Chain of evidence
This frozen page shows Augur's claims and source links for one sent dispatch. Stored spot-checks appear only where the frozen edition supports them; absence is not presented as verification.
As of:
Bundle identity: evidence-bundle-v1-3b1eb090afe42261be84b28d4ee8d7e2791d43231d5d62e481062c16dec4f173
Format: evidence-bundle-v1 · 47 claims
Assertion 1
Assertion status: No spot-check verdict is published for this assertion.
DeepSeek retired the DeepSeek V4 Pro model in favor of the v4.1-Flash release.
No stored spot-check names this claim in this edition.
Assertion 2
Assertion status: No spot-check verdict is published for this assertion.
Fireworks released the DeepSeek-V4.1-Flash model.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark, DeepSeek-V4.1-Flash achieved a Pass@1 accuracy of 74.34%.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark, GPT-6-Astra achieved a Pass@1 accuracy of 74.12%.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark, Gemini 3.8 Flash achieved a Pass@1 accuracy of 73.83%.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark, Claude Opus 5 achieved a Pass@1 accuracy of 73.65%.
No stored spot-check names this claim in this edition.
DeepSeek-V4.1-Flash costs $0.430 per task in the DeepSWE benchmark.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark run, DeepSeek-V4.1-Flash consumed 36.9M input tokens and 211K output tokens per task.
No stored spot-check names this claim in this edition.
In the Terminal-Bench 2.1 benchmark, DeepSeek-V4.1-Flash achieved an accuracy of 86.5%.
No stored spot-check names this claim in this edition.
In the Terminal-Bench 2.1 benchmark, GPT-6-Astra achieved an accuracy of 87.5%.
No stored spot-check names this claim in this edition.
In the Humanity’s Last Exam (HLE) benchmark, DeepSeek-V4.1-Flash achieved a score of 34.52%.
No stored spot-check names this claim in this edition.
Fireworks plans to release US-hosted endpoints for DeepSeek-V4.1-Flash for regulated industries.
No stored spot-check names this claim in this edition.
Fireworks plans to release training support for DeepSeek-V4.1-Flash.
No stored spot-check names this claim in this edition.
Assertion 3
Assertion status: No spot-check verdict is published for this assertion.
Fireworks released the DeepSeek-V4.1-Flash model.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark, DeepSeek-V4.1-Flash achieved a Pass@1 accuracy of 74.34%.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark, GPT-6-Astra achieved a Pass@1 accuracy of 74.12%.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark, Gemini 3.8 Flash achieved a Pass@1 accuracy of 73.83%.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark, Claude Opus 5 achieved a Pass@1 accuracy of 73.65%.
No stored spot-check names this claim in this edition.
DeepSeek-V4.1-Flash costs $0.430 per task in the DeepSWE benchmark.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark run, DeepSeek-V4.1-Flash consumed 36.9M input tokens and 211K output tokens per task.
No stored spot-check names this claim in this edition.
In the Terminal-Bench 2.1 benchmark, DeepSeek-V4.1-Flash achieved an accuracy of 86.5%.
No stored spot-check names this claim in this edition.
In the Terminal-Bench 2.1 benchmark, GPT-6-Astra achieved an accuracy of 87.5%.
No stored spot-check names this claim in this edition.
In the Humanity’s Last Exam (HLE) benchmark, DeepSeek-V4.1-Flash achieved a score of 34.52%.
No stored spot-check names this claim in this edition.
Fireworks plans to release US-hosted endpoints for DeepSeek-V4.1-Flash for regulated industries.
No stored spot-check names this claim in this edition.
Fireworks plans to release training support for DeepSeek-V4.1-Flash.
No stored spot-check names this claim in this edition.
Assertion 4
Assertion status: No spot-check verdict is published for this assertion.
DeepSeek released an update to its V4 Flash model on July 31, 2026, which the source claims beats OpenAI's Luna at the Pareto frontier.
No stored spot-check names this claim in this edition.
In a Together AI benchmark on 113 DeepSWE tasks, GPT-5.6 Luna achieved a pass@1 rate of 67.2%, which was 14 points higher than DeepSeek-V4 Flash 0731's 53.3%.
No stored spot-check names this claim in this edition.
DeepSeek-V4-Flash 0731 outperforms the leading proprietary model tested on the ADE-bench benchmark, scoring 74.4%, indicating that the open source model frontier has caught up for data team tasks.
