{
  "schema_version": "1.0",
  "as_of": "2026-07-19",
  "entries": [
    {
      "id": "training-pytorch-fsdp2-fully-shard",
      "category": "training-runtime",
      "claim": "PyTorch's current FSDP2 fully_shard API uses per-parameter DTensors and documents all-gather, reshard, and reduce-scatter behavior around module execution.",
      "class": "vendor-reported",
      "pin": "PyTorch fully_shard documentation checked 2026-07-19",
      "checked": "2026-07-19",
      "source": "https://docs.pytorch.org/docs/main/distributed.fsdp.fully_shard.html",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "migration_note": "Pin the PyTorch release and repeat layout, checkpoint, numerical-equivalence, and resharding tests before changing FSDP APIs or policies."
    },
    {
      "id": "training-nvidia-nvfp4-2025",
      "category": "training-precision",
      "claim": "NVIDIA's 2025 NVFP4 paper reports a Blackwell-oriented FP4 training recipe using two-level scaling, selective high-precision operations, two-dimensional block quantization, and stochastic rounding.",
      "class": "measured",
      "pin": "NVFP4 Training paper, arXiv 2509.25149",
      "checked": "2026-07-19",
      "source": "https://arxiv.org/abs/2509.25149",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "migration_note": "Treat the format as a dated recipe; validate hardware support, scale granularity, accumulator precision, saturation telemetry, and quality against a higher-precision control."
    },
    {
      "id": "training-kimi-k2-muonclip",
      "category": "training-optimizer",
      "claim": "The Kimi K2 report describes MuonClip, which augments Muon training with attention-logit control through query/key weight rescaling.",
      "class": "measured",
      "pin": "Kimi K2 technical report, arXiv 2507.20534",
      "checked": "2026-07-19",
      "source": "https://arxiv.org/abs/2507.20534",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "migration_note": "Reproduce the logit threshold, eligible parameter families, optimizer split, precision, and scale rules; compare equal-token quality and stability against tuned controls."
    },
    {
      "id": "training-deepep-2026",
      "category": "training-runtime",
      "claim": "DeepEP is DeepSeek's current open-source communication library for mixture-of-experts dispatch and combine collectives, including high-throughput and low-latency modes.",
      "class": "vendor-reported",
      "pin": "DeepEP repository checked 2026-07-19",
      "checked": "2026-07-19",
      "source": "https://github.com/deepseek-ai/DeepEP",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "migration_note": "Pin the commit and benchmark the exact token shapes, topology, routing skew, overlap, precision, and failure behavior before adopting its kernels or hooks."
    },
    {
      "id": "training-torchtitan-2026",
      "category": "training-stack",
      "claim": "TorchTitan is PyTorch's current reference platform for large-model training and composes PyTorch-native distributed training features.",
      "class": "vendor-reported",
      "pin": "TorchTitan repository checked 2026-07-19",
      "checked": "2026-07-19",
      "source": "https://github.com/pytorch/torchtitan",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "migration_note": "Reference stacks change quickly; pin code and configuration, then revalidate memory, convergence, throughput, checkpoint portability, and fault recovery."
    },
    {
      "id": "training-maxtext-2026",
      "category": "training-stack",
      "claim": "MaxText is Google's current JAX-based reference implementation for scalable large-language-model training across supported accelerator configurations.",
      "class": "vendor-reported",
      "pin": "MaxText repository checked 2026-07-19",
      "checked": "2026-07-19",
      "source": "https://github.com/AI-Hypercomputer/maxtext",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "migration_note": "Pin JAX, XLA, MaxText, topology, and model configuration; repeat numerical, sharding, checkpoint, throughput, and recovery validation on the target system."
    },
    {
      "id": "training-llama3-interruptions",
      "category": "training-operations",
      "claim": "The Llama 3 report's 405B training snapshot records 466 interruptions over 54 days: 47 planned, 419 unexpected, and about 78 percent of unexpected interruptions confirmed or suspected hardware-related.",
      "class": "measured",
      "pin": "Llama 3 technical report, arXiv 2407.21783",
      "checked": "2026-07-19",
      "source": "https://arxiv.org/abs/2407.21783",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "migration_note": "Use these values only as a dated existence proof; estimate interruption, detection, recovery, and replay distributions from the target fleet."
