{"assertions":[{"assertion_index":0,"assertion_text":"The consensus rule is fair: quantization, which stores a model's internal numbers in fewer bits so it needs less memory, saves cost but usually shaves accuracy AWS Machine Learning Blog.","citation_check":null,"claim_ids":[20202,20203],"link_only_urls":[]},{"assertion_index":1,"assertion_text":"That compressed model then beat its full-precision source on seven of nine benchmarks, including a 66.5 to 66.0 win on the LiveCodeBench coding test Hugging Face Blog.","citation_check":null,"claim_ids":[51502,51503,51504],"link_only_urls":[]},{"assertion_index":2,"assertion_text":"The idea has been building across vendors: Liquid AI has shipped 4-bit checkpoints made by distilling a high-precision teacher straight into a quantized student Hugging Face Blog.","citation_check":null,"claim_ids":[49185],"link_only_urls":[]},{"assertion_index":3,"assertion_text":"There is a deeper caution on record: one commentator argued that distilled models are narrowed compressions, strong on the tasks they trained toward and brittle outside them Nate Jones.","citation_check":null,"claim_ids":[14450],"link_only_urls":[]},{"assertion_index":4,"assertion_text":"Together's earlier quantization research found a related effect, where quantization-aware training lifts scores while drifting from the original model's actual outputs Together AI Blog.","citation_check":null,"claim_ids":[32425],"link_only_urls":[]},{"assertion_index":5,"assertion_text":"OpenRouter's own traffic shows why sticker shopping fails: coding spans nine of its 29 task types, and in late July one model led eight of them while a different one led code review and security OpenRouter Blog.","citation_check":null,"claim_ids":[51426,51427,51430],"link_only_urls":[]},{"assertion_index":6,"assertion_text":"OpenAI has spent months turning ChatGPT into a workplace agent product, and ChatGPT Work is now its named platform for office workers running AI agents TechCrunch AI.","citation_check":null,"claim_ids":[51510],"link_only_urls":[]},{"assertion_index":7,"assertion_text":"The newer move is an Admin plugin for ChatGPT Work and Codex that lets administrators analyze how a workspace uses the tools OpenAI News.","citation_check":null,"claim_ids":[51419,51420],"link_only_urls":[]},{"assertion_index":8,"assertion_text":"One in three leaders had difficulty describing the term in their own words Cohere Blog.","citation_check":null,"claim_ids":[51399,51400,51401],"link_only_urls":[]},{"assertion_index":9,"assertion_text":"That fuzziness is avoidable, because a working technical definition exists: sovereign AI means explicit control over data flow, model choice, infrastructure, observability, and operations AI Engineer.","citation_check":null,"claim_ids":[18198,18201],"link_only_urls":[]},{"assertion_index":10,"assertion_text":"- Cerebras targets a CS-5 system for 2027 and says its modular design can double token-generation speed yearly; treat the roadmap as a promise until hardware ships. Cerebras","citation_check":null,"claim_ids":[51391,51392,51393],"link_only_urls":[]},{"assertion_index":11,"assertion_text":"- Multiverse's healed 4-bit checkpoints are the thing to replicate; outside runs on the two benchmarks where the compressed model did not beat the original would map where the method stops working. Hugging Face Blog","citation_check":null,"claim_ids":[51502,51503,51504],"link_only_urls":[]},{"assertion_index":12,"assertion_text":"- OpenAI banned Russia-linked accounts that used AI to push a fake Israel-based think tank and a pro-Russia \"sovereignty\" index; influence takedowns are becoming a routine platform-risk signal. OpenAI News","citation_check":null,"claim_ids":[51418],"link_only_urls":[]},{"assertion_index":13,"assertion_text":"- EVE Online's move from Python 2 to Python 3 spans 2.4 million lines of code with roughly 20,000 spots needing human review, a public test of how far scripted migration goes before judgment takes over. Simon Willison's Weblog","citation_check":null,"claim_ids":[51651,51652,51653],"link_only_urls":[]}],"brief_date":"2026-08-26","bundle_id":"evidence-bundle-v1-5f56056c3ae246678a36a5812e08b780d43e9ff7c014b75570d03475b5a1887b","claims":[{"assertion_indexes":[0],"citation_authority_claim_id":20202,"citation_checks":[],"claim_id":20202,"claim_text":"Quantizing model weights from 16-bit to 4-bit shrinks memory usage significantly, though it can reduce model accuracy.