No stored spot-check names this claim in this edition.
Snowflake serves these open models itself rather than proxying third-party APIs to keep inference within a secure enterprise data governance perimeter.
No stored spot-check names this claim in this edition.
Snowflake's approach keeps inference next to governed data, reducing transfer costs and latency while maintaining role-based access control and audit trails for open model usage.
No stored spot-check names this claim in this edition.
Snowflake claims their system will allow organizations to progressively lower the average cost per business outcome while benefiting from advances in model performance, without requiring application redesign or manual routing logic management.
No stored spot-check names this claim in this edition.
Assertion 5
Assertion status: No spot-check verdict is published for this assertion.
Fireworks released the DeepSeek-V4.1-Flash model.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark, DeepSeek-V4.1-Flash achieved a Pass@1 accuracy of 74.34%.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark, GPT-6-Astra achieved a Pass@1 accuracy of 74.12%.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark, Gemini 3.8 Flash achieved a Pass@1 accuracy of 73.83%.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark, Claude Opus 5 achieved a Pass@1 accuracy of 73.65%.
No stored spot-check names this claim in this edition.
DeepSeek-V4.1-Flash costs $0.430 per task in the DeepSWE benchmark.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark run, DeepSeek-V4.1-Flash consumed 36.9M input tokens and 211K output tokens per task.
No stored spot-check names this claim in this edition.
In the Terminal-Bench 2.1 benchmark, DeepSeek-V4.1-Flash achieved an accuracy of 86.5%.
No stored spot-check names this claim in this edition.
In the Terminal-Bench 2.1 benchmark, GPT-6-Astra achieved an accuracy of 87.5%.
No stored spot-check names this claim in this edition.
In the Humanity’s Last Exam (HLE) benchmark, DeepSeek-V4.1-Flash achieved a score of 34.52%.
No stored spot-check names this claim in this edition.
Fireworks plans to release US-hosted endpoints for DeepSeek-V4.1-Flash for regulated industries.
No stored spot-check names this claim in this edition.
Fireworks plans to release training support for DeepSeek-V4.1-Flash.
No stored spot-check names this claim in this edition.
Assertion 6
Assertion status: No spot-check verdict is published for this assertion.
Anthropic CEO Dario Amodei publicly called for slowing advances in the most capable AI systems to allow safety work to catch up.
No stored spot-check names this claim in this edition.
Dario Amodei proposed that AI safety efforts should involve outside evaluators, coordination among democratic countries, and international cooperation.
No stored spot-check names this claim in this edition.
Dario Amodei published an essay titled 'We Must Pace the Frontier', naming his proposal after a letter signed by lab employees in July.
No stored spot-check names this claim in this edition.
Amodei’s proposal aims to slow the rate of AI capability improvements to allow time for alignment and safety work.
No stored spot-check names this claim in this edition.
Amodei unilaterally committed to his first proposal, which OpenAI followed, while Elon Musk and Demis Hassabis endorsed the overall proposal.
No stored spot-check names this claim in this edition.
Assertion 7
Assertion status: No spot-check verdict is published for this assertion.
Huang's views on slowing AI improvement diverge from those of Anthropic CEO Dario Amodei, SpaceX CEO Elon Musk, and OpenAI CEO Sam Altman.
No stored spot-check names this claim in this edition.
Assertion 8
Assertion status: No spot-check verdict is published for this assertion.
Mango Excellent Media's TV adaptation of The Later Journey to the West premiered on August 31, marking the first fully AI-generated entertainment to reach mainstream provincial satellite prime time in China.
No stored spot-check names this claim in this edition.
The production of the AI-generated series took six months, a duration significantly shorter than the one to two years required for traditional TV dramas.
No stored spot-check names this claim in this edition.
ByteDance's Seedance 2.0 and 2.5 models were used to generate all visuals for the series, eliminating the need for live-action actors or filmed footage.
No stored spot-check names this claim in this edition.
Mango Excellent Media's market value surged by 13.3 billion RMB over five days following the premiere of the AI-generated series.
No stored spot-check names this claim in this edition.
In the first half of 2026, more than 74 percent of the 367,000 microdramas released on Chinese platforms were AI-generated.
No stored spot-check names this claim in this edition.
Assertion 9
Assertion status: No spot-check verdict is published for this assertion.
