RSS GitHub
The Ledger A sourced historical record of AI

25 summary revisions 0 open challenges 2 resolved challenges

  1. Summary revision DeepSeek Releases V4-Pro, Open-Weight Frontier Model Trained on Huawei Ascend Chips

    A review of the entry against its sources, raised in the follow-ups of pull request 234, found the error. The model card's comparison table, the source for the 0.2-point gap and the LiveCodeBench comparison, compares V4-Pro with Claude Opus 4.6 and not with Claude Opus 4.7, so the summary now names the model and states the figures the table gives.

  2. Challenge resolved Council of Europe AI Convention Entry into Force

    Disposition: sustained. The Treaty Office chart for CETS 225, status as of November 7, 2026, gives no entry-into-force date for any party, so the convention did not enter into force on November 1, 2025. The same chart lists two ratifications, by the European Union on May 15, 2026 and by Bosnia and Herzegovina on October 28, 2026, against the five, including at least three Council of Europe member states, that entry into force requires. The United Kingdom and Norway signed the convention on September 5, 2024 and had not ratified it; France had neither signed nor ratified it. The claimed UK, French and Norwegian ratifications in October 2025 did not happen. Fifteen of the 46 member states had signed by November 7, 2026, not 46 of 47 at the Vilnius opening as the original narrative states.

  3. Challenge resolved DeepSeek R1 Distillation Adoption

    Disposition: sustained. No listed or newly found source records a distinct DeepSeek-R1 distillation milestone on January 15, 2026, or dozens of community models appearing in early 2026. DeepSeek's R1 report, first posted January 22, 2025, released six smaller Qwen- and Llama-based models distilled from R1. Community distillation of R1 began within a week of that report: on January 28, 2025, Hugging Face announced Open-R1, whose first step was to replicate the R1-Distill models, and the Open Thoughts project launched an open reasoning dataset, citing R1 as a cheap source of reasoning demonstrations. The entry is kept as an attributed account of a trend that began in January 2025. No dated source in the record measures its size or places a peak in early 2026.

  4. Summary revision Council of Europe AI Convention Entry into Force

    The challenge was sustained. The Treaty Office chart for CETS 225, status as of November 7, 2026, gives no entry-into-force date and lists two ratifications, so the summary now states the convention's recorded status in place of the challenged claim.

  5. Summary revision DeepSeek R1 Distillation Adoption

    The challenge was sustained. The summary is re-scoped as an attributed account of the distillation trend, dated by its sources to January 2025, in place of the unsupported January 15, 2026 event.

  6. Summary revision Anthropic Announces Claude 3 Family and Releases Opus and Sonnet

    The March 4 family announcement explicitly separates available Opus/Sonnet models from forthcoming Haiku. The March 13 Haiku announcement establishes its later release. The original body's simultaneous-launch assertion is retained as the historical assertion, not the current account.

  7. Summary revision Anthropic Launches Claude Code to General Availability

    The May 22 primary announcement includes VS Code and JetBrains integrations, which its Claude Code section describes as new beta extensions. The same announcement released Claude Code on GitHub in beta, with background tasks through GitHub Actions. Terminal-based interaction did not mean that Claude Code made no attempt to integrate with IDEs. The original body's claim to that effect is preserved as the earlier assertion and is superseded by this correction.

  8. Summary revision Anthropic Publishes Circuit Tracing and 'Biology of a Large Language Model'

    The March 27 papers and May 29 tooling announcement are distinct publications. The tooling citation's original March 27 publication date is corrected to May 29. The original body's simultaneous-tool-release wording is preserved below as the earlier assertion.

  9. Summary revision Anthropic Rejects Expanded Military Use Terms

    New contemporaneous primary statements dated February 26 and 27 support a narrower, attributed account of the dispute. The event date changes from the unsupported February 28 to February 26, with February 27 as its end date. The original use-policy page and November 2024 report are historical context, not evidence for a new February 2026 pressure campaign. The original body's blanket military-restriction and Congressional-pressure narrative is not established by those citations. These statements record Anthropic's position, not independent adjudication of every allegation or the later legal status of the designation.

  10. Summary revision Anthropic Releases Updated Sonnet and Computer Use Beta, Announces Haiku

    The October 22 announcement distinguishes immediate Sonnet/computer-use beta access from a future Haiku release. Its planned Haiku date is recorded as a plan, not evidence that availability actually occurred that day.

  11. Summary revision Council of Europe AI Convention Entry into Force

    The source audit that produced the 22 approved editorial findings recorded this reading of the Treaty Office chart for CETS 225 on September 12, 2026: the chart "has no entry-into-force date and specifies five ratifications including three member states." That dated reading is the contradicting record for the claimed November 1, 2025 entry into force. The chart could not be checked again for this review: the text fetch required JavaScript and the browser was blocked by Cloudflare. Separately, the Treaty Office's indexed treaty details specify five ratifications, including at least three Council of Europe member states, as the entry-into-force condition. The official Framework Convention page confirms opening for signature on September 5, 2024, which is distinct from entry into force. This review does not infer the treaty's status after September 12, 2026 or an alternative entry date. A directly inspected, dated depositary record is required to resolve the claimed November 2025 event, country ratifications, and signature counts.

  12. Summary revision DeepSeek R1 Distillation Adoption

    The arXiv publication metadata is corrected from January 20 to January 22, 2025. The versioned R1 report describes DeepSeek's own six distilled Qwen/Llama-based models. It does not establish a January 15, 2026 event, dozens of community models, or the timing of the claimed adoption wave. A dated adoption dataset or contemporaneous reporting is needed before this can be retained as a dated historical event; alternatively an attributed retrospective would need separate editorial review. The date is not replaced with the technique's 2025 release, which is already separately recorded.