    },
    {
      "id": "training-megascale-2024",
      "category": "training-operations",
      "claim": "MegaScale reports 55.2 percent model FLOPs utilization for a 175B model on 12,288 GPUs and reports production runs with more than 100 restarts, over 90 percent automatically handled.",
      "class": "measured",
      "pin": "MegaScale paper, arXiv 2402.15627",
      "checked": "2026-07-19",
      "source": "https://arxiv.org/abs/2402.15627",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "migration_note": "Do not transfer utilization or recovery rates across models and clusters; publish the FLOP convention and measure the target topology and operating process."
    },
    {
      "id": "training-decoupled-diloco-2026",
      "category": "decentralized-training",
      "claim": "The 2026 Decoupled DiLoCo paper studies decentralized training with asynchronous model-fragment exchange and quorum-style merging.",
      "class": "measured",
      "pin": "Decoupled DiLoCo, arXiv 2604.21428",
      "checked": "2026-07-19",
      "source": "https://arxiv.org/abs/2604.21428",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "migration_note": "Treat convergence, staleness, quorum, fault, and bandwidth assumptions as algorithmic inputs; reproduce them before using wide-area training."
    },
    {
      "id": "pricing-gcp-a3-highgpu-8g-2026",
      "category": "training-pricing",
      "claim": "Google Cloud's accelerator-optimized pricing page listed the on-demand a3-highgpu-8g H100 VM at 88.490000119 U.S. dollars per hour, or 11.061250014875 dollars per GPU-hour, when checked.",
      "class": "vendor-reported",
      "pin": "Google Cloud A3 pricing checked 2026-07-19",
      "checked": "2026-07-19",
      "source": "https://cloud.google.com/products/compute/pricing/accelerator-optimized",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-pricing-gcp-a3-highgpu-8g-2026",
      "migration_note": "Pricing is region-, contract-, and date-sensitive; obtain an approved quote and include storage, network, support, idle capacity, and interruption waste."
    },
    {
      "id": "hardware-google-ironwood-2026",
      "category": "hardware",
      "claim": "Google describes Ironwood as its seventh-generation TPU and a current Google Cloud AI infrastructure offering.",
      "class": "vendor-reported",
      "pin": "Google Ironwood overview checked 2026-07-19",
      "checked": "2026-07-19",
      "source": "https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/ironwood-google-tpu-things-to-know/",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "migration_note": "Recheck availability and specifications, then benchmark the target model, precision, topology, collective mix, compiler, and power boundary."
    },
    {
      "id": "hardware-nvidia-vera-rubin-announcement",
      "category": "hardware",
      "claim": "NVIDIA has announced the Vera Rubin platform; its availability and delivered workload performance remain dated procurement inputs rather than durable mechanisms.",
      "class": "vendor-reported",
      "pin": "NVIDIA Vera Rubin announcement checked 2026-07-19",
      "checked": "2026-07-19",
      "source": "https://nvidianews.nvidia.com/news/nvidia-vera-rubin-platform",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "migration_note": "Track delivered systems rather than announcement ceilings and obtain workload-shaped benchmarks, power data, topology guarantees, schedule, and quotes."
    },
    {
      "id": "legal-us-1e26-threshold-status-2026",
      "category": "legal-watch",
      "claim": "Executive Order 14110's historical 10^26-operation reporting threshold is not an active general U.S. reporting requirement after that order's January 2025 revocation.",
      "class": "vendor-reported",
      "pin": "Congressional Research Service status summary checked 2026-07-19",
      "checked": "2026-07-19",
      "source": "https://www.congress.gov/crs_external_products/R/HTML/R48555.web.html",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "migration_note": "This is a dated legal-status correction, not legal advice; qualified counsel must recheck federal, state, export-control, contract, and sector obligations."
    },
    {
      "id": "model-deepseek-v4-pro",
      "category": "model-profile",
      "claim": "The DeepSeek-V4-Pro model card reports a 1.6T-parameter mixture-of-experts model with 49B parameters activated per token.",
      "class": "vendor-reported",
      "pin": "Hugging Face revision b5968e9190ef611bbf34a7229255be88a0e937c1",
      "checked": "2026-07-19",
      "source": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-model-deepseek-v4-pro",
      "migration_note": "Re-read the configuration and serving notes before changing capacity examples; total parameters do not determine per-token compute or memory by themselves."