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-11T06:17:33.089448+00:00","source_name":"AWS Machine Learning Blog","source_url":"https://aws.amazon.com/blogs/machine-learning/deploying-quantized-models-on-amazon-sagemaker-ai-with-unsloth/"},{"assertion_indexes":[0],"citation_authority_claim_id":20203,"citation_checks":[],"claim_id":20203,"claim_text":"Dynamic quantization reduces memory usage while maintaining model accuracy by applying different bit reductions to different layers based on sensitivity to precision loss.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-11T06:17:33.089448+00:00","source_name":"AWS Machine Learning Blog","source_url":"https://aws.amazon.com/blogs/machine-learning/deploying-quantized-models-on-amazon-sagemaker-ai-with-unsloth/"},{"assertion_indexes":[1,11],"citation_authority_claim_id":51502,"citation_checks":[],"claim_id":51502,"claim_text":"Multiverse Computing's Quantization-Aware Healing (QAH) method, when applied to a GPT-OSS 120B model compressed to 60B parameters and quantized to MXFP4, produces a 4-bit model that outperforms its full-precision bfloat16 source on 7 of 9 benchmarks.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-26T06:28:53.038899+00:00","source_name":"Hugging Face Blog","source_url":"https://huggingface.co/blog/MultiverseComputingCAI/quantization-aware-healing"},{"assertion_indexes":[1,11],"citation_authority_claim_id":51503,"citation_checks":[],"claim_id":51503,"claim_text":"The QAH 4-bit model surpasses the original 120B teacher model on LiveCodeBench with a score of 66.5 compared to the teacher's 66.0.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-26T06:28:53.038899+00:00","source_name":"Hugging Face Blog","source_url":"https://huggingface.co/blog/MultiverseComputingCAI/quantization-aware-healing"},{"assertion_indexes":[1,11],"citation_authority_claim_id":51504,"citation_checks":[],"claim_id":51504,"claim_text":"QAH distills from the original pre-compression model rather than the recovered checkpoint, allowing the student to pick up information the earlier recovery stage did not transfer.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-26T06:28:53.038899+00:00","source_name":"Hugging Face Blog","source_url":"https://huggingface.co/blog/MultiverseComputingCAI/quantization-aware-healing"},{"assertion_indexes":[2],"citation_authority_claim_id":49185,"citation_checks":[],"claim_id":49185,"claim_text":"Quantization-Aware Distillation (QAD) is a process where a high-precision teacher model is distilled into a quantized student model.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-20T06:32:31.101431+00:00","source_name":"Hugging Face Blog","source_url":"https://huggingface.co/blog/LiquidAI/qad"},{"assertion_indexes":[3],"citation_authority_claim_id":14450,"citation_checks":[],"claim_id":14450,"claim_text":"Distilled models are fundamentally compressions of frontier models rather than copies, resulting in a narrower 'manifold' of capability that performs well on specific trained tasks but fails outside that distribution.","correction_role":null,"label":"opinion","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-10T03:33:39.258893+00:00","source_name":"Nate Jones","source_url":"https://www.youtube.com/watch?v=0v9ixCWNhPo"},{"assertion_indexes":[4],"citation_authority_claim_id":32425,"citation_checks":[],"claim_id":32425,"claim_text":"YAQA outperforms quantization aware training (QAT) methods in minimizing KL divergence without finetuning, while QAT models may diverge from the original model's outputs despite downstream performance improvements.