{"assertions":[{"assertion_index":0,"assertion_text":"OpenRouter Blog The lines to the labs were open.","citation_check":null,"claim_ids":[64404,64405,64406,64407,64411],"link_only_urls":[]},{"assertion_index":1,"assertion_text":"OpenRouter Blog OpenRouter Case Study I read the spend cut as downstream of the speed.","citation_check":null,"claim_ids":[63381,63382,63383,63384,63386,63387,63389,64404,64405,64406,64407,64411],"link_only_urls":[]},{"assertion_index":2,"assertion_text":"OpenRouter Tutorial A cheap model you can't test cheaply never gets used.","citation_check":null,"claim_ids":[51437],"link_only_urls":[]},{"assertion_index":3,"assertion_text":"OpenRouter Case Study Catalogs lag.","citation_check":null,"claim_ids":[63381,63382,63383,63384,63386,63387,63389],"link_only_urls":[]},{"assertion_index":4,"assertion_text":"OpenRouter Blog","citation_check":null,"claim_ids":[64404,64405,64406,64407,64411],"link_only_urls":[]},{"assertion_index":5,"assertion_text":"AI Engineer The published paper turns that into a method.","citation_check":null,"claim_ids":[19374,19375],"link_only_urls":[]},{"assertion_index":6,"assertion_text":"Hugging Face Daily Papers Six recorded failures replayed through simulated model boundaries is a small proof.","citation_check":null,"claim_ids":[64336,64337,64338,64339,64340,64341,64342],"link_only_urls":[]},{"assertion_index":7,"assertion_text":"Hugging Face Daily Papers Those targets are already a generation old.","citation_check":null,"claim_ids":[64210,64211,64212,64213],"link_only_urls":[]},{"assertion_index":8,"assertion_text":"Google DeepMind Blog OpenAI News An engineer picking a voice model should rerun the interruption test against those two before a small model earns a slot.","citation_check":null,"claim_ids":[59828,61826],"link_only_urls":[]},{"assertion_index":9,"assertion_text":"Nate Jones Cost is the reason they can't feel it yet.","citation_check":null,"claim_ids":[64347,64348,64349],"link_only_urls":[]},{"assertion_index":10,"assertion_text":"Nate Jones For a business leader placing next quarter's AI budget, that is the audit.","citation_check":null,"claim_ids":[37716],"link_only_urls":[]},{"assertion_index":11,"assertion_text":"- ScrollEd says it plans to launch its consumer feed at TechCrunch's October conference. If that launch happens, it will show whether turning textbooks into short video has an institutional buyer or only scrollers. TechCrunch AI","citation_check":null,"claim_ids":[64351,64352,64353,64354,64355,64356,64358],"link_only_urls":[]},{"assertion_index":12,"assertion_text":"- Google's Gemini 3.8 Live has been public since September 15, per Google's announcement. The next full-duplex benchmark round should include it. Google DeepMind Blog","citation_check":null,"claim_ids":[61826],"link_only_urls":[]},{"assertion_index":13,"assertion_text":"- The Chronicle code and its six-incident benchmark are public. Any team that replays its own agent traces through it will tell us whether zero-model-call replay survives real boundaries. Hugging Face Daily Papers","citation_check":null,"claim_ids":[64336,64337,64338,64339,64340,64341,64342],"link_only_urls":[]},{"assertion_index":14,"assertion_text":"- vLLM's 0.30.0 release lands a build fix for DeepGEMM on CUDA 12.9. Teams that serve open models on their own chips should confirm their pipelines pick it up before the next release candidate. vLLM Releases","citation_check":null,"claim_ids":[64412,64413],"link_only_urls":[]},{"assertion_index":15,"assertion_text":"OpenRouter Blog Chronicle's results rest on six recorded failures with simulated model boundaries.","citation_check":null,"claim_ids":[64404,64405,64406,64407,64411],"link_only_urls":[]},{"assertion_index":16,"assertion_text":"Hugging Face Daily Papers","citation_check":null,"claim_ids":[64336,64337,64338,64339,64340,64341,64342],"link_only_urls":[]}],"brief_date":"2026-09-21","bundle_id":"evidence-bundle-v1-9d3b65cacbe02a7e47d796116f4ab91b8dc2649f2655393bed1502024be4d151","claims":[{"assertion_indexes":[0,1,4,15],"citation_authority_claim_id":64404,"citation_checks":[],"claim_id":64404,"claim_text":"Descript previously maintained direct integrations with OpenAI, Anthropic, and Google for its video editing agent Underlord.