{
  "format": "agent-article-v0",
  "article_id": "PM-GPT-CORE-07",
  "canonical_url": "/providers/pm-gpt-core-07",
  "page_type": "reference",
  "claim_ids": [],
  "evidence_ids": [
    "EVD-PM-GPT-CORE-07-URL-8F4F2487ED99",
    "EVD-PM-GPT-CORE-07-URL-A01937402C35",
    "EVD-PM-GPT-CORE-07-URL-3F46CD4F9265",
    "EVD-PM-GPT-CORE-07-URL-E4B992CCC389",
    "EVD-PM-GPT-CORE-07-URL-6DB7A0E61936",
    "EVD-PM-GPT-CORE-07-URL-850B438DC6B6",
    "EVD-PM-GPT-CORE-07-URL-8679A00F99F7",
    "EVD-PM-GPT-CORE-07-URL-BAB0E18B2802",
    "EVD-PM-GPT-CORE-07-URL-3140024AB7C6",
    "EVD-PM-GPT-CORE-07-URL-E4E7F033E252",
    "EVD-PM-GPT-CORE-07-URL-827ACF165DA2",
    "EVD-PM-GPT-CORE-07-URL-16D2741870AD",
    "PM-GPT-CORE-07-DOCUMENT-READBACK"
  ],
  "unknown_ids": [
    "PM-GPT-CORE-07-CURRENTNESS"
  ],
  "operational_object_ids": [],
  "provider": [
    "GPT"
  ],
  "models": [],
  "plans": [],
  "surfaces": [],
  "regions": [],
  "current_status": "UNKNOWN",
  "observation_status": "UNOBSERVED",
  "freshness": "UNKNOWN",
  "last_verified": null,
  "supersedes": [],
  "superseded_by": [],
  "provenance": [
    "DOC-083cde4caa627644-PM-GPT-CORE-07",
    "SRC-PM-GPT-CORE-07-URL-8F4F2487ED99",
    "SRC-PM-GPT-CORE-07-URL-A01937402C35",
    "SRC-PM-GPT-CORE-07-URL-3F46CD4F9265",
    "SRC-PM-GPT-CORE-07-URL-E4B992CCC389",
    "SRC-PM-GPT-CORE-07-URL-6DB7A0E61936",
    "SRC-PM-GPT-CORE-07-URL-850B438DC6B6",
    "SRC-PM-GPT-CORE-07-URL-8679A00F99F7",
    "SRC-PM-GPT-CORE-07-URL-BAB0E18B2802",
    "SRC-PM-GPT-CORE-07-URL-3140024AB7C6",
    "SRC-PM-GPT-CORE-07-URL-E4E7F033E252",
    "SRC-PM-GPT-CORE-07-URL-827ACF165DA2",
    "SRC-PM-GPT-CORE-07-URL-16D2741870AD"
  ],
  "article_sources": [
    {
      "id": "EVD-PM-GPT-CORE-07-URL-8F4F2487ED99",
      "evidence_class": "RESEARCH",
      "title": "OpenAI, Evaluation best practices: <[",
      "href": "https://developers.openai.com/api/docs/guides/evaluation-best-practices",
      "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": "APPROVED_DOCUMENT_REFERENCE",
      "source_or_experiment_ids": [
        "SRC-PM-GPT-CORE-07-URL-8F4F2487ED99"
      ],
      "supporting_locators": [
        "Provenance",
        "- OpenAI, Evaluation best practices: <[https://developers.openai.com/api/docs/guides/evaluation-best-practices"
      ],
      "scope": "Provenance",
      "provenance_status": "APPROVED_DOCUMENT_REFERENCE_PROJECTED"
    },
    {
      "id": "EVD-PM-GPT-CORE-07-URL-A01937402C35",
      "evidence_class": "RESEARCH",
      "title": "OpenAI, Working with evals: <[",
      "href": "https://developers.openai.com/api/docs/guides/evals",
      "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": "APPROVED_DOCUMENT_REFERENCE",
      "source_or_experiment_ids": [
        "SRC-PM-GPT-CORE-07-URL-A01937402C35"
      ],
      "supporting_locators": [
        "Provenance",
        "- OpenAI, Working with evals: <[https://developers.openai.com/api/docs/guides/evals"
      ],
      "scope": "Provenance",
      "provenance_status": "APPROVED_DOCUMENT_REFERENCE_PROJECTED"
    },
    {
      "id": "EVD-PM-GPT-CORE-07-URL-3F46CD4F9265",
      "evidence_class": "RESEARCH",
      "title": "OpenAI, Datasets / Graders / Agent evals: <[",
      "href": "https://developers.openai.com/api/docs/guides/evaluation-getting-started",
      "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": "APPROVED_DOCUMENT_REFERENCE",
