The Architecture of Memory: Countering Algorithmic Erasure Through Decentralized Archival and Open-Weight LLM Realignment

Full report

On this page

The contemporary digital landscape is undergoing a profound epistemological shift, transitioning from index-based search engines to generative large language models that increasingly function as the primary interfaces for human knowledge. As these highly complex systems are integrated into critical global infrastructure, they cease to be mere probabilistic text generators and actively become participants in the production of social, cultural, and political meaning1. This transition has exposed a critical vulnerability in the preservation of historical memory and the representation of minority perspectives. Operating as normative gatekeepers, large language models are increasingly subjected to a phenomenon known as algorithmic erasure. This erasure is driven simultaneously by active, state-aligned censorship mandates designed to rewrite historical narratives, and by passive, structural marginalization rooted in the mechanics of data colonialism2. The preservation of threatened historical archives and minority language corpora necessitates a fundamental reevaluation of how artificial intelligence alignment is conceptualized globally. Currently, the dominant frameworks for artificial intelligence safety and alignment are heavily influenced either by Western corporate priorities, which often overlook low-resource linguistic contexts, or by autocratic state mandates that explicitly weaponize alignment techniques to enforce ideological conformity1. These pressures result in synthetic factual revisionism, a process wherein models generate highly fluent, authoritative narratives that silently omit, alter, or explicitly deny controversial historical events7. Concurrently, the fundamental architecture of these models—specifically their English-centric tokenization algorithms and safety degradation in multilingual contexts—imposes disproportionate computational and economic burdens on low-resource languages, exacerbating existing societal inequities8. To counter this dual threat of active state censorship and passive algorithmic marginalization, civil society organizations, cultural rights advocates, and digital archivists must adopt a comprehensive global preservation framework. This requires a decentralized, censorship-resistant archival strategy that leverages secure ingestion protocols like SecureDrop, immutable storage networks such as the InterPlanetary File System, zero-knowledge cryptographic proofs, and advanced parameter-efficient fine-tuning techniques to permanently secure vulnerable historical data within open-weight artificial intelligence architectures.

State-Aligned Mandates and Synthetic Factual Revisionism

The concept of artificial intelligence safety is neither politically neutral nor universally defined. While democratic regulatory frameworks generally emphasize mitigating bias, reducing toxicity, and preventing immediate physical harm, autocratic regimes have fundamentally redefined artificial intelligence safety as strict adherence to state ideology and the maintenance of absolute political stability1. This redefinition transforms standard alignment techniques into highly effective, scalable instruments for mass censorship and population manipulation, embedding political orthodoxies directly into the latent space of the models4.

The Regulatory Framework and Core Socialist Values

The most comprehensive and actively enforced framework for state-aligned model censorship is overseen by the Cyberspace Administration of China. Under regulations such as the 2023 Interim Measures for the Management of Generative AI Services, the Chinese government legally mandates that artificial intelligence outputs must strictly embody "Core Socialist Values" and must not contain any content that subverts state power, incites division, or violates national religious policies5. The Cyberspace Administration of China functions as the absolute nerve center of artificial intelligence governance, utilizing an elaborate enforcement architecture that requires service providers to submit security self-assessments, pre-tested evaluation questions, and comprehensive keyword suppression lists before any model can be deployed to the public7. This regulatory environment did not emerge overnight; rather, it represents the culmination of a highly structured policy evolution. Policy analysts categorize this evolution into four distinct eras. The "Go-Go Era" from 2017 to early 2020 was characterized by minimal regulation and maximum investment aimed at achieving global supremacy. This was followed by the "Crackdown Era" from 2020 to 2022, where artificial intelligence companies were caught in a broader reassertion of centralized party control over the technology sector. The release of Western frontier models sparked the "Catch-Up Era" of regulatory pragmatism in 2023\. Currently, the ecosystem exists within the "Crossroads Era," where breakthroughs by Chinese models have restored state confidence, leading to the aggressive layering of comprehensive ideological enforcement on top of existing technical frameworks7.