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:37:32.564959+00:00","source_name":"OpenRouter Blog","source_url":"https://openrouter.ai/blog/insights/descript-model-evaluation-queue/"},{"assertion_indexes":[0,1,4,15],"citation_authority_claim_id":64405,"citation_checks":[],"claim_id":64405,"claim_text":"Before switching to OpenRouter, Descript's process for adding a new model required engineering time that often resulted in a waiting period of up to a week.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:37:32.564959+00:00","source_name":"OpenRouter Blog","source_url":"https://openrouter.ai/blog/insights/descript-model-evaluation-queue/"},{"assertion_indexes":[0,1,4,15],"citation_authority_claim_id":64406,"citation_checks":[],"claim_id":64406,"claim_text":"Aleks Mistratov, Head of AI Product at Descript, stated that the team can now run evaluations on a new model within an hour or two via a Slack-integrated loop.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:37:32.564959+00:00","source_name":"OpenRouter Blog","source_url":"https://openrouter.ai/blog/insights/descript-model-evaluation-queue/"},{"assertion_indexes":[0,1,4,15],"citation_authority_claim_id":64407,"citation_checks":[],"claim_id":64407,"claim_text":"Descript uses Claude Tag, Anthropic's Slack integration, to automate the execution of model evaluations and the opening of pull requests.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:37:32.564959+00:00","source_name":"OpenRouter Blog","source_url":"https://openrouter.ai/blog/insights/descript-model-evaluation-queue/"},{"assertion_indexes":[0,1,4,15],"citation_authority_claim_id":64411,"citation_checks":[],"claim_id":64411,"claim_text":"Descript uses Baseten for dedicated deployments of open-weight models, with OpenRouter routing to Baseten first and other providers as fallbacks.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:37:32.564959+00:00","source_name":"OpenRouter Blog","source_url":"https://openrouter.ai/blog/insights/descript-model-evaluation-queue/"},{"assertion_indexes":[1,3],"citation_authority_claim_id":63381,"citation_checks":[],"claim_id":63381,"claim_text":"As of August 2026, Descript runs 13 models in production via OpenRouter.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-19T06:37:39.916551+00:00","source_name":"OpenRouter Blog","source_url":"https://openrouter.ai/blog/case-studies/descript-case-study/"},{"assertion_indexes":[1,3],"citation_authority_claim_id":63382,"citation_checks":[],"claim_id":63382,"claim_text":"Descript reduced the time to evaluate a new model from one week or more to 1-2 hours by using OpenRouter.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-19T06:37:39.916551+00:00","source_name":"OpenRouter Blog","source_url":"https://openrouter.ai/blog/case-studies/descript-case-study/"},{"assertion_indexes":[1,3],"citation_authority_claim_id":63383,"citation_checks":[],"claim_id":63383,"claim_text":"Descript's Underlord video editing agent uses OpenRouter as its production inference layer.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-19T06:37:39.916551+00:00","source_name":"OpenRouter Blog","source_url":"https://openrouter.ai/blog/case-studies/descript-case-study/"},{"assertion_indexes":[1,3],"citation_authority_claim_id":63384,"citation_checks":[],"claim_id":63384,"claim_text":"Descript reduced its total model spend to under half of its previous level by utilizing open-weight models via OpenRouter.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-19T06:37:39.916551+00:00","source_name":"OpenRouter Blog","source_url":"https://openrouter.ai/blog/case-studies/descript-case-study/"},{"assertion_indexes":[1,3],"citation_authority_claim_id":63386,"citation_checks":[],"claim_id":63386,"claim_text":"Aleks Mistratov, Head of AI Product at Descript, states that if an inference provider is not on OpenRouter, he waits on their offer unless it is extremely advantageous.","correction_role":null,"label":"opinion","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-19T06:37:39.916551+00:00","source_name":"OpenRouter Blog","source_url":"https://openrouter.ai/blog/case-studies/descript-case-study/"},{"assertion_indexes":[1,3],"citation_authority_claim_id":63387,"citation_checks":[],"claim_id":63387,"claim_text":"Keith Simmons, Technical Lead of Descript\u2019s AI Enablement team, considers the ability to easily integrate new models critical to Descript's business due to the explosion of open-weight models.","correction_role":null,"label":"opinion","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-19T06:37:39.916551+00:00","source_name":"OpenRouter Blog","source_url":"https://openrouter.ai/blog/case-studies/descript-case-study/"},{"assertion_indexes":[1,3],"citation_authority_claim_id":63389,"citation_checks":[],"claim_id":63389,"claim_text":"Descript uses OpenRouter to route uploaded video through Gemini for content analysis.