[
  {
    "id": "O01",
    "title": "Building effective agents",
    "publisher": "Anthropic",
    "url": "https://www.anthropic.com/research/building-effective-agents",
    "published": "2024-12-19",
    "updated": null,
    "type": "vendor engineering guidance",
    "finding": "Added autonomy can improve task performance while increasing cost and latency. Start with a simple architecture and add complexity when the task justifies it.",
    "limitation": "Author experience, not a controlled comparison for Knauf. The page itself warns that tooling has changed since publication.",
    "accessed": "2026-10-22",
    "evidence_status": "opened_primary_source",
    "topic": "optimisation",
    "raw_archive": "not retained; source cards record bounded paraphrases and canonical links"
  },
  {
    "id": "O02",
    "title": "Latency optimization",
    "publisher": "OpenAI",
    "url": "https://developers.openai.com/api/docs/guides/latency-optimization",
    "published": null,
    "updated": null,
    "type": "official API documentation",
    "finding": "Fewer requests, appropriately shorter outputs and classical alternatives can reduce unnecessary model work.",
    "limitation": "Cost and latency do not move proportionally. Page-specific examples are not workload forecasts.",
    "accessed": "2026-10-22",
    "evidence_status": "opened_primary_source",
    "topic": "optimisation",
    "raw_archive": "not retained; source cards record bounded paraphrases and canonical links"
  },
  {
    "id": "O03",
    "title": "RouteLLM: Learning to Route LLMs with Preference Data, v4",
    "publisher": "Ong et al., arXiv / ICLR 2025",
    "url": "https://arxiv.org/html/2406.18665v4",
    "published": "2024-06-26",
    "updated": "2025-02-23",
    "type": "primary research paper, full text read",
    "finding": "Routing between model tiers can improve tested cost/performance trade-offs. Some out-of-domain evaluations performed poorly before targeted training-data augmentation.",
    "limitation": "Tested GPT-4/Mixtral pair and academic benchmarks do not establish a current enterprise savings rate.",
    "accessed": "2026-10-22",
    "evidence_status": "opened_primary_source",
    "topic": "optimisation",
    "raw_archive": "not retained; source cards record bounded paraphrases and canonical links"
  },
  {
    "id": "O04",
    "title": "Automatic Prefix Caching",
    "publisher": "vLLM",
    "url": "https://docs.vllm.ai/en/latest/features/automatic_prefix_caching/",
    "published": null,
    "updated": "2026-10-13",
    "type": "official runtime documentation, rolling latest",
    "finding": "Shared-prefix KV reuse avoids repeated prefill. It does not reduce the decoding work needed to generate new tokens.",
    "limitation": "Requires matching reusable prefixes. No financial saving is guaranteed for a fixed-capacity deployment.",
    "accessed": "2026-10-22",
    "evidence_status": "opened_primary_source",
    "topic": "optimisation",
    "raw_archive": "not retained; source cards record bounded paraphrases and canonical links"
  },
  {
    "id": "O05",
    "title": "Prompt caching",
    "publisher": "OpenAI",
    "url": "https://developers.openai.com/api/docs/guides/prompt-caching",
    "published": null,
    "updated": null,
    "type": "official API documentation",
    "finding": "Cache behavior and usage accounting depend on model generation and retention settings. Current documentation includes cache-write fields and charges.",
    "limitation": "Check model-specific pricing, actual cache hits and writes. No universal discount applies to all repeated content.",
    "accessed": "2026-10-22",
    "evidence_status": "opened_primary_source",
    "topic": "optimisation",
    "raw_archive": "not retained; source cards record bounded paraphrases and canonical links"
  },
  {
    "id": "O06",
    "title": "Batch API",
    "publisher": "OpenAI",
    "url": "https://developers.openai.com/api/docs/guides/batch",
    "published": null,
    "updated": null,
    "type": "official API documentation",
    "finding": "The service documents a 50% discount versus synchronous processing for eligible requests with a 24-hour completion window.",
    "limitation": "Asynchronous only, model/endpoint eligibility and expiry handling matter. The percentage applies to eligible API charges, not total business cost.",
    "accessed": "2026-10-22",
    "evidence_status": "opened_primary_source",
    "topic": "optimisation",
    "raw_archive": "not retained; source cards record bounded paraphrases and canonical links"
