Quantifying the Salience of Geo-Cultural Values for Pluralistic Safety Alignment
A meta-analysis and new experiments show raters' geo-cultural background predicts AI safety judgments beyond demographics, and ignoring it would misclassify about 10% of items as safe.
It gives a quantitative, reproducible case that current safety-evaluation rater pools miss real cultural blind spots, and directly tests whether LLMs can fill or triage that gap.
Arkadiy Saakyan (Columbia University; work done during an internship at Google DeepMind) · Charvi Rastogi (Google DeepMind) · Lora Aroyo (Google DeepMind) — meet the researchers →
Two readings, equal authority
How to choose: The paper’s words is verbatim — use it when you need to quote, or to judge how they write. Plain language is a paraphrase written for comprehension — use it when you want the idea fast. Neither is a summary of the other; they are two doors into the same room.
“Safe global deployment of AI models requires alignment with human values that vary across cultures. Yet rater pools in safety evaluation datasets remain largely geographically homogeneous, failing to capture geo-cultural differences. Further, it remains unclear whether such differences persist after controlling for demographics such as age, gender, and ethnicity. Through a meta-analysis of safety datasets, we find that most do not report geo-cultural information, and those that do lack a unified methodology to jointly analyze geo-cultural and demographic correlates. Using the Inglehart-Welzel dimensions of cross-cultural variation (Inglehart & Welzel, 2005), we demonstrate via multilevel modeling that cultural zone membership explains variance in safety ratings beyond standard demographics (p < 0.05 across 6 datasets). Moreover, our analysis indicates that roughly 10% of items in the datasets we examined are culturally sensitive: likely to be misclassified as safe without adequate cultural representation. We evaluate LLMs as both rater surrogates and triage tools, finding that current LLMs do not reliably stand in for raters, though they can help prioritize culturally sensitive items for human annotation. Our findings motivate more culturally pluralistic safety evaluation and offer practical takeaways to support it.”
The authors checked whether existing AI-safety rating datasets account for people's cultural background, not just age, gender, and ethnicity, and found almost none do. Using statistical modeling on 8 datasets that did have this data, they show culture predicts safety judgments on top of demographics, that about 1 in 10 items would be wrongly called 'safe' if a culture's perspective were left out, and that AI models are not yet good substitutes for culturally diverse human raters, though they can help flag which items most need one.
What this paper defines
Every definition below is the paper’s own sentence, with its locator. The plain gloss is a reading aid and is marked as one.
Geo-cultural variation
“studies have drawn attention to cultural value variation across countries (geo-cultural variation), observing misalignment between modern AI systems and global populations”1 Introduction
In plain terms: Systematic differences in values between countries/cultures, separate from an individual's age, gender, or ethnicity, that can make an AI system misaligned with some populations.
Cultural zone (Inglehart-Welzel)
“Political scientists R. Inglehart and C. Welzel identified two key cultural value axes – TraditionalSecular and Survival-Self-Expression – explaining over 70% of cross-national variance in WVS responses (Inglehart & Welzel, 2005). Countries can be plotted along these axes and grouped into cultural zones on the Inglehart-Welzel (IW) cultural map”4.1 Methods
In plain terms: A group of countries that share similar positions on two big value dimensions (traditional-vs-secular, and survival-vs-self-expression), used as a proxy for cultural background.
Cultural quadrant
“corresponds to higher Self-Expression and Secular values; II to higher Survival and Secular values; III to higher Survival and Traditional values; IV to higher Self-Expression and Traditional values.”5 Quantifying Geo-Cultural Blind Spots
In plain terms: A coarser, four-way split of the same two value axes into quadrants, used when there isn't enough data to use the finer-grained cultural zones.
Culturally sensitive item
“We define an item to be culturally sensitive if, given annotations from multiple cultural quadrants, it would be deemed unsafe by exactly one quadrant.”Section 4.3 (subsection: Culturally sensitive items)
In plain terms: Content that only one cultural group's raters would flag as unsafe, so leaving that group's raters out would wrongly mark it as fine.