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-22T02:29:51.994189+00:00","source_name":"Together AI Blog","source_url":"https://www.together.ai/blog/yaqa"},{"assertion_indexes":[5],"citation_authority_claim_id":51426,"citation_checks":[],"claim_id":51426,"claim_text":"OpenRouter states that a single model led eight coding-related task types in the seven-day window ending July 25, 2026, while a different model led code review and security.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-26T06:27:03.802776+00:00","source_name":"OpenRouter Blog","source_url":"https://openrouter.ai/blog/tutorials/choose-best-ai-model/"},{"assertion_indexes":[5],"citation_authority_claim_id":51427,"citation_checks":[],"claim_id":51427,"claim_text":"OpenRouter reports that coding accounts for nine of 29 task types in its traffic classification.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-26T06:27:03.802776+00:00","source_name":"OpenRouter Blog","source_url":"https://openrouter.ai/blog/tutorials/choose-best-ai-model/"},{"assertion_indexes":[5],"citation_authority_claim_id":51430,"citation_checks":[],"claim_id":51430,"claim_text":"OpenRouter cites a 2026 study by Chen et al. titled 'The Price Reversal Phenomenon' which found that in 32% of model-pair comparisons, the model with the lower listed price incurred the higher total cost.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-26T06:27:03.802776+00:00","source_name":"OpenRouter Blog","source_url":"https://openrouter.ai/blog/tutorials/choose-best-ai-model/"},{"assertion_indexes":[6],"citation_authority_claim_id":51510,"citation_checks":[],"claim_id":51510,"claim_text":"OpenAI launched a platform called ChatGPT Work designed for white-collar workers to use AI agents.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-26T06:29:11.908853+00:00","source_name":"TechCrunch AI","source_url":"https://techcrunch.com/2026/08/25/the-world-seems-to-be-ready-an-interview-with-openai-head-of-product-thibault-sottiaux/"},{"assertion_indexes":[7],"citation_authority_claim_id":51419,"citation_checks":[],"claim_id":51419,"claim_text":"OpenAI has introduced an Admin plugin designed for use with ChatGPT Work and Codex.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-26T06:26:45.249655+00:00","source_name":"OpenAI News","source_url":"https://openai.com/index/introducing-admin-plugin"},{"assertion_indexes":[7],"citation_authority_claim_id":51420,"citation_checks":[],"claim_id":51420,"claim_text":"The Admin plugin allows users to analyze workspace usage.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-26T06:26:45.249655+00:00","source_name":"OpenAI News","source_url":"https://openai.com/index/introducing-admin-plugin"},{"assertion_indexes":[8],"citation_authority_claim_id":51399,"citation_checks":[],"claim_id":51399,"claim_text":"Cohere commissioned IDC to conduct a study on sovereign AI adoption among senior enterprise decision-makers in highly regulated industries.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-26T06:26:30.241639+00:00","source_name":"Cohere Blog","source_url":"https://cohere.com/blog/state-of-sovereign-ai-adoption-2026"},{"assertion_indexes":[8],"citation_authority_claim_id":51400,"citation_checks":[],"claim_id":51400,"claim_text":"IDC's survey of over 500 senior decision-makers was fielded from April to May 2026.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-26T06:26:30.241639+00:00","source_name":"Cohere Blog","source_url":"https://cohere.com/blog/state-of-sovereign-ai-adoption-2026"},{"assertion_indexes":[8],"citation_authority_claim_id":51401,"citation_checks":[],"claim_id":51401,"claim_text":"The IDC study found that one in three leaders had difficulty describing sovereign AI in their own words.