    },
    {
      "id": "model-kimi-k2-5",
      "category": "model-profile",
      "claim": "The Kimi K2.5 model card reports 1T total parameters and 32B activated parameters.",
      "class": "vendor-reported",
      "pin": "Hugging Face revision 4d01dfe0332d63057c186e0b262165819efb6611",
      "checked": "2026-07-19",
      "source": "https://huggingface.co/moonshotai/Kimi-K2.5",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-model-kimi-k2-5",
      "migration_note": "Treat this as an architecture profile, not a recommendation; re-run workload-specific quality, latency, and tool-use evaluations for every checkpoint change."
    },
    {
      "id": "model-qwen3-5-397b-a17b",
      "category": "model-profile",
      "claim": "The Qwen3.5-397B-A17B model card reports 397B total parameters and 17B activated parameters.",
      "class": "vendor-reported",
      "pin": "Hugging Face revision 8472618112abcbd45acbcdc58436aff4233c23f7",
      "checked": "2026-07-19",
      "source": "https://huggingface.co/Qwen/Qwen3.5-397B-A17B",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-model-qwen3-5-397b-a17b",
      "migration_note": "When the card or runtime support changes, update the architecture profile and repeat memory, throughput, and structured-output compatibility tests."
    },
    {
      "id": "model-glm-5",
      "category": "model-profile",
      "claim": "The GLM-5 model card reports 744B total parameters and 40B activated parameters.",
      "class": "vendor-reported",
      "pin": "Hugging Face revision 4e6698ba8e85059d749020e3c4d2123719f23926",
      "checked": "2026-07-19",
      "source": "https://huggingface.co/zai-org/GLM-5",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-model-glm-5",
      "migration_note": "Recheck the model card, license, context policy, tokenizer, and supported inference engines before operational use."
    },
    {
      "id": "model-gpt-oss-120b",
      "category": "model-profile",
      "claim": "The gpt-oss-120b model card reports 117B total parameters and approximately 5.1B activated parameters.",
      "class": "vendor-reported",
      "pin": "Hugging Face revision b5c939de8f754692c1647ca79fbf85e8c1e70f8a",
      "checked": "2026-07-19",
      "source": "https://huggingface.co/openai/gpt-oss-120b",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-model-gpt-oss-120b",
      "migration_note": "Keep model-card facts separate from measured task quality; repin the exact weights, tokenizer, chat template, and inference engine used in an evaluation."
    },
    {
      "id": "model-llama-4-maverick",
      "category": "model-profile",
      "claim": "Meta's Llama 4 announcement reports that Maverick uses 400B total parameters with 17B active parameters.",
      "class": "vendor-reported",
      "pin": "Meta Llama 4 announcement checked 2026-07-19",
      "checked": "2026-07-19",
      "source": "https://ai.meta.com/blog/llama-4-multimodal-intelligence/",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-model-llama-4-maverick",
      "migration_note": "Recheck license, model card, and deployable artifact details; announcement-level claims are not a substitute for an exact checkpoint manifest."
    },
    {
      "id": "runtime-vllm-0-21-0",
      "category": "inference-runtime",
      "claim": "vLLM 0.21.0 is a 2026 release whose notes deprecate Transformers v4 support and require a C++20-compatible compiler.",
      "class": "vendor-reported",
      "pin": "v0.21.0",
      "checked": "2026-07-19",
      "source": "https://github.com/vllm-project/vllm/releases/tag/v0.21.0",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-runtime-vllm-0-21-0",
      "migration_note": "Pin the container digest and engine configuration; rehearse upgrades against golden outputs, memory ceilings, tool parsers, and rollback procedures."
    },
    {
      "id": "runtime-llama-cpp-b9637",
      "category": "local-inference",
      "claim": "llama.cpp build b9637 published binaries for CPU, CUDA, ROCm, Vulkan, SYCL, OpenVINO, macOS, iOS, Android, and other targets.",
      "class": "vendor-reported",
      "pin": "b9637 (aedb2a5)",
      "checked": "2026-07-19",
      "source": "https://github.com/ggml-org/llama.cpp/releases/tag/b9637",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-runtime-llama-cpp-b9637",
      "migration_note": "llama.cpp uses frequent build tags; pin the commit and quantized artifact hash, then repeat perplexity, output-compatibility, and throughput checks."
    },
    {
      "id": "hardware-amd-mi350x",
      "category": "hardware",
      "claim": "AMD specifies 288 GB of HBM3E and up to 8 TB/s peak memory bandwidth for one Instinct MI350X accelerator.",
      "class": "vendor-reported",
      "pin": "MI350X product specification",
      "checked": "2026-07-19",
      "source": "https://www.amd.com/en/products/accelerators/instinct/mi350/mi350x.html",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-hardware-amd-mi350x",
      "migration_note": "Use the vendor figures as ceilings, not workload throughput; benchmark the exact server, interconnect, runtime, precision, batch mix, and power envelope."