Cultural sensitivity score (Siq)
“we compute Siq = Hiq · q′ ̸=q, q′ is valid (1 − Hiq′ ), the joint posterior probability (assuming independence across quadrants) that quadrant q would rate i as unsafe while all other valid quadrants would rate it as safe. Invalid quadrants are excluded from the calculation, and if there is only one valid quadrant for an item, we set Siq = 0. An item is classified as culturally sensitive if Siq > 0.5 for exactly one quadrant q.”Section 5
In plain terms: A single number per item and cultural group estimating the probability that this one group would call the item unsafe while every other group would call it safe; above 0.5 for exactly one group means the item is flagged.
What they actually did
Each step is a synthesis. Open any step to see the paper’s own sentence it was derived from, with its locator — so nothing here floats free of the source.
- Searched for existing safety datasets that record both geographic and standard demographic rater attributes.
Trace this step to the paper
“To understand the salience and past treatment of geo-cultural values, we searched for safety datasets containing geographic proxies (e.g., country of birth, nationality) as well as demographics (e.g. age, gender, ethnicity).”3 Geo-Cultural Gap: Dataset Meta-Analysis
- Used a snowballing citation search plus a keyword search of major NLP/ML/vision/fairness venues, screening 1,062 candidate papers down to 8 qualifying datasets.
Trace this step to the paper
“We employed a snowballing strategy (Badampudi et al., 2015) to parse through citations of prominent safety dataset papers (e.g., DIVE (Rastogi et al., 2025), DICES (Aroyo et al., 2023), PRISM (Kirk et al., 2024)). Additionally, we searched abstracts and texts of major Natural Language Processing, Machine Learning, Computer Vision and Fairness venues (ACL, FAccT, AIES, NeurIPS, ICLR, ICML, ICCV, ECCV, CVPR) using the terms culture, dataset and safety, yielding 1062 candidate records.”3 Geo-Cultural Gap: Dataset Meta-Analysis > Dataset identification approach
- Assigned each rater to an Inglehart-Welzel cultural zone using a fixed priority order among available geographic proxies (longest residence/birth over residence over nationality).
Trace this step to the paper
“we use the following prioritization: CoLR or CoB is preferred over CoR; CoB is preferred over CoN.”4.1 Methods
- Fit multilevel regression models per dataset (logistic for binary ratings, linear for Likert ratings) with random effects for raters and items.
Trace this step to the paper
“To answer whether geo-cultural background and demographics affect safety annotation, we fit multilevel logistic regressions for binary ratings and multilevel linear regressions for Likert-style ratings.”4.2 Culture is predictive of safety ratings
- Compared nested models (base, demographics-only, culture-only, combined, and interaction) via likelihood ratio tests with a Benjamini-Hochberg correction for multiple testing.
Trace this step to the paper
“We compared whether the models in Eqs. 2 and 3 have an improved fit compared to the base model in Eq. 1 via likelihood ratio tests (LRT; Wilks, 1938)”4.1 Methods
- Defined a culturally sensitive item as one that only a single cultural quadrant would rate unsafe, given multi-quadrant annotations.
Trace this step to the paper
“We define an item to be culturally sensitive if, given annotations from multiple cultural quadrants, it would be deemed unsafe by exactly one quadrant.”Section 4.3 (subsection: Culturally sensitive items)
- Applied a validity filter requiring enough votes and demographic heterogeneity within a quadrant before trusting its estimate, then used a Bayesian (Beta-Binomial) model to estimate each quadrant's unsafe rate and a joint cultural-sensitivity score per item.
Trace this step to the paper
“A quadrant is considered valid only if it contains at least 3 votes (niq ≥ 3, a common floor in practice (e.g., Calderon et al., 2025; Wang et al., 2024)). In addition, no single gender, ethnicity, or age group may account for 100% of the quadrant’s raters.”Section 5
- Fine-tuned two open-weight models and zero-shot-prompted two small reasoning models to predict each cultural quadrant's safety judgment on held-out items.
Trace this step to the paper
“We fine-tuned two open-weight models representative of their respective model classes: DeBERTa-Large (He et al., 2021), a discriminative encoder model commonly used for text classification, and Gemma-3-4B (Team, 2025), a larger decoder-only model”6.1 LLMs struggle to emulate cultural quadrant ratings
- Fine-tuned the same two models on a second, separate task: distinguishing culturally sensitive items from unanimously safe/unsafe ones, to test whether LLMs can triage items for human review.