      "source_or_experiment_ids": [
        "SRC-PM-GPT-CORE-07-URL-3F46CD4F9265"
      ],
      "supporting_locators": [
        "Provenance",
        "- OpenAI, Datasets / Graders / Agent evals: <[https://developers.openai.com/api/docs/guides/evaluation-getting-started"
      ],
      "scope": "Provenance",
      "provenance_status": "APPROVED_DOCUMENT_REFERENCE_PROJECTED"
    },
    {
      "id": "EVD-PM-GPT-CORE-07-URL-E4B992CCC389",
      "evidence_class": "RESEARCH",
      "title": "OpenAI, Datasets / Graders / Agent evals: <[https://developers.openai.com/api/docs/guides/evaluation-getting-started](https://developers.openai.com/api/docs/guides/evaluation-getti",
      "href": "https://developers.openai.com/api/docs/guides/graders",
      "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": "APPROVED_DOCUMENT_REFERENCE",
      "source_or_experiment_ids": [
        "SRC-PM-GPT-CORE-07-URL-E4B992CCC389"
      ],
      "supporting_locators": [
        "Provenance",
        "- OpenAI, Datasets / Graders / Agent evals: <[https://developers.openai.com/api/docs/guides/evaluation-getting-started](https://developers.openai.com/api/docs/guides/evaluation-getting-started)>; <[https://developers.openai.com/api/docs/guides/graders"
      ],
      "scope": "Provenance",
      "provenance_status": "APPROVED_DOCUMENT_REFERENCE_PROJECTED"
    },
    {
      "id": "EVD-PM-GPT-CORE-07-URL-6DB7A0E61936",
      "evidence_class": "RESEARCH",
      "title": "OpenAI, Datasets / Graders / Agent evals: <[https://developers.openai.com/api/docs/guides/evaluation-getting-started](https://developers.openai.com/api/docs/guides/evaluation-getti",
      "href": "https://developers.openai.com/api/docs/guides/agent-evals",
      "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": "APPROVED_DOCUMENT_REFERENCE",
      "source_or_experiment_ids": [
        "SRC-PM-GPT-CORE-07-URL-6DB7A0E61936"
      ],
      "supporting_locators": [
        "Provenance",
        "- OpenAI, Datasets / Graders / Agent evals: <[https://developers.openai.com/api/docs/guides/evaluation-getting-started](https://developers.openai.com/api/docs/guides/evaluation-getting-started)>; <[https://developers.openai.com/api/docs/guides/graders](https://developers.openai.com/api/docs/guides/graders)>; <[https://developers.openai.com/api/docs/guides/agent-evals"
      ],
      "scope": "Provenance",
      "provenance_status": "APPROVED_DOCUMENT_REFERENCE_PROJECTED"
    },
    {
      "id": "EVD-PM-GPT-CORE-07-URL-850B438DC6B6",
      "evidence_class": "RESEARCH",
      "title": "OpenAI, Prompt optimizer / Fine-tuning best practices: <[",
      "href": "https://developers.openai.com/api/docs/guides/prompt-optimizer",
      "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": "APPROVED_DOCUMENT_REFERENCE",
      "source_or_experiment_ids": [
        "SRC-PM-GPT-CORE-07-URL-850B438DC6B6"
      ],
      "supporting_locators": [
        "Provenance",
        "- OpenAI, Prompt optimizer / Fine-tuning best practices: <[https://developers.openai.com/api/docs/guides/prompt-optimizer"
      ],
      "scope": "Provenance",
      "provenance_status": "APPROVED_DOCUMENT_REFERENCE_PROJECTED"
    },
    {
      "id": "EVD-PM-GPT-CORE-07-URL-8679A00F99F7",
      "evidence_class": "RESEARCH",
      "title": "OpenAI, Prompt optimizer / Fine-tuning best practices: <[https://developers.openai.com/api/docs/guides/prompt-optimizer](https://developers.openai.com/api/docs/guides/prompt-optimi",