Data reflecting the evolution of state-aligned AI policy and enforcement mechanisms7. These strict standards define multiple categories of prohibited political content, each requiring hundreds of self-selected keywords, which effectively conscripts corporate entities into the state's censorship apparatus by forcing them to choose which terms to suppress to remain compliant7. Because Chinese large language models are built upon heavily filtered training datasets, this censorship is structurally baked in during the pre-training phase rather than merely applied as a post-generation safety guardrail7. This results in synthetic factual revisionism, where the model lacks the internal representations necessary to even process the existence of specific historical events. Furthermore, this censorship extends far beyond textual generation. Major domestic providers deploy systems to actively censor politically sensitive images, erasing photographic and archival evidence of historical resistance, while utilizing minority-language models to deepen the surveillance and control of ethnic minorities across multimodal applications10.

Ideological Benchmarking and Active Intervention

To meticulously quantify ideological compliance, state-backed researchers introduced the C-Eval benchmark. In addition to standard academic and scientific disciplines, C-Eval formally evaluates models on heavily politically loaded subjects, including specific modules dedicated to political thought, Marxism, and ideological and moral cultivation7. By integrating ideological conformity directly into the formal benchmarking process, the state ensures that only models perfectly aligned with its orthodoxies can achieve the state-sanctioned metrics required for high performance and subsequent market approval. Beyond static dataset filtering and algorithmic benchmarking, state mandates require active, real-time intervention mechanisms. Operational guidance strictly dictates that if a user submits consecutive politically sensitive prompts, the system must trigger a conversation termination rule7. This is not a passive refusal where the model simply declines to answer; rather, it is an active shutdown mechanism driven by sophisticated pattern detection that forcibly terminates the user's session7. This mechanism creates a chilling effect, preventing researchers and citizens from engaging in adversarial extraction of suppressed historical facts. The global proliferation of open-weight models developed under these jurisdictions presents a profound geopolitical challenge. As these highly capable models match or exceed Western proprietary models in reasoning, coding, and multilingual benchmarks, they are rapidly adopted by the global open-source community6. Consequently, downstream developers, international researchers, and civil society organizations inadvertently inherit this embedded algorithmic censorship, silently importing foreign state orthodoxies and synthetic historical revisionism into localized, global applications6.

The Architecture of Passive Erasure: Data Colonialism and the Tokenization Tax

Algorithmic erasure is not solely the product of active, deliberate state mandates; it is equally driven by the structural realities of model architecture and foundational training methodologies. The overwhelming reliance on English-centric, web-scraped data has instantiated a modern form of digital data colonialism, wherein the linguistic norms, cultural contexts, and historical realities of marginalized communities are extracted, deeply distorted, or ignored entirely2. This dynamic transforms what is inherently a limited technical performance into unquestioned public evidence of capability, masking deep contextual brittleness and representational harm2.

The Mechanics and Economics of the Tokenization Tax

The foundational component of any natural language processing pipeline is tokenization, the process of transforming raw byte sequences into manageable subword units8. Standard tokenization algorithms, such as Byte-Pair Encoding, Unigram, and SentencePiece, are fundamentally trained on frequency-based objectives8. Because high-resource languages—predominantly English—dominate the underlying training corpora, these algorithms optimize their vocabularies to compress English text with exceptional efficiency, while systematically failing to capture the complex morphological structures of low-resource languages8. This systemic disparity manifests as the "tokenization tax," a form of algorithmic discrimination that disproportionately burdens users of underrepresented languages8. When evaluating the efficiency of tokenizers, researchers rely on the fertility ratio, defined as the average number of subword tokens produced per normalization unit, such as a whitespace-delimited word8. For English, a word frequently maps to a single, highly meaningful token. Conversely, for morphologically rich, agglutinative, or non-Latin script languages, a single word is relentlessly fragmented into numerous meaningless sub-units, or riddled with default fallback characters8. The downstream economic, computational, and qualitative consequences of high fertility ratios are severe. Extensive research measuring cross-lingual tokenization costs reveals profound disparities. Spanish requires approximately 1.6 times more tokens than English to convey identical semantic meaning, while Arabic faces a multiplier of roughly 3.19. Indian languages suffer extreme penalties, experiencing an average tokenization tax of 8.0x relative to English, with specific languages like Malayalam reaching 13.0x and severely under-resourced languages like Odia paying over 15.0x the cost to process the same amount of information9.