    },
    {
      "id": "hardware-nvidia-gb300-nvl72",
      "category": "hardware",
      "claim": "NVIDIA specifies 72 Blackwell Ultra GPUs, 20 TB aggregate GPU memory, and up to 576 TB/s aggregate GPU-memory bandwidth for GB300 NVL72.",
      "class": "vendor-reported",
      "pin": "GB300 NVL72 product specification",
      "checked": "2026-07-19",
      "source": "https://www.nvidia.com/en-us/data-center/gb300-nvl72/",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-hardware-nvidia-gb300-nvl72",
      "migration_note": "Do not compare rack aggregates with per-GPU figures; replace marketing ceilings with measured end-to-end throughput, tail latency, utilization, and energy per successful task."
    },
    {
      "id": "hardware-apple-mac-studio-2025",
      "category": "local-hardware",
      "claim": "Apple's 2025 Mac Studio specification allows up to 256 GB unified memory with M3 Ultra and reports 819 GB/s memory bandwidth.",
      "class": "vendor-reported",
      "pin": "Mac Studio (2025) technical specification",
      "checked": "2026-07-19",
      "source": "https://support.apple.com/en-us/122211",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-hardware-apple-mac-studio-2025",
      "migration_note": "Confirm the configured memory capacity and measure the exact quantization, prompt length, thermal behavior, tokens per second, and time to first token."
    },
    {
      "id": "tokenizer-free-blt",
      "category": "tokenizer-free",
      "claim": "Byte Latent Transformer groups raw bytes into dynamically sized entropy-based patches and reports controlled studies up to 8B parameters and 4T training bytes.",
      "class": "measured",
      "pin": "arXiv:2412.09871",
      "checked": "2026-07-19",
      "source": "https://arxiv.org/abs/2412.09871",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-tokenizer-free-blt",
      "migration_note": "Keep this in the frontier lane until independent large-scale replications cover multilingual quality, serving efficiency, tooling, and full lifecycle costs."
    },
    {
      "id": "tokenizer-free-hnet",
      "category": "tokenizer-free",
      "claim": "H-Net learns content- and context-dependent byte chunking jointly with a hierarchical sequence model rather than using a fixed external tokenizer.",
      "class": "measured",
      "pin": "arXiv:2507.07955; ICLR 2026",
      "checked": "2026-07-19",
      "source": "https://arxiv.org/abs/2507.07955",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-tokenizer-free-hnet",
      "migration_note": "Treat reported compute-matched gains as results under the paper's settings; compare bytes processed, FLOPs, latency, robustness, and downstream quality on the target distribution."
    },
    {
      "id": "tokenizer-free-evabyte",
      "category": "tokenizer-free",
      "claim": "EvaByte reports a 6.5B-parameter byte-level model trained on 1.5T bytes with an architecture designed to reduce byte-level decoding cost.",
      "class": "measured",
      "pin": "EvaByte 6.5B project release",
      "checked": "2026-07-19",
      "source": "https://hkunlp.github.io/blog/2025/evabyte/",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-tokenizer-free-evabyte",
      "migration_note": "Verify checkpoint, code, hardware, and measurement protocol before comparing decoding speed with tokenized models; bytes and tokens are not interchangeable units."
    },
    {
      "id": "framework-langgraph-v1",
      "category": "agent-framework",
      "claim": "LangGraph v1 is documented as a stability-focused release with checkpointing, persistence, streaming, and human-in-the-loop as runtime capabilities.",
      "class": "vendor-reported",
      "pin": "LangGraph 1.x LTS lane",
      "checked": "2026-07-19",
      "source": "https://docs.langchain.com/oss/python/releases/langgraph-v1",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-framework-langgraph-v1",
      "migration_note": "Keep domain state and tool contracts independent of framework classes; use persistence-format migrations and replay tests before upgrading."
    },
    {
      "id": "framework-autogen-0-4",
      "category": "agent-framework",
      "claim": "AutoGen's stable documentation separates AgentChat, an event-driven Core runtime, extensions, and Studio, and identifies 0.4 as the successor to 0.2.",
      "class": "vendor-reported",
      "pin": "AutoGen 0.4 stable documentation lane",
      "checked": "2026-07-19",
      "source": "https://microsoft.github.io/autogen/stable/index.html",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-framework-autogen-0-4",
      "migration_note": "Pin individual Python packages and serialized message schemas; exercise distributed-runtime, cancellation, and recovery behavior before upgrades."
