{ "format": "agent-article-v0", "article_id": "EA-094", "canonical_url": "/tools/ea-094-ai-latency-slo", "page_type": "reference", "claim_ids": [], "evidence_ids": [ "EVD-EA-094-URL-06C934AF218E", "EVD-EA-094-URL-7B9A3E66DBB9", "EA-094-R11-RESEARCH", "EA-094-R11-WEB-READBACK" ], "unknown_ids": [ "EA-094-UNKNOWN-01", "EA-094-UNKNOWN-CURRENTNESS" ], "operational_object_ids": [], "provider": [], "models": [], "plans": [], "surfaces": [], "regions": [], "current_status": "UNKNOWN", "observation_status": "SPEC_ONLY", "freshness": "UNKNOWN", "last_verified": null, "supersedes": [], "superseded_by": [], "provenance": [ "COMM-082-EA0094", "SRC-EA-094-URL-06C934AF218E", "SRC-EA-094-URL-7B9A3E66DBB9", "WEBBODY-EA0094-R8-A0DFFE91" ], "article_sources": [ { "id": "EVD-EA-094-URL-06C934AF218E", "evidence_class": "RESEARCH", "title": "1. [SRE practice] Google SRE Workbook, “Implementing SLOs”, Steven Thurgood et al., accessed 2026-09-16.", "href": "https://sre.google/workbook/implementing-slos/", "supports": [], "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。", "origin_class": "SOURCE_PACKET_REFERENCE", "source_or_experiment_ids": [ "SRC-EA-094-URL-06C934AF218E" ], "supporting_locators": [ "EVIDENCE TABLE", "1. [SRE practice] Google SRE Workbook, “Implementing SLOs”, Steven Thurgood et al., accessed 2026-09-16. https://sre.google/workbook/implementing-slos/" ], "scope": "EVIDENCE TABLE", "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED" }, { "id": "EVD-EA-094-URL-7B9A3E66DBB9", "evidence_class": "RESEARCH", "title": "2. [SRE practice] Google SRE Book, “Service Level Objectives”, Chris Jones, John Wilkes, Niall Murphy et al., accessed 2026-09-16.", "href": "https://sre.google/sre-book/service-level-objectives/", "supports": [], "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。", "origin_class": "SOURCE_PACKET_REFERENCE", "source_or_experiment_ids": [ "SRC-EA-094-URL-7B9A3E66DBB9" ], "supporting_locators": [ "EVIDENCE TABLE", "2. [SRE practice] Google SRE Book, “Service Level Objectives”, Chris Jones, John Wilkes, Niall Murphy et al., accessed 2026-09-16. https://sre.google/sre-book/service-level-objectives/" ], "scope": "EVIDENCE TABLE", "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED" }, { "id": "EA-094-R11-RESEARCH", "evidence_class": "RESEARCH", "title": "EA-094 監査済みresearch / canonical revision binding", "supports": [ "- AI latencyは `queue wait + retrieval + prompt assembly + provider admission + TTFT + generation + tool waits + validation + approval + commit` に分解する。\n- interactive、background、batch、human-in-loopを別SLO classにする。\n- model TTFT/TPOTだけでなく end-to-end completion、time-to-useful-first-result、deadline success ratioを持つ。\n- P50/P95/P99とthreshold-based good-event ratioを併用し、平均のみを採用しない。" ], "limit": "canonical監査・revision bindingの存在を示す。外部ソースcurrentnessが本文またはcontrol-planeで確定していない場合は、currentnessをUNKNOWNのまま保持する。", "origin_class": "AUDITED_CANONICAL_RESEARCH_PACKET", "source_or_experiment_ids": [ "COMM-082-EA0094" ], "supporting_locators": [ "COMM-082-EA0094", "EA-094-R5-LEGACY-MIGRATION-20261008-R1", "content/staging/EA-094.md" ], "scope": "監査済みEA本文・canonical revision・既存Source Graph relationの範囲。欠損currentnessは推測で閉じない。", "provenance_status": "AUDIT_PASS_CANONICAL_BINDING" }, { "id": "EA-094-R11-WEB-READBACK", "evidence_class": "IMPL", "title": "EA-094 Web body identity / canonical readback", "supports": [ "canonical staging bodyのSHA-256一致とR9 canonical ledger readbackをProduction projectionの入力として検証する。" ], "limit": "body identity/readbackは本文claim自体の真偽や外部ソースの最新性を追加保証しない。", "checked_at": null, "origin_class": "IMPLEMENTATION_READBACK", "source_or_experiment_ids": [ "WEBBODY-EA0094-R8-A0DFFE91" ], "supporting_locators": [ "WEBBODY-EA0094-R8-A0DFFE91", "content/staging/EA-094.md", "EA-094-R5-LEGACY-MIGRATION-20261008-R1" ], "scope": "R9/R10 canonical readback後のProduction生成時body identity。", "provenance_status": "CANONICAL_READBACK_BODY_HASH_MATCH" } ], "research_sections": [ { "heading": "WHY_THIS_MATTERS_TO_UPDATE", "text": "model API latencyだけを測ると、queue、retrieval、tool、retry、streaming、human approvalを含む実際の業務待ち時間を見失う。