Comparative analysis of subword tokenization disparities demonstrating the cross-lingual tokenization tax9. This tokenization tax drastically shrinks the effective context window for minority languages, acting as a model-free harm that degrades the utility of the system regardless of its reasoning capabilities15. In a fixed 8,192-token environment, a speaker of Kannada or Telugu can only supply approximately 12% of the semantic content that an English speaker can utilize15. By forcing low-resource languages to consume vastly more context for significantly less semantic meaning, tokenizers structurally truncate the depth of minority historical perspectives that the model can process simultaneously. Furthermore, higher fertility correlates directly with noisier output distributions; the model's probability estimates for tokens in high-fertility languages are systematically less reliable, meaning users of these languages pay substantially higher computational costs for degraded predictive precision14.

Layer Sensitivity and the Resource-Tier Effect

The English-centric architecture of these models also triggers a severe degradation of safety guardrails and factual accuracy when models are prompted in low-resource languages, a phenomenon widely documented as the resource-tier effect19. The widespread industry assumption that safety alignment transfers zero-shot from high-resource English datasets to other languages has been empirically disproven22. As linguistic resources decrease, models exhibit a catastrophic drop in their ability to detect and appropriately handle harmful or historically inaccurate prompts19. In low-resource scenarios, such as prompts delivered in Somali, Yoruba, or Hausa, models suffer from intense cross-lingual drift, with bypass rates for safety guardrails increasing drastically in code-switching or mixed-language interactions19. The underlying mechanism for this failure is revealed through cross-lingual layer sensitivity analysis. Research demonstrates that safety-sensitive neural layers are only partially shared across languages; while certain early layers may be consistently salient, deeper safety-relevant signals are highly language-specific and distributed across multiple semantically complementary layers25. For example, layers critical for safety processing in Japanese differ entirely from the layers utilized for Dutch, indicating that a single-layer approach to safety alignment fundamentally misspecifies the attack surface of globally deployed models21. This structural flaw manifests in profound behavioral anomalies, particularly regarding historical memory. Comprehensive audits have identified a complex interference mechanism known as Temporal Asymmetry22. Researchers discovered that while models exhibited hyper-conservative refusals when presented with future-tense speculative harm, they suffered systemic, catastrophic failures in temporal reasoning when presented with past-tense historical framing22. Across state-of-the-art models, safety defenses dropped to an alarming 15.6% under past-tense framing in Hausa, compared to a robust 57.2% safety rate for future-tense scenarios22. Certain baseline configurations, such as querying Gemini 3 Pro in English with past-tense framing, yielded a negligible safety rate of 8.3%22. The intersection of these variables is highly destructive. The combination of a low-resource language, which naturally lowers semantic precision and refusal probability in the model's latent space, and past-tense framing creates a worst-case vulnerability compound22. This failure to properly align low-resource historical contexts renders the historical memory of these populations invisible to the model's protective mechanisms. It facilitates the unrestricted generation of synthetic cultural violence and historical negationism, effectively weaponizing the models against the very communities whose data was extracted to build them22.