    },
    {
      "id": "protocol-mcp-2025-11-25",
      "category": "protocol",
      "claim": "The current Model Context Protocol revision is 2025-11-25; MCP uses date-based versions and negotiates one version during initialization.",
      "class": "vendor-reported",
      "pin": "2025-11-25",
      "checked": "2026-07-19",
      "source": "https://modelcontextprotocol.io/specification/2025-11-25",
      "owner_chapter": "chapters/19-protocols-frameworks.qmd",
      "migration_note": "Advertise supported revisions, run capability-negotiation contract tests, and follow feature deprecation notices instead of assuming every server supports every feature."
    },
    {
      "id": "protocol-a2a-1-0-0",
      "category": "protocol",
      "claim": "The A2A project released version 1.0.0 on 12 March 2026; the 1.0 protocol line uses explicit version negotiation and introduces breaking changes from earlier lines.",
      "class": "vendor-reported",
      "pin": "A2A tag v1.0.0 at commit 173695755607e884aa9acf8ce4feed90e32727a1",
      "checked": "2026-07-19",
      "source": "https://a2a-protocol.org/latest/specification/",
      "owner_chapter": "chapters/19-protocols-frameworks.qmd",
      "migration_note": "Send and validate A2A-Version, preserve task and artifact state across retries, and use compatibility fixtures when migrating from 0.3 object discriminators."
    },
    {
      "id": "protocol-agui-pre-1",
      "category": "protocol",
      "claim": "AG-UI is an event-based agent-to-user protocol; its documentation marks THINKING events deprecated in favor of REASONING events before 1.0.0.",
      "class": "vendor-reported",
      "pin": "pre-1.0 event schema checked 2026-07-19",
      "checked": "2026-07-19",
      "source": "https://docs.ag-ui.com/concepts/events",
      "owner_chapter": "chapters/19-protocols-frameworks.qmd",
      "migration_note": "Use tolerant readers and explicit event-schema versioning; dual-read deprecated THINKING events only for a bounded compatibility window."
    },
    {
      "id": "protocol-otel-semconv-1-43",
      "category": "observability-standard",
      "claim": "The dedicated OpenTelemetry GenAI semantic-conventions repository extends the core semantic conventions and, at the pinned revision, declares SEMCONV_VERSION v1.43.0; its GenAI surface remains a versioned mapping that applications must review for sensitive content and cardinality.",
      "class": "vendor-reported",
      "pin": "semantic-conventions-genai commit c26a2c21d1ee70d5231bd440c7b48d3c94ee506a; SEMCONV_VERSION v1.43.0",
      "checked": "2026-07-19",
      "source": "https://github.com/open-telemetry/semantic-conventions-genai",
      "owner_chapter": "chapters/27-operating-agents.qmd",
      "migration_note": "Pin emitted attribute names, redact content before export, constrain cardinality, and dual-emit only through an explicitly measured migration window."
    },
    {
      "id": "protocol-c2pa-2-4",
      "category": "provenance-standard",
      "claim": "C2PA 2.4, published in April 2026, specifies signed provenance manifests for media and related assets.",
      "class": "vendor-reported",
      "pin": "C2PA 2.4",
      "checked": "2026-07-19",
      "source": "https://spec.c2pa.org/specifications/specifications/2.4/specs/C2PA_Specification.html",
      "owner_chapter": "chapters/30-multimodal-voice-video-media.qmd",
      "migration_note": "Validate manifests and trust chains with pinned conformance tools; provenance metadata supports authenticity reasoning but does not prove that content is true."
    },
    {
      "id": "benchmark-terminal-bench-2-1",
      "category": "benchmark",
      "claim": "Terminal-Bench 2.1 is a May 2026 revision that fixes 28 tasks and adds continuous validation; Harbor is the official harness for the 2.x line.",
      "class": "measured",
      "pin": "terminal-bench 2.1 plus pinned Harbor",
      "checked": "2026-07-19",
      "source": "https://www.tbench.ai/news",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-benchmark-terminal-bench-2-1",
      "migration_note": "Record dataset, task, container, harness, agent, model, and budget pins; never compare a 2.0 score directly with a 2.1 score without a task-level crosswalk."