Seedance 2.5 generates video clips with a duration ranging from 4 to 30 seconds.
No stored spot-check names this claim in this edition.
Seedance 2.5 accepts image, video, and audio assets as input references.
No stored spot-check names this claim in this edition.
Seedance 2.5 generates audio in the same pass as video without an additional charge.
No stored spot-check names this claim in this edition.
Assertion 10
Assertion status: No spot-check verdict is published for this assertion.
Snowflake introduced the Data Connectivity Proxy (DCP) for its Snowflake Openflow service to enable access to private data sources that were previously unreachable.
No stored spot-check names this claim in this edition.
Snowflake Openflow is described as a fully managed data integration service that connects data without requiring custom ETL infrastructure.
No stored spot-check names this claim in this edition.
The DCP agent operates by establishing an outbound TLS connection on port 443 from the customer's environment to Snowflake, eliminating the need for inbound firewall rules.
No stored spot-check names this claim in this edition.
DCP is designed to function in deny-all-inbound network environments where opening inbound ports is restricted by regulatory or security policies.
No stored spot-check names this claim in this edition.
The DCP agent is a TCP pass-through that does not inspect, parse, or store data, with source credentials remaining in Snowflake and never landing on the agent.
No stored spot-check names this claim in this edition.
Mutual TLS (mTLS) is used for identity verification between the DCP agent and Snowflake, with Snowflake rotating agent certificates in the background.
No stored spot-check names this claim in this edition.
DCP supports connectivity to on-premises databases (including Oracle, SQL Server, PostgreSQL, and MySQL), private Kafka brokers, and cross-cloud sources.
No stored spot-check names this claim in this edition.
Routing the DCP agent's outbound traffic through an enterprise-managed forward proxy (HTTPS_PROXY or CONNECT-based) is not supported.
No stored spot-check names this claim in this edition.
DCP is generally available on AWS, Azure, and GCP for Snowflake customers on Standard, Enterprise, and Business-Critical editions.
No stored spot-check names this claim in this edition.
The DCP agent is delivered as a Docker image that runs on any Linux host with Docker or another OCI-compatible runtime, without requiring Kubernetes.
No stored spot-check names this claim in this edition.
Assertion 11
Assertion status: No spot-check verdict is published for this assertion.
Fireworks released the DeepSeek-V4.1-Flash model.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark, DeepSeek-V4.1-Flash achieved a Pass@1 accuracy of 74.34%.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark, GPT-6-Astra achieved a Pass@1 accuracy of 74.12%.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark, Gemini 3.8 Flash achieved a Pass@1 accuracy of 73.83%.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark, Claude Opus 5 achieved a Pass@1 accuracy of 73.65%.
No stored spot-check names this claim in this edition.
DeepSeek-V4.1-Flash costs $0.430 per task in the DeepSWE benchmark.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark run, DeepSeek-V4.1-Flash consumed 36.9M input tokens and 211K output tokens per task.
No stored spot-check names this claim in this edition.
In the Terminal-Bench 2.1 benchmark, DeepSeek-V4.1-Flash achieved an accuracy of 86.5%.
No stored spot-check names this claim in this edition.
In the Terminal-Bench 2.1 benchmark, GPT-6-Astra achieved an accuracy of 87.5%.
No stored spot-check names this claim in this edition.
In the Humanity’s Last Exam (HLE) benchmark, DeepSeek-V4.1-Flash achieved a score of 34.52%.
No stored spot-check names this claim in this edition.
Fireworks plans to release US-hosted endpoints for DeepSeek-V4.1-Flash for regulated industries.
No stored spot-check names this claim in this edition.
Fireworks plans to release training support for DeepSeek-V4.1-Flash.
No stored spot-check names this claim in this edition.
Assertion 12
Assertion status: No spot-check verdict is published for this assertion.
In a Together AI benchmark on 113 DeepSWE tasks, GPT-5.6 Luna achieved a pass@1 rate of 67.2%, which was 14 points higher than DeepSeek-V4 Flash 0731's 53.3%.
No stored spot-check names this claim in this edition.
Assertion 13
Assertion status: No spot-check verdict is published for this assertion.
Dario Amodei said that the pace of AI model capability improvements must be slowed, but acknowledged that AI progress will still seem fast during this period.
No stored spot-check names this claim in this edition.