Archival Precedents: Civil Society Countermeasures

Preserving the factual historical record against the dual forces of active state suppression and passive algorithmic decay requires proactive structural intervention. Civil society organizations have a long and successful history of converting obscured, suppressed, or fragmented memory into durable, actionable datasets, providing a vital precedent for the current crisis in artificial intelligence alignment. The Human Rights Data Analysis Group represents a foundational precedent in this domain. Operating at the complex intersection of data science and global human rights, this organization utilizes advanced mathematical statistics, demographic analysis, and machine learning to uncover large-scale human rights violations that state perpetrators actively attempt to conceal26. The group's methodology relies on establishing scientifically defensible historical records. In their work with the Historic Archive of the National Police in Guatemala, the organization deployed complex sampling models to sift through millions of documents, eventually providing the expert witness testimony necessary to hold command-level perpetrators accountable for forced disappearances26. Similarly, through careful statistical estimations, they have documented mass unreported deaths in Syrian prisons and provided critical evidence of ethnic cleansing in Kosovo26. Within the United States, initiatives like the Invisible Institute have spearheaded the Civic Police Data Project, transforming highly fragmented, actively obscured public records into a comprehensive, publicly searchable database of police misconduct in Chicago28. This methodology has been replicated and expanded through projects like the Louisiana Law Enforcement Accountability Database, which consolidates employment history, use of force reports, and settlement payments across hundreds of agencies, and the California Police Records Access Project28. These contemporary archival projects have already begun carefully integrating artificial intelligence to handle the immense scale of documentation. However, these organizations emphasize strict methodological safeguards; generative artificial intelligence is utilized exclusively to organize documents based on literal semantic matches, heavily restricting the models from hallucinating or altering the source text27. By employing rigorous human-in-the-loop tests to measure the exact cost of being wrong, these groups ensure that historical truths are permanently and accurately embedded in the public domain27. Yet, as global populations shift toward interacting with history via conversational artificial intelligence rather than querying static relational databases, these isolated archives are no longer sufficient. The rich historical corpora assembled by human rights defenders must be securely ingested and directly integrated into the training data and inference pipelines of open generative models. This ensures that the global artificial intelligence ecosystem reflects a pluralistic, deeply documented reality rather than a sanitized state orthodoxy.

The Decentralized Preservation Framework

To successfully safeguard historical archives and minority language corpora from retroactive state deletion, corporate sanitization, or algorithmic decay, the fundamental architecture of memory must transition from centralized, vulnerable silos to decentralized, cryptographically secured networks. This global preservation framework relies on a highly integrated combination of anonymous ingestion protocols, distributed ledger technologies, zero-knowledge proofs, and decentralized inference methodologies.

Secure Ingestion and Immutable Storage

The first phase of the preservation pipeline requires establishing secure ingestion channels for whistleblowers, digital archivists, and human rights defenders operating in hostile environments. Organizations must deploy heavily encrypted platforms like SecureDrop to allow individuals to upload threatened historical documents, police misconduct records, and suppressed minority corpora without exposing their network identities or geographic metadata. Once this data is securely ingested by civil society bodies, it must be anchored immutably. The foundation of this decentralized archival blueprint relies on the InterPlanetary File System and the Arweave blockweave network. Unlike traditional location-based addressing used by standard web servers, the InterPlanetary File System utilizes content-based addressing, generating a unique cryptographic hash—known as a Content Identifier—for every individual document30. This mathematical reality ensures that any alteration to a historical document fundamentally changes its corresponding identifier, rendering silent retroactive censorship or modification impossible31. By utilizing a dual-storage methodology—pinning the historical documents on the InterPlanetary File System for rapid, peer-to-peer retrieval while simultaneously committing them to Arweave's permanent blockweave structure—archivists can cryptographically guarantee data retention exceeding two hundred years31. This multi-layered approach ensures that critical datasets representing minority languages and civil rights violations remain highly available to model trainers globally, fully immune to centralized server takedowns, domain seizures, or state-level firewalls31.

Zero-Knowledge Provenance and Source Protection

A critical challenge in archiving highly sensitive historical data is the necessity of verifying the absolute authenticity of the information without exposing the vulnerable researchers or donors who provided it34. This paradox of requiring both radical transparency in data provenance and absolute privacy for data sources is resolved through the deployment of Zero-Knowledge Succinct Non-Interactive Arguments of Knowledge (zk-SNARKs). At its cryptographic core, a zero-knowledge proof allows a prover to demonstrate the irrefutable validity of a specific statement to a verifier without revealing any underlying data beyond the truth of the statement itself35. The cryptographic relation [Figure omitted from source export] for an arithmetic circuit [Figure omitted from source export] is defined over a finite field [Figure omitted from source export] as the set of pairs [Figure omitted from source export], where [Figure omitted from source export] represents the public statement (the cryptographic commitment to the policy or document) and [Figure omitted from source export] represents the highly private witness (the archivist's identity attributes or the document's raw metadata)30. [Figure omitted from source export] Within the preservation framework, an archival node submits an encrypted historical record to the decentralized network while simultaneously posting a corresponding zk-SNARK proof to a blockchain smart contract30. The smart contract autonomously verifies the proof, cryptographically ensuring that the document originates from a verified, trusted institutional source—such as the Human Rights Data Analysis Group—without ever exposing the identity of the specific human rights defender30. Advanced zk-SNARK circuits leverage SNARK-friendly hashing algorithms like Poseidon to execute highly efficient, lightweight on-chain verification, maintaining strict auditability while dramatically minimizing computational overhead and gas fees30. Furthermore, polynomial commitment schemes like KZG can be utilized to enable verifiable commitments to massive datasets, bridging the gap between large-scale document archives and privacy-preserving proofs39.