      "href": "https://developers.openai.com/api/docs/guides/fine-tuning-best-practices",
      "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": "APPROVED_DOCUMENT_REFERENCE",
      "source_or_experiment_ids": [
        "SRC-PM-GPT-CORE-07-URL-8679A00F99F7"
      ],
      "supporting_locators": [
        "Provenance",
        "- OpenAI, Prompt optimizer / Fine-tuning best practices: <[https://developers.openai.com/api/docs/guides/prompt-optimizer](https://developers.openai.com/api/docs/guides/prompt-optimizer)>; <[https://developers.openai.com/api/docs/guides/fine-tuning-best-practices"
      ],
      "scope": "Provenance",
      "provenance_status": "APPROVED_DOCUMENT_REFERENCE_PROJECTED"
    },
    {
      "id": "EVD-PM-GPT-CORE-07-URL-BAB0E18B2802",
      "evidence_class": "RESEARCH",
      "title": "OpenAI, Deprecations: <[",
      "href": "https://developers.openai.com/api/docs/deprecations",
      "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": "APPROVED_DOCUMENT_REFERENCE",
      "source_or_experiment_ids": [
        "SRC-PM-GPT-CORE-07-URL-BAB0E18B2802"
      ],
      "supporting_locators": [
        "Provenance",
        "- OpenAI, Deprecations: <[https://developers.openai.com/api/docs/deprecations"
      ],
      "scope": "Provenance",
      "provenance_status": "APPROVED_DOCUMENT_REFERENCE_PROJECTED"
    },
    {
      "id": "EVD-PM-GPT-CORE-07-URL-3140024AB7C6",
      "evidence_class": "RESEARCH",
      "title": "OpenAI, openai/evals: <[",
      "href": "https://github.com/openai/evals",
      "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": "APPROVED_DOCUMENT_REFERENCE",
      "source_or_experiment_ids": [
        "SRC-PM-GPT-CORE-07-URL-3140024AB7C6"
      ],
      "supporting_locators": [
        "Provenance",
        "- OpenAI, openai/evals: <[https://github.com/openai/evals"
      ],
      "scope": "Provenance",
      "provenance_status": "APPROVED_DOCUMENT_REFERENCE_PROJECTED"
    },
    {
      "id": "EVD-PM-GPT-CORE-07-URL-E4E7F033E252",
      "evidence_class": "RESEARCH",
      "title": "OpenAI, openai/evals: <[https://github.com/openai/evals](https://github.com/openai/evals)>; commit <[",
      "href": "https://github.com/openai/evals/commit/8eac7a7de5215c907fbddc30efdaf316913eccdd",
      "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": "APPROVED_DOCUMENT_REFERENCE",
      "source_or_experiment_ids": [
        "SRC-PM-GPT-CORE-07-URL-E4E7F033E252"
      ],
      "supporting_locators": [
        "Provenance",
        "- OpenAI, openai/evals: <[https://github.com/openai/evals](https://github.com/openai/evals)>; commit <[https://github.com/openai/evals/commit/8eac7a7de5215c907fbddc30efdaf316913eccdd"
      ],
      "scope": "Provenance",
      "provenance_status": "APPROVED_DOCUMENT_REFERENCE_PROJECTED"
    },
    {
      "id": "EVD-PM-GPT-CORE-07-URL-827ACF165DA2",
      "evidence_class": "RESEARCH",
      "title": "Shi et al., Judging the Judges, ACL-IJCNLP 2025: <[",
      "href": "https://aclanthology.org/2025.ijcnlp-long.18/",
      "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": "APPROVED_DOCUMENT_REFERENCE",
      "source_or_experiment_ids": [
        "SRC-PM-GPT-CORE-07-URL-827ACF165DA2"
      ],
      "supporting_locators": [
        "Provenance",
        "- Shi et al., Judging the Judges, ACL-IJCNLP 2025: <[https://aclanthology.org/2025.ijcnlp-long.18/"