  },
  {
    "id": "O07",
    "title": "How we built our multi-agent research system",
    "publisher": "Anthropic",
    "url": "https://www.anthropic.com/research/multi-agent-research-system",
    "published": "2025-06-13",
    "updated": null,
    "type": "vendor system report",
    "finding": "The authors report that their agents consumed roughly four times chat tokens and multi-agent systems roughly fifteen times in their data.",
    "limitation": "Directional evidence about added work, not a universal multiplier. The reported performance gains use an internal research evaluation.",
    "accessed": "2026-10-22",
    "evidence_status": "opened_primary_source",
    "topic": "optimisation",
    "raw_archive": "not retained; source cards record bounded paraphrases and canonical links"
  },
  {
    "id": "O08",
    "title": "Evaluation best practices",
    "publisher": "OpenAI",
    "url": "https://developers.openai.com/api/docs/guides/evaluation-best-practices",
    "published": null,
    "updated": null,
    "type": "official engineering guidance",
    "finding": "Define acceptance criteria, use representative cases and domain expertise, compare changes and evaluate continuously.",
    "limitation": "Use the general method. Example thresholds and platform tools are not Knauf requirements.",
    "accessed": "2026-10-22",
    "evidence_status": "opened_primary_source",
    "topic": "optimisation",
    "raw_archive": "not retained; source cards record bounded paraphrases and canonical links"
  },
  {
    "id": "O09",
    "title": "Quantization",
    "publisher": "vLLM",
    "url": "https://docs.vllm.ai/en/latest/features/quantization/",
    "published": null,
    "updated": null,
    "type": "official runtime documentation, rolling latest",
    "finding": "Quantisation trades numerical precision for lower memory use. Supported methods differ by hardware.",
    "limitation": "The table establishes support, not guaranteed performance or acceptable quality for a particular model.",
    "accessed": "2026-10-22",
    "evidence_status": "opened_primary_source",
    "topic": "optimisation",
    "raw_archive": "not retained; source cards record bounded paraphrases and canonical links"
  },
  {
    "id": "O10",
    "title": "Batchers",
    "publisher": "NVIDIA Triton Inference Server",
    "url": "https://docs.nvidia.com/deeplearning/triton-inference-server/user-guide/docs/user_guide/batcher.html",
    "published": null,
    "updated": null,
    "type": "official server documentation",
    "finding": "Dynamic batching groups requests for throughput and exposes queue-delay limits. Tuning should observe a latency budget.",
    "limitation": "Documented dynamic batcher targets stateless models. LLM runtimes may use distinct continuous-batching mechanisms.",
    "accessed": "2026-10-22",
    "evidence_status": "opened_primary_source",
    "topic": "optimisation",
    "raw_archive": "not retained; source cards record bounded paraphrases and canonical links"
  },
  {
    "id": "O11",
    "title": "Optimization and Tuning",
    "publisher": "vLLM",
    "url": "https://docs.vllm.ai/en/stable/configuration/optimization/",
    "published": null,
    "updated": null,
    "type": "official runtime documentation, rolling stable",
    "finding": "Chunked prefill and scheduler settings affect throughput, time to first token and inter-token latency.",
    "limitation": "Defaults and suggested settings are version and workload dependent. Do not treat them as universal tuning prescriptions.",
    "accessed": "2026-10-22",
    "evidence_status": "opened_primary_source",
    "topic": "optimisation",
    "raw_archive": "not retained; source cards record bounded paraphrases and canonical links"
  },
  {
    "id": "O12",
    "title": "Speculative Decoding, v0.21.0",
    "publisher": "vLLM",
    "url": "https://docs.vllm.ai/en/v0.21.0/features/speculative_decoding/",
    "published": null,
    "updated": null,
    "type": "versioned official runtime documentation",
    "finding": "The documented target includes reducing inter-token latency for medium-to-low-QPS, memory-bound workloads.",
    "limitation": "Version-pinned mechanism example, not a claim about the latest release or universal throughput improvement.",
    "accessed": "2026-10-22",
    "evidence_status": "opened_primary_source",
    "topic": "optimisation",
    "raw_archive": "not retained; source cards record bounded paraphrases and canonical links"
  }
]