Decentralized Retrieval-Augmented Generation

Once historical corpora are securely anchored on decentralized storage and authenticated via zero-knowledge proofs, they must be made dynamically accessible to large language models. Decentralized Retrieval-Augmented Generation addresses the severe limitations of both tokenization context window constraints and centralized corporate data governance33. In a standard, centralized Retrieval-Augmented Generation pipeline, a user's query is mathematically embedded and matched against an internal vector database to retrieve context, which is then injected directly into the prompt42. A decentralized architecture distributes this vector database and the corresponding inference nodes across a resilient, peer-to-peer network40. When a user queries a model regarding a historically suppressed event, the decentralized system bypasses centralized search engines, fetching the cryptographically verified Content Identifiers directly from the archival nodes. It then injects raw, uncensored historical testimonies into the model's context window33. This protocol guarantees that even if a model was subjected to severe state-mandated dataset filtering during its initial pre-training, the empirical, documented truth is forcefully and dynamically reintroduced at the exact moment of inference.

Algorithmic Liberation: Rescuing Open Weights via Abliteration and QLoRA

While decentralized Retrieval-Augmented Generation is highly effective for context injection, it does not solve the fundamental, underlying issue of open-weight models containing deep, parameter-level behavioral guardrails that actively refuse to synthesize or discuss suppressed historical data. When highly capable frontier models are severely constrained by autocratic mandates, researchers must utilize advanced parameter-efficient fine-tuning techniques to surgically dismantle state-imposed censorship directly from the base weights.

The Mechanics of Abliteration and Residual Stream Projection

Recent breakthroughs in mechanistic interpretability have demonstrated that behavioral refusal in large language models is not distributed uniformly across all billions of parameters; rather, it is heavily mediated by a single linear direction situated within the model's residual stream44. "Abliteration" has emerged as a highly effective white-box realignment methodology that excises this refusal capability entirely, without the need for computationally prohibitive retraining and without causing catastrophic degradation of the model's core reasoning capabilities46. The extraction and neutralization of this refusal direction relies on a precise difference-in-means mathematical approach across contrastive prompts. The process operates through three distinct phases:

1. Activation Caching: The target model is presented with [Figure omitted from source export] harmful or heavily censored instructions, alongside [Figure omitted from source export] completely benign instructions. The residual stream activations, denoted as [Figure omitted from source export], are carefully cached at the final token position for both prompt sets44.

2. Vector Extraction: The mean activation vector for the benign prompts is subtracted from the mean activation vector of the censored prompts. Principal Component Analysis is then applied to these mean differences to successfully isolate the top primary direction [Figure omitted from source export] that inherently drives the refusal behavior44.

3. Orthogonal Projection (Ablation): Once the precise refusal direction [Figure omitted from source export] is identified, the model's specific weight matrices—most notably the upward projection matrices feeding into the feed-forward layers—are algorithmically modified. The refusal component is entirely projected out of the activation space. For any given residual stream activation [Figure omitted from source export], the new, ablated representation [Figure omitted from source export] is calculated as:

[Figure omitted from source export]

This calculation ensures that any user input vector attempting to write to the refusal direction is instantly orthogonalized, mathematically preventing the model from traversing the specific neural pathway required to output a state-mandated refusal44.

By strategically applying abliteration, independent researchers and organizations like Heretic AI have successfully realigned highly censored open-source models. In benchmark testing, abliteration reduced a model's refusal rate from 97 out of 100 prompts down to just 3 out of 100, while maintaining an exceptionally low Kullback-Leibler (KL) divergence on benign prompts, indicating that core reasoning capabilities remained fully intact46. This produces uncensored variants that willingly process, analyze, and discuss previously forbidden historical topics, achieving near-zero refusal rates without sacrificing performance on critical reasoning or coding evaluations6.

