{"assertions":[{"assertion_index":0,"assertion_text":"Switching on thinking, the mode where a model reasons step by step before answering, recovered roughly 40 to 65 percent of the stored-but-unretrieved facts in models tuned for it, but 11 to 12 percent still refused to surface even then, and the 2 to 5 percent of facts that were never encoded stay out of reach no matter how hard the model thinks Google Research Blog.","citation_check":null,"claim_ids":[45724,45725,45728],"link_only_urls":[]},{"assertion_index":1,"assertion_text":"It extends earlier Google work showing models like Gemini-2.5 could recall answers that were nearly unrecoverable with reasoning switched off Google Research Blog.","citation_check":null,"claim_ids":[26892],"link_only_urls":[]},{"assertion_index":2,"assertion_text":"It also fits the wider record: models have long scored better on benchmarks than in everyday use, and a Princeton study concluded a whole generation of flagships was not meaningfully more reliable than its predecessors, contradicting the vendors' launch claims Nate Jones Latent Space.","citation_check":null,"claim_ids":[1316,1318,13507],"link_only_urls":[]},{"assertion_index":3,"assertion_text":"More than 180 organizations have signed the Code of Practice that turns those duties into concrete steps EU AI Act Newsletter.","citation_check":null,"claim_ids":[45701,45702,45703],"link_only_urls":[]},{"assertion_index":4,"assertion_text":"RingCentral joined the customer roster, saying it uses ChatGPT Work and Codex, OpenAI's coding agent, to speed AI product development and pool operating knowledge across engineering and operations teams OpenAI News.","citation_check":null,"claim_ids":[45679,45680],"link_only_urls":[]},{"assertion_index":5,"assertion_text":"OpenAI also published its own research arguing enterprises are moving from AI assistance to AI execution OpenAI News.","citation_check":null,"claim_ids":[45699],"link_only_urls":[]},{"assertion_index":6,"assertion_text":"Cursor released Grok 4.5, its coding-and-agents model developed jointly with SpaceXAI, on July 8; five weeks later the pair have shipped Grok 4.6, which Cursor says is meant for agents that run longer and for work with heavier visual and interactive elements, and the cadence is the real news Cursor Blog Latent Space Cursor Blog.","citation_check":null,"claim_ids":[10623,10633,27373,27374,27380,45659],"link_only_urls":[]},{"assertion_index":7,"assertion_text":"- Google Research's encode-versus-recall gap invites replication; the same measurement run on open-weight models, the kind anyone can download and run, would show whether the finding holds beyond these two flagships. Google Research Blog","citation_check":null,"claim_ids":[45724,45725,45728],"link_only_urls":[]},{"assertion_index":8,"assertion_text":"- The European Commission's signatory list for the transparency Code of Practice, now past 180 organizations, is the roster to scan for holdout labs. EU AI Act Newsletter","citation_check":null,"claim_ids":[45701,45702,45703],"link_only_urls":[]},{"assertion_index":9,"assertion_text":"- Cursor's claim that Grok 4.6 suits long-running agents needs third-party agent evaluations before anyone standardizes on it. Cursor Blog","citation_check":null,"claim_ids":[45659],"link_only_urls":[]},{"assertion_index":10,"assertion_text":"- Frontier labs now face evidence that encrypted reasoning traces from their APIs can be decoded and replayed to weaker models; their response will show how seriously they treat leakage of a model's hidden reasoning steps. Latent Space","citation_check":null,"claim_ids":[45719],"link_only_urls":[]}],"brief_date":"2026-08-13","bundle_id":"evidence-bundle-v1-9dc6564936a0f45d5a534e85d2c71de52b359a85834f1d9977f840b5c928229e","claims":[{"assertion_indexes":[0,7],"citation_authority_claim_id":45724,"citation_checks":[],"claim_id":45724,"claim_text":"The authors state that Gemini-3-Pro and GPT-5 encode 95\u201398% of facts but fail to directly recall 26\u201334% of them.