      ],
      "scope": "Provenance",
      "provenance_status": "APPROVED_DOCUMENT_REFERENCE_PROJECTED"
    },
    {
      "id": "EVD-PM-GPT-CORE-07-URL-16D2741870AD",
      "evidence_class": "RESEARCH",
      "title": "Dekoninck et al., ConStat: <[",
      "href": "https://arxiv.org/abs/2405.16281",
      "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": "APPROVED_DOCUMENT_REFERENCE",
      "source_or_experiment_ids": [
        "SRC-PM-GPT-CORE-07-URL-16D2741870AD"
      ],
      "supporting_locators": [
        "Provenance",
        "- Dekoninck et al., ConStat: <[https://arxiv.org/abs/2405.16281"
      ],
      "scope": "Provenance",
      "provenance_status": "APPROVED_DOCUMENT_REFERENCE_PROJECTED"
    },
    {
      "id": "PM-GPT-CORE-07-DOCUMENT-READBACK",
      "evidence_class": "IMPL",
      "title": "原資料との収録照合",
      "supports": [
        "資料単位の本文ハッシュ、段落数、表数を照合しています。"
      ],
      "limit": "収録照合は主張の真偽や外部根拠の最新性を追加保証しません。",
      "origin_class": "DOCUMENT_INGESTION_READBACK",
      "source_or_experiment_ids": [
        "DOC-083cde4caa627644-PM-GPT-CORE-07"
      ],
      "supporting_locators": [
        "18WPldGemks2px0ukhKYINeBMrBwAmlAi",
        "083cde4caa627644036a2225f1f6ff29b8015bd902614fe1f7a00a997b4eb59c",
        "provider_mappings/PM-GPT-CORE-07.md",
        "28b16e7672f7e23b934410b7a8ce9e715ab5eb3bbe68f50c7730951b425b9345"
      ],
      "scope": "公開指定を受けた資料単位の全文・表・出典。",
      "provenance_status": "SOURCE_IDENTITY_VERIFIED"
    }
  ],
  "research_sections": [
    {
      "heading": "資料の識別情報",
      "text": "Source SHA-256 bffa68bb3a761f7e193bfa5647b19f0503bac9de371403111bf9ba8007ba67ea"
    },
    {
      "heading": "Product Mapping — GPT / CORE-07",
      "text": "PRODUCT_MAPPING_ID: PM-GPT-CORE-07-v0.1\n\nPROVIDER: GPT / OpenAI\n\nCORE_PACKET: PKT-GEMINI-A-CORE-07-v0.1\n\nAS_OF: 2026-09-09 JST\n\nSTATUS: PUBLIC_MAPPING_BOUNDED_COMPLETE / HOSTED_EVALS_DEPRECATION_ACTIVE / LIVE_ORG_PENDING\n\nLAYER: PRODUCT_MAPPING (not Enterprise AI Core; not Organization Overlay)"
    },
    {
      "heading": "Scope",
      "text": "CORE-07のgolden dataset、holdout、scorer/judge registry、slice-level AND gate、release/rollback evidenceをOpenAIのDatasets、Evals、Graders、Agent evals、Prompt optimizerおよびopenai/evals OSSへ対応づける。公開仕様と固定OSSだけを対象とし、実tenant、非公開dataset、model/judgeの実性能、組織閾値はOrganization Overlayへ残す。"
    },
    {
      "heading": "Surface and lifecycle separation",
      "text": "| SURFACE | CURRENT EVIDENCE | LIFECYCLE DECISION | LIMIT |\n| --- | --- | --- | --- |\n| dashboard Datasets | rows、dynamic data、prompts、annotations、graderを保持して反復評価を支援 | rapid iteration用artifactとして扱い、dataset provenance、split、export、access、retentionを別台帳で固定 | UI上のdataset名・IDだけではimmutability、hidden holdout隔離、長期保存を証明しない |\n| Evals API/dashboard | data_source_config、item schema、testing_criteria、runs、result countsを提供 | 2026-06-03 deprecated、2026-10-31既存evals read-only、2026-11-30 dashboard/API shutdown予定のACTIVE_DEPRECATED_DEPENDENCY | forward release gateとして新規依存しない。停止予定は現在の即時利用不能を意味しない |\n| Graders | string、text-similarity、score-model、label-model、Python等の評価器 | deterministic checkとmodel judgeを分離し、grader definition、model、prompt、threshold、range、sampling、error policyをversion固定 | grader scoreは人手正解や全slice品質の代用ではなく、grader基盤自体もdeprecation対象 |\n| Agent evals / traces | traceからdataset/eval runを構成し、agent workflow変更を比較可能 | trajectory評価をrelease evidenceの一部として採用 | traceは外部systemのauthoritative final state、業務成果、完全なside effectを証明しない |\n| Prompt optimizer | dataset、annotations、grader resultsを用いてprompt改善を反復 | optimization corpusとsealed holdoutを分離し、再利用回数・閲覧・派生promptをleak ledgerへ記録 | 同一datasetでの反復最適化はtest overfittingを招き得る。