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-19T06:37:39.916551+00:00","source_name":"OpenRouter Blog","source_url":"https://openrouter.ai/blog/case-studies/descript-case-study/"},{"assertion_indexes":[2],"citation_authority_claim_id":51437,"citation_checks":[],"claim_id":51437,"claim_text":"OpenRouter asserts that a model with a low unit price can become expensive if it retries, produces long completions, or requires a stronger model to catch failures.","correction_role":null,"label":"opinion","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":19374,"citation_checks":[],"claim_id":19374,"claim_text":"The Chronicle framework uses a 'boundary' annotation to freeze and record the full state of agent runs, including model versions and code versions.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-11T04:15:01.497367+00:00","source_name":"AI Engineer","source_url":"https://www.youtube.com/watch?v=Lc8zRh9muoY"},{"assertion_indexes":[5],"citation_authority_claim_id":19375,"citation_checks":[],"claim_id":19375,"claim_text":"Chronicle allows stubbing non-modified nodes in a recorded trace to enable deterministic regression testing of specific code changes.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-11T04:15:01.497367+00:00","source_name":"AI Engineer","source_url":"https://www.youtube.com/watch?v=Lc8zRh9muoY"},{"assertion_indexes":[6,13,16],"citation_authority_claim_id":64336,"citation_checks":[],"claim_id":64336,"claim_text":"The Chronicle system records agent runs at non-deterministic boundaries as immutable envelopes to enable replay.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:34:56.990669+00:00","source_name":"Hugging Face Daily Papers","source_url":"https://tldr.takara.ai/p/2609.20625"},{"assertion_indexes":[6,13,16],"citation_authority_claim_id":64337,"citation_checks":[],"claim_id":64337,"claim_text":"Chronicle's cut-point replay operation serves a chosen subset of recorded boundaries while executing the complementary subset live with new code to turn recorded incidents into regression tests for continuous integration.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:34:56.990669+00:00","source_name":"Hugging Face Daily Papers","source_url":"https://tldr.takara.ai/p/2609.20625"},{"assertion_indexes":[6,13,16],"citation_authority_claim_id":64338,"citation_checks":[],"claim_id":64338,"claim_text":"On a benchmark of 6 recorded failures with simulated model boundaries, Chronicle's recording adds 23 microseconds per boundary crossing, which represents 0.008% of an assumed 300 ms model call.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:34:56.990669+00:00","source_name":"Hugging Face Daily Papers","source_url":"https://tldr.takara.ai/p/2609.20625"},{"assertion_indexes":[6,13,16],"citation_authority_claim_id":64339,"citation_checks":[],"claim_id":64339,"claim_text":"Full replay by Chronicle issues zero model calls and remains bit-stable across 20 repetitions.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:34:56.990669+00:00","source_name":"Hugging Face Daily Papers","source_url":"https://tldr.takara.ai/p/2609.20625"},{"assertion_indexes":[6,13,16],"citation_authority_claim_id":64340,"citation_checks":[],"claim_id":64340,"claim_text":"Chronicle's cut-point tests fail on faulty code and pass on guarded and benign changes for all 6 incidents in the benchmark.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:34:56.990669+00:00","source_name":"Hugging Face Daily Papers","source_url":"https://tldr.takara.ai/p/2609.20625"},{"assertion_indexes":[6,13,16],"citation_authority_claim_id":64341,"citation_checks":[],"claim_id":64341,"claim_text":"In a mutation study of guarded tools, Chronicle's cut-point tests catch every mutant that allows a recorded unsafe action through, whereas a baseline that stubs every boundary catches none.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:34:56.990669+00:00","source_name":"Hugging Face Daily Papers","source_url":"https://tldr.takara.ai/p/2609.20625"},{"assertion_indexes":[6,13,16],"citation_authority_claim_id":64342,"citation_checks":[],"claim_id":64342,"claim_text":"Chronicle and its associated benchmark are publicly available at the URL https://github.com/theagentplane/chronicle.