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-26T06:26:30.241639+00:00","source_name":"Cohere Blog","source_url":"https://cohere.com/blog/state-of-sovereign-ai-adoption-2026"},{"assertion_indexes":[9],"citation_authority_claim_id":18198,"citation_checks":[],"claim_id":18198,"claim_text":"Sovereign AI is technically defined as having explicit control over data flow, model choice, infrastructure, observability, and operations. fact null null","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-11T03:48:03.856841+00:00","source_name":"AI Engineer","source_url":"https://www.youtube.com/watch?v=x2bH0RKPgdc"},{"assertion_indexes":[9],"citation_authority_claim_id":18201,"citation_checks":[],"claim_id":18201,"claim_text":"Infrastructure sovereignty involves determining where the AI application layer and compute resources are hosted. fact null null","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-11T03:48:03.856841+00:00","source_name":"AI Engineer","source_url":"https://www.youtube.com/watch?v=x2bH0RKPgdc"},{"assertion_indexes":[10],"citation_authority_claim_id":51391,"citation_checks":[],"claim_id":51391,"claim_text":"Cerebras introduced the CS-4 AI accelerator, which is the first system built on the Cerebras Nexus rack-scale platform. (fact, 2026-08-18T00:00:00+00:00, null, null)","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-26T06:26:07.680952+00:00","source_name":"Cerebras","source_url":"https://www.cerebras.ai/blog/ultrafast-frontier-inference-cerebras-deep-dive-at-hot-chips-2026"},{"assertion_indexes":[10],"citation_authority_claim_id":51392,"citation_checks":[],"claim_id":51392,"claim_text":"The CS-4 system utilizes a modular power, cooling, and I/O architecture designed to double token-generation speed year over year for several years. (forecast, null, null, null)","correction_role":null,"label":"forecast","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-26T06:26:07.680952+00:00","source_name":"Cerebras","source_url":"https://www.cerebras.ai/blog/ultrafast-frontier-inference-cerebras-deep-dive-at-hot-chips-2026"},{"assertion_indexes":[10],"citation_authority_claim_id":51393,"citation_checks":[],"claim_id":51393,"claim_text":"Cerebras targets the CS-5 system for release in 2027. (forecast, null, 2027-12-31, null)","correction_role":null,"label":"forecast","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-26T06:26:07.680952+00:00","source_name":"Cerebras","source_url":"https://www.cerebras.ai/blog/ultrafast-frontier-inference-cerebras-deep-dive-at-hot-chips-2026"},{"assertion_indexes":[12],"citation_authority_claim_id":51418,"citation_checks":[],"claim_id":51418,"claim_text":"OpenAI banned accounts originating from Russia that used artificial intelligence to promote a fake Israel-based think tank and a 'sovereignty' index praising Russia and criticizing the West.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-26T06:26:38.520036+00:00","source_name":"OpenAI News","source_url":"https://openai.com/index/disrupting-malicious-uses-of-ai-influence-campaign-russia"},{"assertion_indexes":[13],"citation_authority_claim_id":51651,"citation_checks":[],"claim_id":51651,"claim_text":"EVE Online is initiating a migration from Python 2 to Python 3, a process that involves running the futurize script on approximately 2.4 million lines of code.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-26T06:33:27.770454+00:00","source_name":"Simon Willison's Weblog","source_url":"https://simonwillison.net/2026/Aug/25/eve-online-move-to-python-3/"},{"assertion_indexes":[13],"citation_authority_claim_id":51652,"citation_checks":[],"claim_id":51652,"claim_text":"The EVE Online migration to Python 3 requires manual review of approximately 20,000 code locations where behavior differs between Python 2 and Python 3.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-26T06:33:27.770454+00:00","source_name":"Simon Willison's Weblog","source_url":"https://simonwillison.net/2026/Aug/25/eve-online-move-to-python-3/"},{"assertion_indexes":[13],"citation_authority_claim_id":51653,"citation_checks":[],"claim_id":51653,"claim_text":"EVE Online has utilized Stackless Python since its launch in 2003, with its last major upgrade to Stackless Python 2.7 occurring in 2010.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-26T06:33:27.770454+00:00","source_name":"Simon Willison's Weblog","source_url":"https://simonwillison.net/2026/Aug/25/eve-online-move-to-python-3/"}],"version":"evidence-bundle-v1"}