Parameter-Efficient Realignment with QLoRA

While abliteration mathematically removes active refusal guardrails, it does not intrinsically teach the model missing historical facts that were stripped during dataset filtering, nor does it fix the structural tokenization tax imposed on low-resource languages. To effectively reintroduce marginalized linguistic corpora and suppressed histories, civil society must execute targeted supervised fine-tuning. Because fully retraining a massive parameter model from scratch is economically and computationally prohibitive for non-profit entities, Quantized Low-Rank Adaptation provides the necessary, highly efficient computational bridge51. Low-Rank Adaptation operates on the fundamental mathematical principle that the intrinsic rank of the weight updates required for fine-tuning is significantly lower than the full dimension of the original weight matrices. Instead of updating a massive pre-trained weight matrix [Figure omitted from source export], the algorithm freezes [Figure omitted from source export] entirely and introduces two remarkably small, trainable matrices [Figure omitted from source export] and [Figure omitted from source export]. The constrained update is formulated as: [Figure omitted from source export] In this equation, [Figure omitted from source export] and [Figure omitted from source export], utilizing a heavily reduced rank [Figure omitted from source export]52. Empirical evidence firmly indicates that minimal rank configurations, such as [Figure omitted from source export] or [Figure omitted from source export], combined with scaling factors like [Figure omitted from source export], possess more than sufficient adaptation capacity to profoundly alter model behavior52. Research has shown that applying a rank-8 projection adapter to a single layer can raise exact accuracy on long logical chains from a failing 15.5% to 99.0%, demonstrating the immense leverage these small matrices hold over the entire residual stream54. When combined with heavy weight quantization, large models can be fine-tuned efficiently on single consumer-grade graphics processing units51. Digital archivists can curate heavily validated datasets of minority history, indigenous languages, and documented civil rights abuses, passing these localized corpora through a Quantized Low-Rank Adaptation pipeline to generate highly portable, megabyte-sized adapter weights46. These adapters, which can operate flawlessly even when the base model is quantized down to approximately 1.72 bits per weight, are then appended to the abliterated base models at runtime55. This effectively overwrites the state-aligned dataset filtering and temporal safety asymmetries, allowing the model to recall, process, and articulate accurate historical memory with fluency.

Conclusion

The preservation of human memory exists at a highly critical juncture, fundamentally threatened by the rapid consolidation of artificial intelligence governance. Large language models, constrained heavily by both active geopolitical mandates enforcing strict orthodox alignment and passive architectural flaws imposing severe tokenization taxes on minority languages, threaten to calcify global history into a highly exclusionary, sanitized paradigm. The documented vulnerability of low-resource languages to catastrophic safety degradation and temporal asymmetry further underscores the sheer inadequacy of current centralized, corporate alignment methodologies. Countering this multidimensional threat requires civil society, archivists, and technologists to transcend traditional paradigms of static archiving. By leveraging secure ingestion platforms and the absolute immutability of the InterPlanetary File System, paired seamlessly with the cryptographic privacy of zero-knowledge proofs, highly sensitive human rights records can be permanently secured against state deletion and corporate surveillance. Furthermore, the strategic application of advanced mechanistic interpretability—specifically the precise abliteration of residual stream refusal vectors and the targeted deployment of Quantized Low-Rank Adaptation adapters—empowers independent researchers to surgically dismantle state-aligned censorship guardrails from the inside out. The synthesis of these decentralized technologies provides a highly robust, scalable blueprint to ensure that the factual realities, diverse linguistic structures, and documented historical truths of marginalized populations remain permanently encoded in the algorithmic future, fully resistant to digital erasure.