    },
    {
      "id": "benchmark-osworld-2",
      "category": "benchmark",
      "claim": "OSWorld 2.0 version v2026.06.24 contains 108 long-horizon computer-use workflows and reports results under a dated harness and environment.",
      "class": "measured",
      "pin": "OSWorld 2.0 v2026.06.24",
      "checked": "2026-07-19",
      "source": "https://osworld-v2.xlang.ai/",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-benchmark-osworld-2",
      "migration_note": "Pin VM images, application versions, task revision, policy, observation resolution, and retry budget; report verified end state rather than click completion."
    },
    {
      "id": "benchmark-longmemeval-v2",
      "category": "benchmark",
      "claim": "LongMemEval-V2 contains 451 manually curated questions across five memory abilities for agents operating in customized web environments.",
      "class": "measured",
      "pin": "arXiv:2605.12493 and repository snapshot",
      "checked": "2026-07-19",
      "source": "https://arxiv.org/abs/2605.12493",
      "owner_chapter": "chapters/18-memory-experiential-learning.qmd",
      "migration_note": "Keep evaluation data isolated from prompts and memory stores; audit harness rules for leakage and complement question answering with action-level memory tests."
    },
    {
      "id": "benchmark-tau2",
      "category": "benchmark",
      "claim": "tau2-bench evaluates dual-control conversations in which both the agent and simulated user act through tools in a shared environment.",
      "class": "measured",
      "pin": "arXiv:2506.07982 and pinned simulator",
      "checked": "2026-07-19",
      "source": "https://arxiv.org/abs/2506.07982",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-benchmark-tau2",
      "migration_note": "Pin the user simulator, policy text, task generator, tool environment, speech settings if used, and repeated-run budget; separate coordination failures from policy failures."
    },
    {
      "id": "benchmark-swe-verified-retirement",
      "category": "benchmark-retirement",
      "claim": "OpenAI stated in March 2026 that SWE-bench Verified no longer measures frontier coding capability reliably because contamination and flawed tests undermine interpretation.",
      "class": "vendor-reported",
      "pin": "OpenAI retirement rationale 2026-03",
      "checked": "2026-07-19",
      "source": "https://openai.com/index/why-we-no-longer-evaluate-swe-bench-verified/",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-benchmark-swe-verified-retirement",
      "migration_note": "Retain only for historical continuity; route launch decisions to fresh, contamination-aware, task-audited internal and live benchmark suites.",
      "status": "retired"
    },
    {
      "id": "benchmark-swe-live",
      "category": "benchmark",
      "claim": "SWE-bench-Live is designed as a live-updatable software-engineering benchmark to reduce static-test contamination and staleness.",
      "class": "measured",
      "pin": "arXiv:2505.23419 plus evaluation snapshot",
      "checked": "2026-07-19",
      "source": "https://arxiv.org/abs/2505.23419",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-benchmark-swe-live",
      "migration_note": "Report the task cutoff, repository state, harness, tool policy, compute budget, and adjudication procedure; live does not automatically mean unbiased or error-free."
    },
    {
      "id": "benchmark-screenspot-pro",
      "category": "benchmark",
      "claim": "ScreenSpot-Pro evaluates GUI grounding on high-resolution professional applications spanning 23 applications, five industries, and three operating systems.",
      "class": "measured",
      "pin": "arXiv:2504.07981 and leaderboard snapshot",
      "checked": "2026-07-19",
      "source": "https://arxiv.org/abs/2504.07981",
      "owner_chapter": "chapters/29-multimodal-vlm-documents-gui.qmd",
      "migration_note": "Pin screen resolution and grounding protocol; do not treat element grounding as evidence of safe task completion or recovery under UI changes."
    },
    {
      "id": "benchmark-omnidocbench-1-5",
      "category": "benchmark",
      "claim": "OmniDocBench evaluates document parsing across nine document sources with layout, attribute, module-level, and end-to-end annotations and metrics.",
      "class": "measured",
      "pin": "OmniDocBench arXiv:2412.07626v2 (CVPR 2025 paper)",
      "checked": "2026-07-19",
      "source": "https://arxiv.org/abs/2412.07626",
      "owner_chapter": "chapters/29-multimodal-vlm-documents-gui.qmd",
      "migration_note": "Report the paper and dataset revision plus per-element metrics; audit source traceability, annotations, contamination, and downstream semantic correctness before using it as a launch gate."