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-13T06:24:13.276207+00:00","source_name":"Google Research Blog","source_url":"https://research.google/blog/empty-shelves-or-lost-keys-recall-is-the-bottleneck-for-parametric-factuality/"},{"assertion_indexes":[0,7],"citation_authority_claim_id":45725,"citation_checks":[],"claim_id":45725,"claim_text":"The authors claim that even with 'thinking' enabled, Gemini-3-Pro and GPT-5 fail to recall 11\u201312% of facts.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-13T06:24:13.276207+00:00","source_name":"Google Research Blog","source_url":"https://research.google/blog/empty-shelves-or-lost-keys-recall-is-the-bottleneck-for-parametric-factuality/"},{"assertion_indexes":[0,7],"citation_authority_claim_id":45728,"citation_checks":[],"claim_id":45728,"claim_text":"The authors state that 'thinking' recovers roughly 40\u201365% of encoded-but-not-directly-known facts in thinking-optimized models.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-13T06:24:13.276207+00:00","source_name":"Google Research Blog","source_url":"https://research.google/blog/empty-shelves-or-lost-keys-recall-is-the-bottleneck-for-parametric-factuality/"},{"assertion_indexes":[1],"citation_authority_claim_id":26892,"citation_checks":[],"claim_id":26892,"claim_text":"Models such as Gemini-2.5 and Qwen3-32B successfully recall answers that are virtually unrecoverable when their reasoning capabilities are disabled.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-18T16:24:32.565429+00:00","source_name":"Google Research Blog","source_url":"https://research.google/blog/thinking-to-recall-how-reasoning-unlocks-parametric-knowledge-in-llms/"},{"assertion_indexes":[2],"citation_authority_claim_id":1316,"citation_checks":[],"claim_id":1316,"claim_text":"Google released Gemma 4 Quantization-Aware Training (QAT) checkpoints for local deployment.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-06-08T13:12:00.245877+00:00","source_name":"Latent Space","source_url":"https://www.latent.space/p/ainews-not-much-happened-today-6b8"},{"assertion_indexes":[2],"citation_authority_claim_id":1318,"citation_checks":[],"claim_id":1318,"claim_text":"A Princeton ICML 2026 paper update concludes that GPT 5.5, Gemini 3.1 Pro/3.5 Flash, and Claude Opus 4.7 are not meaningfully more reliable than previous models.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-06-08T13:12:00.245877+00:00","source_name":"Latent Space","source_url":"https://www.latent.space/p/ainews-not-much-happened-today-6b8"},{"assertion_indexes":[2],"citation_authority_claim_id":13507,"citation_checks":[],"claim_id":13507,"claim_text":"Current large language models perform significantly better on benchmarks than in practical, everyday scenarios.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-10T00:59:33.387078+00:00","source_name":"Nate Jones","source_url":"https://www.youtube.com/watch?v=DcrXHTOxi3I"},{"assertion_indexes":[3,8],"citation_authority_claim_id":45701,"citation_checks":[],"claim_id":45701,"claim_text":"The European Commission's AI Office, alongside national authorities, began enforcing the EU AI Act and its new transparency requirements starting on 2 August 2026.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-13T06:23:45.773934+00:00","source_name":"EU AI Act Newsletter","source_url":"https://artificialintelligenceact.substack.com/p/the-eu-ai-act-newsletter-108-enforcement"},{"assertion_indexes":[3,8],"citation_authority_claim_id":45702,"citation_checks":[],"claim_id":45702,"claim_text":"Under the new transparency rules taking effect on 2 August 2026, chatbots and interactive systems must disclose to users that they are interacting with AI, deepfakes must be labeled, and AI-generated or altered content must carry machine-readable marks for detection.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-13T06:23:45.773934+00:00","source_name":"EU AI Act Newsletter","source_url":"https://artificialintelligenceact.substack.com/p/the-eu-ai-act-newsletter-108-enforcement"},{"assertion_indexes":[3,8],"citation_authority_claim_id":45703,"citation_checks":[],"claim_id":45703,"claim_text":"The European Commission published a list of more than 180 organizations that have signed the Code of Practice operationalizing the AI Act's transparency rules.