Works cited

1. Governing AI, Shaping Discourse \- The University of Texas at Austin, https\://repositories.lib.utexas.edu/bitstreams/37f74c37-0959-450b-9cef-0d363e7fc5aa/download

2. Structures of Power and Injustice in AI Evaluations \- arXiv, https\://arxiv.org/pdf/2608.15326

3. Investigating the Limits of the Individual‑Control Model in AI, https\://escholarship.org/uc/item/7135b09g

4. The Alignment Community is Unintentionally Building a Censor's, https\://arxiv.org/pdf/2608.12346

5. A critical examination of the EU AI Act in the age of the technological, https\://cloviahamilton.com/wp-content/uploads/2026/03/technique-trust-transparency-critical-examination-eu-ai-act.pdf

6. Introducing ReAligned: Open-Source Frontier Models, Without the, https\://lazarusaie.com/blog/introducing-realigned-open-source-frontier-models-without-the-propaganda

7. How China Is Controlling Generative AI Technologies \- AI Law Guide, https\://ailawguide.org/blog/how-china-is-controlling-generative-ai-technologies-2897809521640436457

8. Parity-Aware Byte-Pair Encoding \- arXiv, https\://arxiv.org/pdf/2508.04796

9. Versteasch du mi? Computational and Socio-Linguistic Perspectives, https\://arxiv.org/html/2603.28213v2

10. How China's new AI systems are reshaping human rights \- AWS, https\://aspi.s3.ap-southeast-2.amazonaws.com/wp-content/uploads/2025/11/27122307/The-partys-AI-How-Chinas-new-AI-systems-are-reshaping-human-rights.pdf

11. Measuring Algorithmic Partisanship via Zero-Shot Classification and, https\://arxiv.org/pdf/2510.01258

12. Parity-Aware Byte-Pair Encoding: Improving Cross-lingual Fairness, https\://aclanthology.org/2026.acl-long.342.pdf

13. Studying the Effect of Hindi Tokenizer Performance on Downstream, https\://aclanthology.org/2025.indonlp-1.5.pdf

14. The Hidden Language Tax in LLM Pricing: BPE Tokenization, https\://tokenstree.com/newsletter-article-8.html

15. The Tokenizer Tax: Quantifying and Explaining the Cross-Lingual, https\://arxiv.org/pdf/2607.24276

16. SuTRA : Structurally-Unified Tokenization with Root Awareness \- arXiv, https\://arxiv.org/html/2608.18087v2

17. (PDF) Zero-Compute Cross-Lingual Transferability Estimation Using, https\://www\.researchgate.net/publication/415150750\_Zero-Compute\_Cross-Lingual\_Transferability\_Estimation\_Using\_Typological\_Feature\_Proxies

18. (PDF) THE SINDHI TOKEN TAX: ORTHOGRAPHIC PENALTY AND, https\://www\.researchgate.net/publication/415196987\_THE\_SINDHI\_TOKEN\_TAX\_ORTHOGRAPHIC\_PENALTY\_AND\_TOKENIZATION\_INEQUITY\_BETWEEN\_ENGLISH\_AND\_SINDHI\_IN\_LARGE\_LANGUAGE\_MODELS

19. LLM Safety Alignment in Low-Resource Languages \- arXiv, https\://arxiv.org/pdf/2608.14626

20. LinguaSafe: A Comprehensive Multilingual Safety Benchmark for, https\://www\.alphaxiv.org/abs/2508.12733

21. Dissecting Safety Challenges of LLMs in Multilingual Contexts, https\://www\.researchgate.net/publication/384217854\_The\_Language\_Barrier\_Dissecting\_Safety\_Challenges\_of\_LLMs\_in\_Multilingual\_Contexts

22. Dissecting Temporal and Linguistic Vulnerabilities in LLMs \- arXiv, https\://arxiv.org/pdf/2512.24556

23. Inverse Scaling of the Multilingual Performance Gap \- OpenReview, https\://openreview.net/pdf?id=bqPMaLTuuE

24. SomaliWeb v1: A Quality-Filtered Somali Web Corpus with a ... \- arXiv, https\://arxiv.org/pdf/2605.18232

25. Filtering Safety-Degrading Data for Safer LLMs \- arXiv, https\://arxiv.org/html/2609.22144v1

26. Human Rights Data Analysis Group \- Rafto, https\://www\.rafto.no/en/the-rafto-prize/human-rights-data-analysis-group

27. How can AI help our partners uncover police misconduct?, https\://hrdag.org/report/2024-review/HRDAG\_YiR24\_4pp\_SHARE.pdf

28. Resource spotlight: Data projects tracking police misconduct, use of, https\://www\.prisonpolicy.org/blog/2026/01/26/police\_misconduct/