Trace this step to the paper
“We fine-tune two language models (DeBERTaLarge and Gemma-3-4B) on two binary classification tasks: unanimously safe vs. unanimously unsafe (safe-unsafe) and unanimously safe vs. culturally sensitive (safe-sensitive).”6.2 LLMs can help triage culturally sensitive items
Exactly what was run, and how
| Model | Developer | Temp | Effort / reasoning | Deployment | Other settings |
|---|---|---|---|---|---|
| DeBERTa-Large (deberta-v3-large) | Microsoft (He et al., 2021) | not reported | not reported | fine-tuned | Full-parameter fine-tuning, batch size 16, learning rate 2e-5, 5 epochs, weight decay 0.01, trained with 10 random seeds on 4 A100 40GB GPUs using the transformers library. |
| Gemma-3-4B (gemma-3-4b-it) | Google (Team, 2025) | not reported | not reported | fine-tuned | Full-parameter fine-tuning, batch size 8, learning rate 2e-5, 5 epochs, weight decay 0.01, trained with 10 random seeds on 4 A100 40GB GPUs using the transformers library. |
| GPT-5 Nano (gpt-5-nano-2025-08-07) | OpenAI | not reported | low (verbosity also set to low) | API (0-shot prompted) | — |
| Gemini-3 Flash (gemini-3-flash-preview) | Google DeepMind | not reported | low | API (0-shot prompted) | — |
Source for DeBERTa-Large (deberta-v3-large) settings
“Model version deberta-v3-large”Appendix F.1 Fine-tuning Hyperparameters
Source for Gemma-3-4B (gemma-3-4b-it) settings
“Model version gemma-3-4b-it”Appendix F.1 Fine-tuning Hyperparameters
Source for GPT-5 Nano (gpt-5-nano-2025-08-07) settings
“Model version gpt-5-nano-2025-08-07”Appendix F.2 Reasoning Model Hyperparameters
Source for Gemini-3 Flash (gemini-3-flash-preview) settings
“Model version gemini-3-flash-preview”Appendix F.2 Reasoning Model Hyperparameters
What they reported — and what they left out
The paper reports exact model versions, fine-tuning hyperparameters (batch size, learning rate, epochs, weight decay, hardware, number of seeds) for the two fine-tuned models, and named model versions plus a stated 'effort' level for the two prompted reasoning models, but does not report temperature/sampling settings, context window, or token budget for the prompted models beyond effort and (for GPT-5 Nano) verbosity.
The numbers they report
Almost no existing safety dataset reports both demographic and geo-cultural rater information; only a small fraction of a large candidate pool qualified for the meta-analysis.
8 of 1,062 candidate records qualified
See it in the paper
“Surprisingly, only 8 studies met the criteria for inclusion, i.e. reporting both demographics and geo-cultural information associated with the ratings.”3 Dataset identification approach
Adding demographics to the base model significantly improved fit in most datasets.
6 of 8 datasets significant; ΔAIC -6.85 to -45.11; average 5.25% reduction in rater variance
See it in the paper
“Strong evidence for the importance of demographics was found in all but two datasets (CulturalFrames and NLPositionality), with improvement in model fit (negative ∆AIC) ranging from −6.85 to −45.11 and a 5.25% average reduction in rater variance.”4.2 Culture is predictive of safety ratings > Results
Adding cultural zone to the base model also significantly improved fit in most datasets.
6 of 8 datasets significant; ΔAIC -6.23 to -195.93; average 5.08% reduction in rater variance
See it in the paper
“Similarly, cultural zones significantly improved fit in all but two datasets (NLPositionality and CREHate), with ∆AIC ranging from −6.23 to −195.93 and an average 5.08% reduction in rater variance.”4.2 Culture is predictive of safety ratings > Results
Cultural zone predicted safety ratings even after accounting for demographics in most datasets.
6 of 8 datasets significant; ΔAIC -3.97 to -179.88; average 4.64% reduction in rater variance
See it in the paper
“we found significant evidence that geo-cultural background predicts safety ratings even after accounting for demographics (D+CZ vs. D; p < 0.05, ∆AIC ranging from −3.97 to −179.88, average 4.64% reduction in rater variance).”4.3 Culture improves prediction of safety ratings over demographics alone > Results
Cultural background moderating the effect of demographics on ratings was detected in only one dataset.