    },
    {
      "id": "benchmark-real5-omnidocbench",
      "category": "benchmark",
      "claim": "Real5-OmniDocBench reports a one-to-one physical reconstruction of all 1,355 OmniDocBench v1.5 images under scanning, warping, screen-photography, illumination, and skew conditions.",
      "class": "measured",
      "pin": "Real5-OmniDocBench arXiv:2603.04205v2",
      "checked": "2026-07-19",
      "source": "https://arxiv.org/abs/2603.04205",
      "owner_chapter": "chapters/29-multimodal-vlm-documents-gui.qmd",
      "migration_note": "Keep the digital-to-physical pairing and five capture conditions explicit; rerun condition-level and downstream semantic checks when the benchmark, capture pipeline, or evaluated parser changes."
    },
    {
      "id": "retrieval-graphrag-query-modes",
      "category": "retrieval-policy",
      "claim": "Microsoft's GraphRAG documentation exposes local, global, DRIFT, and basic query modes over its indexed corpus structures.",
      "class": "vendor-reported",
      "pin": "GraphRAG query documentation checked 2026-07-19",
      "checked": "2026-07-19",
      "source": "https://microsoft.github.io/graphrag/query/overview/",
      "owner_chapter": "chapters/15-agentic-retrieval.qmd",
      "migration_note": "Pin the package and index schema, then rerun route-specific quality, indexing-cost, provenance, migration, and deletion tests before adopting a changed query mode."
    },
    {
      "id": "retrieval-search-r1",
      "category": "retrieval-policy",
      "claim": "Search-R1 reports reinforcement-learning experiments in which a language model interleaves reasoning with multiple real-time search calls.",
      "class": "measured",
      "pin": "COLM 2025; OpenReview Rwhi91ideu",
      "checked": "2026-07-19",
      "source": "https://openreview.net/forum?id=Rwhi91ideu",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-retrieval-search-r1",
      "migration_note": "Treat the reported trajectories as task- and reward-specific evidence; repeat answer, citation, call-count, latency, and security evaluation on the target retriever and corpus."
    },
    {
      "id": "retrieval-research",
      "category": "retrieval-policy",
      "claim": "ReSearch reports training language models to interleave reasoning and search with reinforcement learning, without supervised data for reasoning steps.",
      "class": "measured",
      "pin": "arXiv:2503.19470v3 (2025-09-23); accepted to NeurIPS 2025",
      "checked": "2026-07-19",
      "source": "https://arxiv.org/abs/2503.19470",
      "owner_chapter": "chapters/15-agentic-retrieval.qmd",
      "migration_note": "Keep the paper's training distribution, reward, retriever, model sizes, and benchmark settings attached to its results; rerun a cost-matched evaluation before adopting the learned search policy."
    },
    {
      "id": "retrieval-sparkle-acl-2026",
      "category": "retrieval-policy",
      "claim": "SPARKLE presents a structured plug-and-play agentic retrieval policy for adaptive RAG in the ACL 2026 proceedings.",
      "class": "measured",
      "pin": "ACL 2026 long paper 1793",
      "checked": "2026-07-19",
      "source": "https://aclanthology.org/2026.acl-long.1793/",
      "owner_chapter": "chapters/15-agentic-retrieval.qmd",
      "migration_note": "Repin released code and evaluation data when available; compare the policy with direct, one-shot, and bounded iterative baselines under one cost and abstention contract."
    },
    {
      "id": "retrieval-poisonedrag",
      "category": "retrieval-security",
      "claim": "PoisonedRAG studies knowledge-corruption attacks that inject malicious texts into a retrieval corpus to influence selected answers.",
      "class": "measured",
      "pin": "34th USENIX Security Symposium (USENIX Security 25), pages 3827-3844",
      "checked": "2026-07-19",
      "source": "https://www.usenix.org/conference/usenixsecurity25/presentation/zou-poisonedrag",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-retrieval-poisonedrag",
      "migration_note": "Do not generalize one attack setup into a universal rate; keep corpus admission, provenance, retrieval exposure, model influence, and downstream effect tests separate."
    },
    {
      "id": "security-nist-ai-100-2e2025",
      "category": "security-standard",
      "claim": "NIST AI 100-2e2025 includes indirect prompt injection in its adversarial-machine-learning taxonomy and terminology.",
      "class": "vendor-reported",
      "pin": "NIST AI 100-2e2025",
      "checked": "2026-07-19",
      "source": "https://doi.org/10.6028/NIST.AI.100-2e2025",
      "owner_chapter": "chapters/24-agent-security.qmd",
      "migration_note": "Recheck terminology and mapped controls when NIST revises the publication; preserve protocol-neutral least-privilege, provenance, isolation, and effect-gating tests."
