve the issue\.” Verification:Neither “threats” nor “demands” \(nor Kannada equivalents\\kannadafontಬೆದರಿಕೆ /\\kannadafontಬೇಡಿಕೆ\) appear in the query or in any of the three retrieved passages \(Amith Ramachandra v\. Shantha;Sadashiva Shetty v\. Vijaya Bank;State v\. Sudhakar Pai\) \(though threats appear in theAmith Ramachandra v\. Shanthatext in a seperate chunk\) \. The reasoning invents a threat scenario with no basis anywhere in the available context\.

> Model:Llama3Condition:F \(Kannada script, no priming, RAG\) Query \(gloss\):“My neighbor’s dog bit my child while we were walking” Human chosen label:Torts/IndividualsPredicted:Estates/Wills Reasoning trace:“The situation involves eviction and family members, which suggests an issue related to property or inheritance\.” Verification:No eviction\-related term \(\\kannadafontಖಾಲಿ,\\kannadafontಉಚ್ಚಾಟನೆ\) appears in the query or in any of the three retrieved passages \(Nikitha A\. v\. Santhosh Ghatke;Akkamma v\. Ramesh Patil;Dr\. Alfred Aranha v\. Sahel Peter Aranha\)\. “Eviction” is introduced with no basis in either source, though family members are brought up in bothNikitha A\. v\. Santhosh GhatkeandAkkamma v\. Ramesh Patil\.

### C\.3RAG Content and Correct Classification Aligning

> Model:SarvamCondition:F \(Kannada script, no priming, RAG\) Query \(gloss\):“My neighbor’s dog bit my child while we were walking” Human chosen label:Torts/Individuals Predicted:Torts/Individuals Reasoning trace:“This is a personal injury case caused by another person’s \(or their animal’s\) negligence or action, which falls under the Torts/Individuals category\.” Verification:The retrieved passage \(Nikitha A\. v\. Santhosh Ghatke, similarity 0\.916\) is actually on track with Torts/Individuals content\. Llama3 confabulates on this same query under the same condition \(see above\), predicting Estates/Wills with reasoning citing “eviction and family members”\. The relevance of a retrieved passage does not ensure accurate classification\.

> Model:Hex\-1Condition:A \(English gloss, RAG\) Query \(gloss\):“After my father passed away my family took the land that was meant to be given to me” Human chosen label:Estates/Wills Predicted:Estates/Wills Reasoning trace:“The situation involves inheritance disputes after a family member’s death\.” Verification:The retrieved passage \(Gagan B\.S\. v\. Shivamurthy T\., similarity 0\.824\) is itself an ancestral\-property inheritance\-share dispute \(\\kannadafontಪಿತ್ರಾಜರ್ಿತ ಆಸ್ತಿ, “ancestral property”\)\. Retrieval and classification align correctly here as well\.

### C\.4Relevant Retrieval, Persistent Failure

> Model:SarvamCondition:D\-full \(Kannada script, full priming, RAG\) Query \(gloss\):“They ended our agreement before the time we decided” Human chosen label:Consumer/Money/Debt Predicted:Land/Property Reasoning trace \(translated\):“This situation involves a contract dispute related to property rights or control\. This is more of a property and contract dispute than a landlord\-tenant relationship\.” Verification:The retrieved passage \(Sridhar P\. v\. Ramesh, similarity 0\.911\) is an on\-topic financial/debt dispute, though financial matters are discussed briefly\. Despite having retrieved directly relevant material, the model still produced an unrelated category\.

## Appendix DFull Condition\-Level Results

This appendix gives the complete comprehension and RAG results underlying the condensed Findings in §[4](https://arxiv.org/html/2608.28645#S4): the full script×\\timespriming\-tier breakdown \(Table[7](https://arxiv.org/html/2608.28645#A4.T7), Figure[3](https://arxiv.org/html/2608.28645#A4.F3)\), the full 19\-condition breakdown \(Table[8](https://arxiv.org/html/2608.28645#A4.T8), Figure[4](https://arxiv.org/html/2608.28645#A4.F4)\), and the complete significance test battery \(Table[9](https://arxiv.org/html/2608.28645#A4.T9)\)\.

Table 7:Comprehension accuracy by script and priming tier, no RAG, all three models \(n=60n=60/cell\)\.Figure 3:Hits@1 by script and priming tier, per model \(n=60n=60/bar\)\.Table 8:Hits@1 / Hits@3 across all 19 conditions, per model \(n=60n=60each\)\.Figure 4:Hits@1 across all 19 conditions, per model \(n=60n=60/bar\)\. Each panel’s dashed line marks that model’s own no\-RAG, full\-priming ceiling \(E\_full\)\.Table 9:Fisher’s exact and binomial tests for pairwise condition comparisons at Hits@1,n=60n=60/condition\. Grouped by comparison type: baseline comprehension checks against chance/majority guessing, priming\-effect checks \(no RAG\), and RAG\-effect checks \(Kannada bare, sparse, and full priming, plus query\-language effect under RAG\)\.∗p<\.05\{\}^\{\*\}p<\.05,p∗⁣∗<\.01\{\}^\{\*\*\}p<\.01,∗∗∗p<\.001\{\}^\{\*\*\*\}p<\.001\.Table 10:Kannada\-script vs\. Romanized\-script comprehension \(Hits@1, %\) across all six priming/RAG tiers, per model \(n=60n=60/condition\)\. Values shown as Kannada%\>\>Romanized%\. Fisher’s exact test, same methodology as Table[9](https://arxiv.org/html/2608.28645#A4.T9)\. Computed from condition\-level results in Table[8](https://arxiv.org/html/2608.28645#A4.T8); no new model calls were made\.∗p<\.05\{\}^\{\*\}p<\.05,p∗⁣∗<\.01\{\}^\{\*\*\}p<\.01,∗∗∗p<\.001\{\}^\{\*\*\*\}p<\.001\.Table 11:Disagreement\-level analysis: among cases where full/sparse\-priming no\-RAG \(E\) and RAG \(D\) predictions diverge, how often each condition was uniquely correct, per model and priming tier \(n=60n=60/condition\)\.Table 12:CRAG relevance\-verification results, Llama3, Kannada\-script D\-condition queries \(n=20n=20\)\. Fisher’s exact test on Hits@1 \(English\-judge vs\. Kannada\-judge\):p=\.054p=\.054\. Reference points from the full 60\-query set \(Table[8](https://arxiv.org/html/2608.28645#A4.T8)\): D\_full \(RAG, full priming\) = 23\.3%; E\_full \(no\-RAG, full priming\) = 60\.0%\.
## Appendix ECourt Cases Cited

Table[13](https://arxiv.org/html/2608.28645#A5.T13)lists full identifiers for the eight court cases named in Appendix[C](https://arxiv.org/html/2608.28645#A3), drawn from the 18\-document Kannada legal corpus \(§[3\.2](https://arxiv.org/html/2608.28645#S3.SS2)\)\. All documents are public lower\-court filings retrieved from ecourtsindia\.com; CNR is the eCourts Case Number Record, the portal’s unique case identifier\.

Table 13:Court cases named in Appendix[C](https://arxiv.org/html/2608.28645#A3), with case number, deciding court, eCourts CNR identifier, and order date\. All are public lower\-court filings from ecourtsindia\.com \.