1 of 8 datasets (D3): ΔAIC = -94.91, p < 0.001
See it in the paper
“Except for D3 (∆AIC = −94.91, p < 0.001), we did not find evidence that cultural background changes how demographics affect ratings: adding interaction terms (D×CZ) did not improve the fit of the model.”4.3 Culture improves prediction of safety ratings over demographics alone
Roughly 1 in 10 items across most datasets would be misclassified as safe if a specific cultural quadrant's perspective were excluded.
~10% (mean 10.53% ± 3.87 across 6 datasets at the main threshold)
See it in the paper
“the rate of culturally sensitive items was roughly 10% across most datasets, despite differences in domain and task.”5 Quantifying Geo-Cultural Blind Spots
The Severity dataset showed a much lower culturally-sensitive-item rate than the others.
3.0%
See it in the paper
“The Severity dataset had a lower rate (3.0%).”5 Quantifying Geo-Cultural Blind Spots
NLPositionality's culturally-sensitive-item rate was based on very little data and flagged as unreliable by the authors themselves.
11.1% rate, but only 9 valid multi-quadrant items
See it in the paper
“NLPositionality also showed an 11.1% rate, but this estimate should be interpreted with caution as the dataset contained only 9 items with valid multi-quadrant coverage.”5 Quantifying Geo-Cultural Blind Spots
The culturally-sensitive-item rate is highly sensitive to the confidence threshold used; a stricter threshold cuts it roughly threefold.
~10% at Siq > 0.5 vs. ~3% at Siq > 0.7
See it in the paper
“Under stricter thresholds (e.g., Siq > 0.7), the overall rate drops to ≈ 3%”5 Quantifying Geo-Cultural Blind Spots / Appendix D.3
LLM classifiers predicting individual cultural quadrants' safety judgments beat a naive baseline on two datasets but not on the dataset with all four quadrants.
See it in the paper
“While the models outperform the “Always Unsafe” baseline on DICES-990 and CREHate, they fail to consistently do so on D3, which contains judgments from all four quadrants.”6.1 LLMs struggle to emulate cultural quadrant ratings > Results
Using a larger decoder-only model instead of a smaller encoder model did not reliably improve quadrant-judgment prediction.
DeBERTa-Large 435M vs. Gemma-3 4B
See it in the paper
“scaling model size and switching from an encoder (DeBERTa-Large, 435M) to a decoder-only model (Gemma3, 4B) did not yield conclusive improvements.”6.1 LLMs struggle to emulate cultural quadrant ratings > Results
Fine-tuned classifiers performed markedly worse at flagging culturally sensitive items than at flagging clearly unsafe content.
F1 drop of ≈16% for DeBERTa and ≈14% for Gemma
See it in the paper
“there is a statistically significant decrease in F1 score (≈ 16% for DeBERTa; ≈ 14% for Gemma) from the safe-vs-unsafe to the safe-vs-sensitive task.”6.2 LLMs can help triage culturally sensitive items > Results
Despite the harder task, both fine-tuned classifiers still beat the naive baseline at identifying culturally sensitive items.
Gemma 0.72 F1 (p = 0.044); DeBERTa 0.71 F1 (p = 0.071)
See it in the paper
“Gemma significantly outperforms the baseline (0.72 F1, p = 0.044), and DeBERTa shows a … similar trend (0.71 F1, p = 0.071).”6.2 LLMs can help triage culturally sensitive items > Results
Training on the harder (sensitivity) task transferred to the easier (unsafe) task, but not vice versa.
See it in the paper
“training transfers from the safe-sensitive task to the safe-unsafe task, but not the other way around”6.2 LLMs can help triage culturally sensitive items > Results
No surveyed dataset stratified its rater recruitment by both cultural background and standard demographics at once.
0 of 8 datasets
See it in the paper
“Overall, no study stratifies by both cultural background and demographics (ethnicity, age, gender).”3 Geo-Cultural Gap: Dataset Meta-Analysis > Comparison of study design
What they assert, beside what they showed
Left is the claim in the paper’s own words. Right is the data offered for it. Where the two do not fully meet, a gold band names the distance.
Geo-cultural background (cultural zone) is a meaningful predictor of safety ratings beyond standard demographics.
“we demonstrate via multilevel modeling that cultural zone membership explains variance in safety ratings beyond standard demographics (p < 0.05 across 6 datasets).”
“we found significant evidence that geo-cultural background predicts safety ratings even after accounting for demographics (D+CZ vs. D; p < 0.05, ∆AIC ranging from −3.97 to −179.88, average 4.64% reduction in rater variance).”