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-13T06:23:45.773934+00:00","source_name":"EU AI Act Newsletter","source_url":"https://artificialintelligenceact.substack.com/p/the-eu-ai-act-newsletter-108-enforcement"},{"assertion_indexes":[4],"citation_authority_claim_id":45679,"citation_checks":[],"claim_id":45679,"claim_text":"RingCentral utilizes ChatGPT Work and Codex to accelerate its AI product development.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-13T06:23:03.509248+00:00","source_name":"OpenAI News","source_url":"https://openai.com/index/ringcentral"},{"assertion_indexes":[4],"citation_authority_claim_id":45680,"citation_checks":[],"claim_id":45680,"claim_text":"RingCentral uses ChatGPT Work and Codex to centralize operational intelligence across its engineering and operations teams.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-13T06:23:03.509248+00:00","source_name":"OpenAI News","source_url":"https://openai.com/index/ringcentral"},{"assertion_indexes":[5],"citation_authority_claim_id":45699,"citation_checks":[],"claim_id":45699,"claim_text":"OpenAI research indicates that enterprises are adopting agentic AI, utilizing ChatGPT and Codex.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-13T06:23:26.734648+00:00","source_name":"OpenAI News","source_url":"https://openai.com/index/how-enterprises-put-ai-to-work"},{"assertion_indexes":[6],"citation_authority_claim_id":10623,"citation_checks":[],"claim_id":10623,"claim_text":"xAI launched Grok 4.5, a model positioned for coding and agent workflows, on July 8, 2026.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-09T06:29:45.745032+00:00","source_name":"Latent Space","source_url":"https://www.latent.space/p/ainews-spacexai-launches-grok-45"},{"assertion_indexes":[6],"citation_authority_claim_id":10633,"citation_checks":[],"claim_id":10633,"claim_text":"xAI and Cursor partnered to train Grok 4.5.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-09T06:29:45.745032+00:00","source_name":"Latent Space","source_url":"https://www.latent.space/p/ainews-spacexai-launches-grok-45"},{"assertion_indexes":[6],"citation_authority_claim_id":27373,"citation_checks":[],"claim_id":27373,"claim_text":"Cursor has released Grok 4.5, developed jointly with SpaceXAI.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-18T16:36:48.589620+00:00","source_name":"Cursor Blog","source_url":"https://cursor.com/blog/grok-4-5"},{"assertion_indexes":[6],"citation_authority_claim_id":27374,"citation_checks":[],"claim_id":27374,"claim_text":"Grok 4.5 is characterized as Cursor's most intelligent model to date.","correction_role":null,"label":"opinion","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-18T16:36:48.589620+00:00","source_name":"Cursor Blog","source_url":"https://cursor.com/blog/grok-4-5"},{"assertion_indexes":[6],"citation_authority_claim_id":27380,"citation_checks":[],"claim_id":27380,"claim_text":"Grok 4.5 is available on Cursor desktop, web, iOS, CLI, and SDK.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-18T16:36:48.589620+00:00","source_name":"Cursor Blog","source_url":"https://cursor.com/blog/grok-4-5"},{"assertion_indexes":[6,9],"citation_authority_claim_id":45659,"citation_checks":[],"claim_id":45659,"claim_text":"Cursor and SpaceXAI released the Grok 4.6 model on August 12, 2026.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-13T06:22:49.941988+00:00","source_name":"Cursor Blog","source_url":"https://cursor.com/blog/grok-4-6"},{"assertion_indexes":[10],"citation_authority_claim_id":45719,"citation_checks":[],"claim_id":45719,"claim_text":"A research paper demonstrates that encrypted reasoning traces from frontier API models can be decoded and replayed to weaker models to extract the hidden chain-of-thought content.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-13T06:23:56.185600+00:00","source_name":"Latent Space","source_url":"https://www.latent.space/p/ainews-how-to-steal-a-reasoning-trace"}],"version":"evidence-bundle-v1"}