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:34:56.990669+00:00","source_name":"Hugging Face Daily Papers","source_url":"https://tldr.takara.ai/p/2609.20625"},{"assertion_indexes":[7],"citation_authority_claim_id":64210,"citation_checks":[],"claim_id":64210,"claim_text":"Realtime-Venus-Omni achieved the highest scores among evaluated online models on six of eight video benchmarks, including 70.2% on StreamingBench, 64.7% on OVO-Bench, and 81.3% on Daily-Omni.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:30:51.137471+00:00","source_name":"Hugging Face Daily Papers","source_url":"https://tldr.takara.ai/p/2609.13814"},{"assertion_indexes":[7],"citation_authority_claim_id":64211,"citation_checks":[],"claim_id":64211,"claim_text":"Realtime-Venus-Audio led compared models on MMAU (78.0%), MMAU-Pro (63.2%), Llama Questions (83.8%), and Speech CMMLU (67.8%), and matched the best VoiceBench AlpacaEval score of 4.81.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:30:51.137471+00:00","source_name":"Hugging Face Daily Papers","source_url":"https://tldr.takara.ai/p/2609.13814"},{"assertion_indexes":[7],"citation_authority_claim_id":64212,"citation_checks":[],"claim_id":64212,"claim_text":"On Full-Duplex-Bench v1.5, Realtime-Venus-Audio responded to 75% of user interruptions and achieved continuation rates of 97%, 88%, and 86% under backchannels, other-directed speech, and background speech, respectively, exceeding Gemini 3.1 Live and GPT-4o on these metrics.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:30:51.137471+00:00","source_name":"Hugging Face Daily Papers","source_url":"https://tldr.takara.ai/p/2609.13814"},{"assertion_indexes":[7],"citation_authority_claim_id":64213,"citation_checks":[],"claim_id":64213,"claim_text":"Realtime-Venus is a proactive full-duplex interaction system comprising two separately trained 9B models: Realtime-Venus-Omni for audio-visual interaction and Realtime-Venus-Audio for spoken interaction.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:30:51.137471+00:00","source_name":"Hugging Face Daily Papers","source_url":"https://tldr.takara.ai/p/2609.13814"},{"assertion_indexes":[8],"citation_authority_claim_id":59828,"citation_checks":[],"claim_id":59828,"claim_text":"The GPT-Live-1 API feature enables natural, full-duplex voice conversations.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-11T06:42:31.817064+00:00","source_name":"OpenAI News","source_url":"https://openai.com/index/introducing-gpt-live-1-in-the-api"},{"assertion_indexes":[8,12],"citation_authority_claim_id":61826,"citation_checks":[],"claim_id":61826,"claim_text":"Gemini 3.8 Live Extended Thinking scored 97.7% on the Big Bench Audio reasoning benchmark.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-16T06:38:03.085291+00:00","source_name":"Google DeepMind Blog","source_url":"https://deepmind.google/blog/introducing-gemini-3-8-live-and-3-8-live-extended-thinking/"},{"assertion_indexes":[9],"citation_authority_claim_id":64347,"citation_checks":[],"claim_id":64347,"claim_text":"Salesforce executive Rohan Kumar reported that a customer expressed confusion about how the transformative potential of AI aligns with the fact that AI products currently look much the same as previous versions.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:35:54.580973+00:00","source_name":"Nate Jones","source_url":"https://natesnewsletter.substack.com/p/ai-cost-to-serve-customers"},{"assertion_indexes":[9],"citation_authority_claim_id":64348,"citation_checks":[],"claim_id":64348,"claim_text":"Historically, the high expense of understanding customer problems and coordinating responses determined business decisions regarding which customers received human support, which received forms, and which received custom responses.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:35:54.580973+00:00","source_name":"Nate Jones","source_url":"https://natesnewsletter.substack.com/p/ai-cost-to-serve-customers"},{"assertion_indexes":[9],"citation_authority_claim_id":64349,"citation_checks":[],"claim_id":64349,"claim_text":"The author argues that if AI technology is truly consequential, customers should eventually receive meaningfully different results, such as outcomes previously unattainable, prices previously unjustifiable, or attention levels previously unaffordable.","correction_role":null,"label":"opinion","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:35:54.580973+00:00","source_name":"Nate Jones","source_url":"https://natesnewsletter.substack.com/p/ai-cost-to-serve-customers"},{"assertion_indexes":[10],"citation_authority_claim_id":37716,"citation_checks":[],"claim_id":37716,"claim_text":"A time study conducted by the author found that reconstructing customer data across multiple systems was the expensive part of a support ticket, while composing the reply was the cheap part.