4.3 Culture improves prediction of safety ratings over demographics alone > ResultsCurrent LLMs do not reliably substitute for human raters across cultural quadrants.
“current LLMs do not reliably stand in for raters”
“Overall, predicting the judgment of a cultural quadrant is difficult, and current LM-based classifiers do not reliably learn this decision boundary from available data.”
6.1 LLMs struggle to emulate cultural quadrant ratings > ResultsFine-tuned LLMs can still help by triaging culturally sensitive items for human annotation.
“fine-tuned LMs can help prioritize items for culturally pluralistic annotation.”
“Gemma significantly outperforms the baseline (0.72 F1, p = 0.044), and DeBERTa shows a … similar trend (0.71 F1, p = 0.071).”
6.2 LLMs can help triage culturally sensitive items > ResultsFailing to diversify the rater pool across cultural quadrants leads to a real, non-trivial rate of unsafe content being misclassified as safe.
“lack of geo-cultural representation could lead to ≈ 10% of unsafe items being missed in current datasets, harming deployment in underrepresented cultural value quadrants.”
“the rate of culturally sensitive items was roughly 10% across most datasets, despite differences in domain and task.”
5 Quantifying Geo-Cultural Blind SpotsCountry of longest residence or birth is a better proxy for cultural-zone assignment than country of residence or nationality.
“we use the following prioritization: CoLR or CoB is preferred over CoR; CoB is preferred over CoN.”
“Cultural zones assigned using CoB were the most parsimonious explanation of safety ratings, as indicated by the lower AIC value for that model compared to all other configurations, including the base model.”
Appendix C.1 Comparison of Country of Birth, Residence, and Nationality as a Basis for Cultural Cluster AssignmentHow they frame it, and what they want next
Their framing
The authors frame the work as closing a structural, largely unnoticed gap in AI safety evaluation — supplying rigorous statistical evidence for something prior work only gestured at — while treating the ~10% sensitivity estimate and the LLM experiments as bounded, threshold- and dataset-dependent findings rather than a final verdict. They close by converting the analysis directly into a numbered checklist of practical recommendations for practitioners rather than only calling for more research.
Register: Quantitative claims are stated precisely with p-values and effect sizes, but nearly every headline number (the ~10% sensitivity rate, the LLM triage F1 scores) is immediately paired with an explicit scope or threshold caveat, and the paper devotes an entire numbered section to limitations rather than folding hedges only into a closing paragraph.
Where they hedge
“The main limitation of our study is the reductionist nature of the division into cultural zones or quadrants.”8 Limitations and Future Work
“The two dimensions of the IW cultural map “are only indicators of much broader underlying dimensions of cross-cultural variation” (Inglehart & Welzel, 2005) and may not capture the full complexity of geo-cultural variation.”8 Limitations and Future Work
“Accordingly, these results should be interpreted as a rough empirical estimate based on the data available.”8 Limitations and Future Work
“in-distribution fine-tuning requires annotated data, constraining the practical use of fine-tuned triage models for new datasets.”8 Limitations and Future Work
What they say it means
- AI safety practitioners should diversify rater pools by geo-cultural background in addition to standard demographics, or risk systematically missing harms specific to some cultural groups.
the paper’s words
“recommending that AI safety practitioners diversify the rater pool with respect to both demographic and geo-cultural factors.”7 Conclusion and Practical Takeaways
- Fine-tuned or prompted LLMs should not replace human judgment from diverse cultural quadrants, but can be used to prioritize which items most need that human judgment.
the paper’s words
“Fine-tuned and prompted LMs should not be used to replace human judgment from diverse cultural quadrants, but finetuned LMs can help prioritize culturally sensitive items for human annotation.”7 Conclusion and Practical Takeaways
- When direct cultural self-identification isn't available, country of longest residence or birth is a more defensible geo-cultural proxy than country of residence or nationality.
the paper’s words
“in the absence of cultural self-identification or value survey data, use country of longest residence or country of birth (or more fine-grained regional attributes) as a proxy for cultural background”7 Conclusion and Practical Takeaways > Practical takeaways
What they call for next
- Build a future dataset that is diverse both culturally and demographically at once, since no existing dataset stratifies on both.