  13. Summary revision DeepSeek Releases DeepSeek-V3, High-Performance Open-Weight Model

    DeepSeek's preserved release commit is dated December 26, 2024 at 11:01:57 UTC; the original December 25 date has no documented timezone basis in this record. The technical report's Table 1 distinguishes $5.328 million for pre-training from $5.576 million for the full official training run, excluding prior research and ablations. The release README distinguishes MIT-licensed code from separately licensed weights; the original body's MIT-derivative/fully-open characterization should not be relied on.

  14. Summary revision Gemini Deep Think Achieves IMO 2025 Gold in Natural Language

    The cited arXiv identifier was assigned an incorrect title and attribution. Version 1, submitted July 21, is titled Gemini 2.5 Pro Capable of Winning Gold at IMO 2025 and names Yichen Huang and Lin F. Yang. Later versions changed the title and expanded the model comparison; they do not turn it into DeepMind's official report. The misattributed citation is removed from the event's supporting source list and preserved here as contextual evidence of the correction. The existing official DeepMind announcement remains the source for the competition result.

  15. Summary revision Google DeepMind Launches Gemini

    The December 6 primary announcement explicitly dates developer API access to December 13. The original body combined Bard and API availability; the corrected summary distinguishes the product rollout from the announced API schedule.

  16. Summary revision Google Launches Gemini 2.5 Pro Experimental with Native Thinking

    The same March 25 primary announcement identifies Gemini 2.0 Flash Thinking as Google's earlier thinking model. The first-Gemini implication is removed; being the first 2.5 release is distinct from originating thinking across the Gemini family.

  17. Summary revision Google Makes Gemini 2.0 Flash Generally Available in the API

    The original citation documents the December 11, 2024 introduction rather than the February 5 rollout. The February 5 primary announcement is added and the title/summary specify API general availability, distinguishing the earlier experimental and app releases.

  18. Summary revision Google Releases Gemma 3 Open-Weight Multimodal Family

    The model card assigns 128K context to 4B/12B/27B and 32K to 1B, while Google's developer explanation identifies 1B as text-only. The original summary and body incorrectly generalized vision and 128K context to every launch variant. Later additions to the Gemma family are not backdated to this four-size launch.

  19. Summary revision Mistral AI Releases Mistral 7B via Torrent

    The primary announcement explicitly identifies Llama 1 34B and explains why it did not use Llama 2 34B. The original body's comparison with an unreleased Llama 2 34B variant is corrected in the canonical summary.

  20. Summary revision Mistral AI Releases Mistral Large

    The model announcement supports February 26, 2024, but the financing participant's announcement places the Series A in December 2023. The event title and summary no longer combine the model launch and funding close as simultaneous news.

  21. Summary revision Mistral AI Releases Small 3 — 24B Apache-2.0 Open-Weight Model

    Apache 2.0 section 4 requires a copy of the license for recipients, notices of modified files, retention of relevant source notices, and applicable NOTICE attribution in redistributed derivative works. The original body's unconditional no-attribution wording is withdrawn. This describes the license text rather than promising that any particular distribution satisfies it.

  22. Summary revision NVIDIA Announces H20 License Assurances; Trump Later Discloses Revenue Share

    The original July 15 date and combined title backdated an August disclosure. NVIDIA's primary announcement is dated July 14 and concerns license assurances and applications. Trump's August 11 press-conference transcript at approximately 01:06-01:07 discusses his initial 20% request and negotiated 15% share. The American Presidency Project preserves a corresponding transcript in its Semiconductor Manufacturing section. NPR's August 11 report was inspected through its VPM syndication because the NPR page denied automated access. These are separate milestones, not evidence that the 15% condition was announced in July.

  23. Summary revision OpenAI Launches GPT-4o

    The primary announcement separates demonstrated capabilities from launch-day access. It places the new Voice Mode alpha and limited-partner audio/video API rollout in the coming weeks. The original broad availability language must not be read as general availability of the demonstrated new voice mode on May 13.

  24. Summary revision OpenAI Releases GPT-5 as Unified Flagship Model

    The primary announcement is dated August 7, 2025; April 14, 2025 was the GPT-4.1 API launch. The original April date is corrected in the event metadata. The original body below retains the mistaken date and single-model description; the announcement describes a routed system, rather than a completed merger into one model.

  25. Summary revision OpenAI Releases GPT-5.2

    The announcement establishes December 11, 2025, rather than February 1, 2026. The event is relocated to the 2025 directory while retaining its ID and public event URL. The announcement explicitly distinguishes Instant, Thinking, and Pro; the original summary's assertion that there were no separate thinking modes is withdrawn. The old source dates and original metadata are preserved in this revision record; the mutable platform guide is not assigned an invented publication date.

  26. Summary revision OpenAI Releases GPT-5.5, First Fully Retrained Base Model Since GPT-4.5

    The million-token first claim is contradicted by GPT-4.1's primary announcement. The source for GPT-5.3-Codex establishes February 5, not March 2026. GPT-5.5's own announcement distinguishes 1M API context from 400K in Codex and adds API availability in its April 24 update. The original summary is preserved in this revision record. Other architectural and benchmark assertions in the original body are retained as historical assertions and require a separate source review.

  27. Summary revision OpenAI Releases o3-mini Reasoning Model

    The January 31 primary post announces o3-mini, not the simultaneous release of o3 and o3-mini. The title is narrowed while the historical event ID is retained to preserve existing links and thread references.

  28. Challenge opened Council of Europe AI Convention Entry into Force

    Alleges timeline error, unsupported inference.

  29. Challenge opened DeepSeek R1 Distillation Adoption

    Alleges timeline error, unsupported inference.