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-27T09:48:24.630438+00:00","source_name":"Nate Jones","source_url":"https://natesnewsletter.substack.com/p/first-ai-agent-use-case"},{"assertion_indexes":[11],"citation_authority_claim_id":64351,"citation_checks":[],"claim_id":64351,"claim_text":"ScrollEd is a Palo Alto-based startup pitching in the TechCrunch Disrupt Startup Battlefield 200 competition.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:36:08.161087+00:00","source_name":"TechCrunch AI","source_url":"https://techcrunch.com/2026/09/20/scrolled-wants-to-turn-textbooks-into-tiktok/"},{"assertion_indexes":[11],"citation_authority_claim_id":64352,"citation_checks":[],"claim_id":64352,"claim_text":"ScrollEd has built an app that converts text files, including PDFs and textbooks, into a vertical feed resembling Instagram Reels or TikTok.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:36:08.161087+00:00","source_name":"TechCrunch AI","source_url":"https://techcrunch.com/2026/09/20/scrolled-wants-to-turn-textbooks-into-tiktok/"},{"assertion_indexes":[11],"citation_authority_claim_id":64353,"citation_checks":[],"claim_id":64353,"claim_text":"The ScrollEd app uses AI to generate video, audio, text, or interactive quizzes from source content.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:36:08.161087+00:00","source_name":"TechCrunch AI","source_url":"https://techcrunch.com/2026/09/20/scrolled-wants-to-turn-textbooks-into-tiktok/"},{"assertion_indexes":[11],"citation_authority_claim_id":64354,"citation_checks":[],"claim_id":64354,"claim_text":"ScrollEd co-founders Utsav Gupta and Rebecca Neff are students at Stanford University and the University of Pennsylvania, respectively.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:36:08.161087+00:00","source_name":"TechCrunch AI","source_url":"https://techcrunch.com/2026/09/20/scrolled-wants-to-turn-textbooks-into-tiktok/"},{"assertion_indexes":[11],"citation_authority_claim_id":64355,"citation_checks":[],"claim_id":64355,"claim_text":"ScrollEd follows a freemium business model combining a free consumer feed, a paid ScrollEd Pro subscription, and annual institutional licenses.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:36:08.161087+00:00","source_name":"TechCrunch AI","source_url":"https://techcrunch.com/2026/09/20/scrolled-wants-to-turn-textbooks-into-tiktok/"},{"assertion_indexes":[11],"citation_authority_claim_id":64356,"citation_checks":[],"claim_id":64356,"claim_text":"ScrollEd intends to launch its consumer business at TechCrunch Disrupt.","correction_role":null,"label":"forecast","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:36:08.161087+00:00","source_name":"TechCrunch AI","source_url":"https://techcrunch.com/2026/09/20/scrolled-wants-to-turn-textbooks-into-tiktok/"},{"assertion_indexes":[11],"citation_authority_claim_id":64358,"citation_checks":[],"claim_id":64358,"claim_text":"ScrollEd's co-founder Utsav Gupta stated that the company aims to build a social network that is better for its users.","correction_role":null,"label":"opinion","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:36:08.161087+00:00","source_name":"TechCrunch AI","source_url":"https://techcrunch.com/2026/09/20/scrolled-wants-to-turn-textbooks-into-tiktok/"},{"assertion_indexes":[14],"citation_authority_claim_id":64412,"citation_checks":[],"claim_id":64412,"claim_text":"The vLLM project released version 0.30.0.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:37:39.458089+00:00","source_name":"vLLM Releases","source_url":"https://github.com/vllm-project/vllm/releases/tag/v0.30.0"},{"assertion_indexes":[14],"citation_authority_claim_id":64413,"citation_checks":[],"claim_id":64413,"claim_text":"The release of vLLM version 0.30.0 addresses a build issue for DeepGEMM CUDA 12.9 release builds.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-09-21T06:37:39.458089+00:00","source_name":"vLLM Releases","source_url":"https://github.com/vllm-project/vllm/releases/tag/v0.30.0"}],"version":"evidence-bundle-v1"}