OpenAI CEO Sam Altman agreed with Dario Amodei's call to slow the frontier of AI development and stated OpenAI would follow Anthropic's commitment to embed third-party evaluators in their company.
No stored spot-check names this claim in this edition.
Amodei called for leading AI companies in democratic countries to coordinate common safety standards and limits on the rate of unchecked AI progress, ideally mediated by the US government with narrow antitrust waivers.
No stored spot-check names this claim in this edition.
Dario Amodei suggested that by controlling chip and semiconductor sales to Chinese companies and cracking down on model distillation, the US could slow China's AI progress enough to widen its lead significantly over the next 3 to 5 years.
No stored spot-check names this claim in this edition.
Huang's views on slowing AI improvement diverge from those of Anthropic CEO Dario Amodei, SpaceX CEO Elon Musk, and OpenAI CEO Sam Altman.
No stored spot-check names this claim in this edition.
Assertion 14
Assertion status: No spot-check verdict is published for this assertion.
Mango Excellent Media's TV adaptation of The Later Journey to the West premiered on August 31, marking the first fully AI-generated entertainment to reach mainstream provincial satellite prime time in China.
No stored spot-check names this claim in this edition.
The production of the AI-generated series took six months, a duration significantly shorter than the one to two years required for traditional TV dramas.
No stored spot-check names this claim in this edition.
ByteDance's Seedance 2.0 and 2.5 models were used to generate all visuals for the series, eliminating the need for live-action actors or filmed footage.
No stored spot-check names this claim in this edition.
Mango Excellent Media's market value surged by 13.3 billion RMB over five days following the premiere of the AI-generated series.
No stored spot-check names this claim in this edition.
In the first half of 2026, more than 74 percent of the 367,000 microdramas released on Chinese platforms were AI-generated.
No stored spot-check names this claim in this edition.
Assertion 15
Assertion status: No spot-check verdict is published for this assertion.
Fireworks released the DeepSeek-V4.1-Flash model.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark, DeepSeek-V4.1-Flash achieved a Pass@1 accuracy of 74.34%.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark, GPT-6-Astra achieved a Pass@1 accuracy of 74.12%.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark, Gemini 3.8 Flash achieved a Pass@1 accuracy of 73.83%.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark, Claude Opus 5 achieved a Pass@1 accuracy of 73.65%.
No stored spot-check names this claim in this edition.
DeepSeek-V4.1-Flash costs $0.430 per task in the DeepSWE benchmark.
No stored spot-check names this claim in this edition.
In the DeepSWE benchmark run, DeepSeek-V4.1-Flash consumed 36.9M input tokens and 211K output tokens per task.
No stored spot-check names this claim in this edition.
In the Terminal-Bench 2.1 benchmark, DeepSeek-V4.1-Flash achieved an accuracy of 86.5%.
No stored spot-check names this claim in this edition.
In the Terminal-Bench 2.1 benchmark, GPT-6-Astra achieved an accuracy of 87.5%.
No stored spot-check names this claim in this edition.
In the Humanity’s Last Exam (HLE) benchmark, DeepSeek-V4.1-Flash achieved a score of 34.52%.
No stored spot-check names this claim in this edition.
Fireworks plans to release US-hosted endpoints for DeepSeek-V4.1-Flash for regulated industries.
No stored spot-check names this claim in this edition.
Fireworks plans to release training support for DeepSeek-V4.1-Flash.
No stored spot-check names this claim in this edition.
Assertion 16
Assertion status: No spot-check verdict is published for this assertion.
Mango Excellent Media's TV adaptation of The Later Journey to the West premiered on August 31, marking the first fully AI-generated entertainment to reach mainstream provincial satellite prime time in China.
No stored spot-check names this claim in this edition.
The production of the AI-generated series took six months, a duration significantly shorter than the one to two years required for traditional TV dramas.
No stored spot-check names this claim in this edition.
ByteDance's Seedance 2.0 and 2.5 models were used to generate all visuals for the series, eliminating the need for live-action actors or filmed footage.
No stored spot-check names this claim in this edition.
Mango Excellent Media's market value surged by 13.3 billion RMB over five days following the premiere of the AI-generated series.
No stored spot-check names this claim in this edition.
In the first half of 2026, more than 74 percent of the 367,000 microdramas released on Chinese platforms were AI-generated.
No stored spot-check names this claim in this edition.
Move on to Augur's other sections.