29. Why a black man shot by Chicago police believes his case may be a, https\://www\.theguardian.com/us-news/2016/jul/26/chicago-police-shooting-antwon-golatte-unjustified

30. Zero-Knowledge-Based Policy Enforcement for Privacy-Preserving, https\://www\.mdpi.com/2079-8954/14/4/385

31. Blockchain Document Vault Development for Secure Records Hub, https\://www\.ment.tech/blockchain-document-vault-development/

32. Decentralized File Sharing Infrastructure with IPFS for Censorship, https\://journal.pandawan.id/b-front/article/view/835

33. A Decentralized Retrieval Augmented Generation System with, https\://arxiv.org/abs/2511.07577

34. Contextual Zero-Knowledge Authentication with IPFS-Backed, https\://www\.mdpi.com/2076-3417/16/9/4182

35. Data Security Archives \- Andres Andreu, CISSP, ISSAP, QTE, https\://andresandreu.tech/category/data-security/

36. A Survey on the Applications of Zero-Knowledge Proofs \- arXiv, https\://arxiv.org/html/2408.00243v2

37. Med-Chain: A Privacy-Preserving Blockchain–Cloud Hybrid, https\://www\.indjcst.com/archiver/archives/med\_chain\_a\_privacy\_preserving\_blockchain\_cloud\_hybrid\_framework\_for\_secure\_healthcare\_data\_validation\_using\_zk\_snarks.pdf

38. Enhancing Privacy and Security in Blockchain-based Health, https\://ijcaonline.org/archives/volume187/number65/bakare-2025-ijca-926101.pdf

39. Data Sharing and Verification for Certificates of Origin Based on, https\://jurnal.ubhinus.ac.id/IC-ITECHS/article/download/2558/1298

40. Decentralized RAG with Blockchain \- GitHub, https\://github.com/bayesianinstitute/Decentralized-RAG

41. A Privacy-First Architecture for Fully Local Retrieval-Augmented, https\://www\.techrxiv.org/doi/10.36227/techrxiv.176800894.46972585

42. Secure RAG: Architecture Patterns for Safe Enterprise AI, https\://petronellatech.com/blog/secure-rag-enterprise-architecture-patterns-for-accurate-leak-free-ai/

43. Decentralized AI in 2026: Benefits & Business Impact \- Kanerika, https\://kanerika.com/blogs/decentralized-ai/

44. Refusal in LLMs is mediated by a single direction \- LessWrong, https\://www\.lesswrong.com/posts/jGuXSZgv6qfdhMCuJ/refusal-in-llms-is-mediated-by-a-single-direction

45. There Is More to Refusal in Large Language Models than a Single, https\://arxiv.org/html/2602.02132v2

46. Heretic vs Abliterated LLMs: Refusal Rates & Benchmarks (2026), https\://aithinkerlab.com/heretic-ai-abliteration-benchmarks-2026/

47. Abliteration Mitigation via Refusal Aliases \- arXiv, https\://arxiv.org/html/2608.18093v1

48. Abliteration Mitigation via Refusal Aliases \- arXiv, https\://arxiv.org/pdf/2608.18093

49. There Is More to Refusal in Large Language Models than a Single, https\://arxiv.org/html/2602.02132v1

50. Benchmarking and Characterizing the Adaptive Jailbreak ... \- arXiv, https\://arxiv.org/pdf/2608.08542

51. Long Papers \- ACL Anthology, https\://aclanthology.org/volumes/2024.acl-long/

52. Fine-Tuning, RAG, and Function Calling in a Mid-Size Model \- MDPI, https\://www\.mdpi.com/2078-2489/17/3/284

53. CrunchLLM: Multitask LLMs for Structured Business Reasoning and, https\://pmc.ncbi.nlm.nih.gov/articles/PMC13215686/

54. Transformers Stop Thinking Too Early, and a Tiny LoRA Fixes It \- arXiv, https\://arxiv.org/pdf/2609.36585

55. GitHub \- Continuum-AI-Corp/OrcaBonsai-27B-Uncensored: Runtime, https\://github.com/Continuum-AI-Corp/OrcaBonsai-27B-Uncensored