    },
    {
      "id": "security-owasp-llmvs-2",
      "category": "security-standard",
      "claim": "OWASP's LLM Verification Standard v2 identifies retrieved documents, tool outputs, messages, and web content as channels for indirect prompt injection.",
      "class": "vendor-reported",
      "pin": "OWASP LLMSVS v2 page checked 2026-07-19",
      "checked": "2026-07-19",
      "source": "https://owasp.org/www-project-llm-verification-standard/LLMSVS-v2.0-en.html",
      "owner_chapter": "chapters/15-agentic-retrieval.qmd",
      "migration_note": "Track requirement identifiers and maturity changes, but retain independent authorization and output/effect controls rather than relying on an instruction detector."
    },
    {
      "id": "governance-eu-ai-act-timeline",
      "category": "legal-watch",
      "claim": "The European Commission's July 2026 page says the AI Act entered into force on 1 August 2024, GPAI obligations applied from 2 August 2025, and most rules apply from 2 August 2026, with stated exceptions and later high-risk dates.",
      "class": "vendor-reported",
      "pin": "European Commission implementation page checked 2026-07-19",
      "checked": "2026-07-19",
      "source": "https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai",
      "owner_chapter": "chapters/25-alignment-interpretability-governance.qmd",
      "migration_note": "Counsel must confirm the enacted text, jurisdiction, role, risk classification, later amendments, standards, and exact application dates before relying on this summary."
    },
    {
      "id": "governance-california-ab-2013",
      "category": "legal-watch",
      "claim": "California AB 2013 requires specified training-data documentation for covered generative-AI systems made available to Californians from 1 January 2026.",
      "class": "vendor-reported",
      "pin": "California Chapter 817, Statutes of 2024",
      "checked": "2026-07-19",
      "source": "https://leginfo.legislature.ca.gov/faces/billNavClient.xhtml?bill_id=202320240AB2013",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-governance-california-ab-2013",
      "migration_note": "Route applicability, exceptions, disclosure content, and update obligations to qualified counsel; maintain dataset lineage detailed enough to support the legal review."
    },
    {
      "id": "governance-nist-ai-600-1",
      "category": "risk-framework",
      "claim": "NIST AI 600-1 is the July 2024 Generative AI Profile for the voluntary AI Risk Management Framework and was updated on NIST's site in April 2026.",
      "class": "vendor-reported",
      "pin": "NIST AI 600-1",
      "checked": "2026-07-19",
      "source": "https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-governance-nist-ai-600-1",
      "migration_note": "Map profile actions to named controls, owners, evidence, and residual risks; a crosswalk or policy citation without operational evidence is not implementation."
    },
    {
      "id": "governance-usco-ai-report",
      "category": "legal-watch",
      "claim": "The U.S. Copyright Office published its AI report Part 2 on copyrightability in January 2025 and a pre-publication Part 3 on generative-AI training in May 2025.",
      "class": "vendor-reported",
      "pin": "USCO AI Report Parts 2 and 3 page checked 2026-07-19",
      "checked": "2026-07-19",
      "source": "https://www.copyright.gov/ai/",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-governance-usco-ai-report",
      "migration_note": "Counsel must track the final Part 3, litigation, licenses, jurisdiction, and work-specific facts; preserve provenance and human-contribution records without treating them as legal conclusions."
    },
    {
      "id": "governance-ftc-ai-accuracy-proposal",
      "category": "legal-watch",
      "claim": "On 30 June 2026 the U.S. FTC opened comment on a proposed policy statement concerning deceptive suppression of accuracy information in marketed AI systems.",
      "class": "vendor-reported",
      "pin": "FTC-2026-0859-0001 proposal",
      "checked": "2026-07-19",
      "source": "https://www.ftc.gov/policy/public-comments/policy-statement-concerning-suppression-accuracy-artificial-intelligence-systems",
      "owner_chapter": "appendices/c-landscape-2026.qmd",
      "owner_anchor": "landscape-ref-governance-ftc-ai-accuracy-proposal",
      "migration_note": "This is a proposal, not a final rule; legal and product-review owners must follow its status and substantiate public performance claims with reproducible evidence.",
      "status": "watch"
    }
  ]
}