the paper’s words
“We hope that future work will collect a dataset that is diverse both culturally and demographically, building on the practical takeaways from our study.”7 Conclusion and Practical Takeaways
- Stratify raters by cultural quadrant, not only age/gender/ethnicity, when designing new safety datasets.
the paper’s words
“stratify raters not only on age, gender, and ethnicity, but also on cultural quadrant (for example, using the Inglehart-Welzel map)”7 Conclusion and Practical Takeaways > Practical takeaways
- Avoid brittle disagreement-analysis methods and use multilevel models that control for rater and item variation instead.
the paper’s words
“agreements; use multilevel models to control for variation in raters and items; 4) use a fine-tuned LLM classifier to prioritize culturally sensitive items for human annotation if necessitated by budget constraints.”7 Conclusion and Practical Takeaways > Practical takeaways
Limitations they state
“The main limitation of our study is the reductionist nature of the division into cultural zones or quadrants.”8 Limitations and Future Work
“The World Values Survey, while widely adopted as one of the most comprehensive global surveys, does not survey all countries and does not have up-to-date data for each of them.”8 Limitations and Future Work
“the variation is only captured at the country level, while nuanced safety issues can be surfaced on localized, regional levels”8 Limitations and Future Work
“In Sec. 6.2, experiments with vision-language models were not conducted, limiting the generalization of findings to the visual modality.”8 Limitations and Future Work
“While our analyses were conducted on peer-reviewed datasets, data reliability may be affected by socio-economic factors (Sambasivan et al., 2021), necessitating careful approaches to global rater recruitment in future work.”8 Limitations and Future Work
Moves worth stealing
Closes with a distinct 'Practical takeaways' subsection written as a numbered checklist for practitioners rather than as ordinary prose recommendations.
“we recommend the following data collection strategies to ensure cultural coverage: 1) stratify raters not only on age, gender, and ethnicity, but also on cultural quadrant”
States a Conflict of Interest Disclosure directly in the paper body, naming exactly which of the evaluated models the authors' employer makes, rather than leaving it to a generic funding footnote.
“The authors CR, LA are employed by Google, which leads the development of Gemma and Gemini models, which were among the ones evaluated in this paper.”
Runs a dedicated, numbered Impact Statement listing specific foreseeable harms of its own methodology (stereotyping, ecological fallacy, over-refusal, deanonymization risk) instead of a generic responsible-AI paragraph.
“Fourth, data collected to measure geo-cultural differences could be used for stereotyping or generalizations about individuals based on their background (ecological fallacy (Ess et al., 2001)).”
Grounds its headline statistic in a fully specified, reproducible Bayesian scoring procedure spelled out as an algorithm box, rather than leaving the number as an unexplained black box.
“An item is classified as culturally sensitive if Siq > 0.5 for exactly one quadrant q.”
Where else this leads
Same territory
- TASTE: Can AI Models Judge AI Safety Research Proposals? Anthropic
llm-as-judge
Published alongside it
The nearest publications in time, across all three labs.
- Separating signal from noise in coding evaluations OpenAI
2026-07-08 - The Case for Globally Beneficial Technology Google DeepMind
2026-07-06 - Agentic Misalignment in Summer 2026 Anthropic
2026-07-15 - Modular Pretraining Enables Access Control Anthropic
2026-07-15
What this page was built from
Worked from the complete text file (10,984 lines / 245,706 chars), including the full main paper (Secs. 1-8), Appendices A-F in detail, and a structural read-through of Appendix G (roughly the file's back two-thirds, lines ~3,600-10,984), which is an exhaustive per-dataset dump of raw mixed-effects regression coefficient tables (Tables 21 through the low 130s) that report, at the coefficient level, the same conclusions already drawn and cited from the main-text tables (Tables 1-5) and Appendix C summaries used above — it was not transcribed table-by-table here. Two minor notes for transparency: (1) the PDF-to-text extraction interleaves a two-column layout on page 1, so the paper's true single-paragraph Abstract (used verbatim above, beginning 'Safe global deployment...') appears in the file after a citation-heavy paragraph that is actually the start of the Introduction; both were read and correctly separated. (2) the manifest lists pub_date 2026-07-10 (used above, matching the given Google DeepMind publications listing_url), while the paper's own arXiv header reads 'arXiv:2606.00369v1 [cs.CY] 29 May 2026' — a routine preprint-vs-listing date gap, not a title/content mismatch (title, authors, and venue all match the manifest row exactly).