平均値はtail latencyと失敗したuser journeyを隠す。" }, { "heading": "SCOPE_ROUTING", "text": "Enterprise AI Core / performance / reliability / observability / benchmark" }, { "heading": "CLAIM_OR_RESEARCH_QUESTION", "text": "AI SLOはmodel-call単位ではなくcritical user journeyとstage breakdownを併用すべきか。" }, { "heading": "EVIDENCE TABLE", "text": "1. [SRE practice] Google SRE Workbook, “Implementing SLOs”, Steven Thurgood et al., accessed 2026-09-16. https://sre.google/workbook/implementing-slos/ — SLO/SLI/error budget、user-centric journey、latency threshold、90%/99%等複数grade、dependency modelingを提示。 relation=support.\n2. [SRE practice] Google SRE Book, “Service Level Objectives”, Chris Jones, John Wilkes, Niall Murphy et al., accessed 2026-09-16. https://sre.google/sre-book/service-level-objectives/ — SLI/SLOをuser-relevant behaviorとして定義し、複数percentile/thresholdとworkload class別objectiveを例示。 relation=support/context." }, { "heading": "FINDINGS", "text": "- AI latencyは `queue wait + retrieval + prompt assembly + provider admission + TTFT + generation + tool waits + validation + approval + commit` に分解する。\n- interactive、background、batch、human-in-loopを別SLO classにする。\n- model TTFT/TPOTだけでなく end-to-end completion、time-to-useful-first-result、deadline success ratioを持つ。\n- P50/P95/P99とthreshold-based good-event ratioを併用し、平均のみを採用しない。" }, { "heading": "LIMITS / COUNTEREVIDENCE / FAILURE ENVELOPE", "text": "速い誤答はSLO成功ではない。latencyだけを最適化するとquality/safety/costを劣化させ得る。streamingはperceived latencyを改善しても業務完了時刻を短縮しない場合がある。retry/fallbackはavailabilityを上げつつtail latencyを悪化させる。" }, { "heading": "AVAILABILITY_GATES", "text": "workload class、model/provider/region、streaming/tool availability、queue/admission policy、provisioned/shared capacity。" }, { "heading": "IMPLEMENTATION_STATE", "text": "将来Core候補 `AI_JOURNEY_SLO_CONTRACT`: journey ID, workload class, start/end semantic points, stage spans, deadline, TTFT, TPOT, end-to-end duration, success/quality gate, percentile/threshold objectives, error budget, dependency attribution, cost。" }, { "heading": "BENCHMARK_OR_EVAL", "text": "固定prompt長/output長だけでなくshort/long context、tool 0/1/N、RAG、retry、concurrency、cold/warmを層別。P50/P95/P99、deadline-good-event ratio、quality-qualified latency、costを同時測定。" }, { "heading": "SECURITY / PRIVACY / AUDIT / COST / OPERABILITY", "text": "traceにはprompt本文を必須とせず、hash/size/classificationで機微情報を避ける。SLO breachをprovider責任へ自動帰属せずstage spanで根拠化。" }, { "heading": "SAFE_WORDING_FOR_FUTURE_BOOK", "text": "「AI性能はtokens/secだけでは測れない。利用者が必要な結果を得るまでのjourneyをSLO化し、queue・retrieval・model・tool・retry・approvalを分解してtail latencyを管理する。」" }, { "heading": "UPDATE_WORK_HANDOFF", "text": "EA-093 admission/retry telemetryを同じtraceへ接続。EA-091 routeごとのquality-qualified latencyを比較し、fallbackの速度だけでrouteを選ばない。" }, { "heading": "OPEN_UNKNOWN / STOP_CONDITION", "text": "各providerの公開SLA/latency guarantee、OpenTelemetry GenAI semantic conventionsの公開後current revisionは再pin対象。tenant workload SLO値はorganization-specificでUNKNOWNのまま。SLO contractとfixture確立につき停止条件充足。\n\n---" } ], "human_url": "/tools/ea-094-ai-latency-slo", "markdown_url": "/agent/article.md?path=%2Ftools%2Fea-094-ai-latency-slo", "json_url": "/agent/article.json?path=%2Ftools%2Fea-094-ai-latency-slo" }