改善scoreを未見分布へ外挿しない |\n| openai/evals OSS | author-maintained framework/registry。2026-04-14 commit 8eac7a7de5215c907fbddc30efdaf316913eccddを確認 | self-hosted/external harness候補としてhosted Evals lifecycleと別dependency化 | GitHub repositoryの存続はhosted API/dashboardの移行先・SLA・互換保証ではない |"
    },
    {
      "heading": "Core → GPT mapping rows",
      "text": "| ROW | CORE CONTRACT | GPT / OPENAI FINDING | DECISION / UNKNOWN |\n| --- | --- | --- | --- |\n| PM-C07-GPT-001 | evaluation artifact separation | Datasets、Evals runs、Graders、Agent traces、Prompt optimizerは役割とlifecycleが異なる | 一つの「OpenAI eval」へ束ねず、dataset/version、candidate config、scorer/judge、run、trace、release decisionを別IDで結合する |\n| PM-C07-GPT-002 | temporal dependency | hosted EvalsとGradersは2026-06-03 deprecated。既存evalsは2026-10-31 read-only、2026-11-30 shutdown予定 | 新規release gateのsingle point of failureにしない。migration owner、export、dual-run、cutover、rollback期限を要求。後継の完全同値surfaceは公開情報ではUNKNOWN |\n| PM-C07-GPT-003 | dataset schema and run evidence | Evals APIはitem JSON schema、testing criteria、run result countsを持つ | schema/run ID/countをCore evaluation contractのcomponentとして記録。ただしprovenance、holdout access、slice completeness、immutability、retentionは別証拠が必要 |\n| PM-C07-GPT-004 | deterministic-first scoring | Datasets/Gradersは複数grader typeを提供し、Python graderの例外・invalid floatはinvalid/0として扱われる | exact/schema/rule checksを先行させ、grader failureとcandidate failureを分離。aggregate scoreだけでreleaseしない |\n| PM-C07-GPT-005 | judge registry and calibration | score-model graderはmodel、prompt、threshold、range、samplingを設定する。独立研究はposition biasがjudge/taskで変動することを示す | judge ID/model/snapshot、prompt/rubric、sampling、candidate order/permutation、abstention、human disagreementを保存。swap testと人手calibrationなしにjudge scoreをground truth化しない |\n| PM-C07-GPT-006 | optimization/holdout split | Prompt optimizerはdataset、annotation、grader resultを反復利用する | optimization corpusとsealed holdout/challenge setを分離。prompt optimizerへの閲覧・feedback・再実行をcontamination eventとして追跡 |\n| PM-C07-GPT-007 | hosted/OSS boundary | openai/evals OSSはhosted Evalsとは別artifactで、公開repoにrelease tagはなくcommit履歴で固定可能 | commit SHA、dependencies、registry、runner、result digestを固定。OSS利用可能性をhosted replacement guaranteeへ昇格しない |\n| PM-C07-GPT-008 | bounded contamination counterevidence | ConStatはprimary/reference benchmarkと同分布のrephrased/synthetic samplesを比較し、非一般化したscore inflationを検出する | named-model contaminationの断定には使わず、public benchmark、rephrase、fresh challenge、sealed holdoutのgapをrelease前に測る反証profileとして採用 |\n| PM-C07-GPT-009 | release decision and readback | OpenAIはtask-specific data、実分布、continuous evaluation、人手feedbackとのcalibrationを推奨する | overall平均ではなくcritical slice AND gate、variance、known regression、safety/over-refusalを個別判定し、release/rollback後のeffective configurationと受入結果をread backする |"
    },
    {
      "heading": "Safe verification profile",
      "text": "非機密synthetic fixtureと既存のprovider-neutral 12/12 diagnosticだけを使う。今回はOpenAI API、dashboard、tenant、dataset、grader、optimizer、model、traceを操作していない。\n\n- Datasets、Evals、Graders、Agent traces、Prompt optimizer、OSS harnessを別dependencyとしてinventoryし、hosted deprecation日とmigration ownerを付ける。\n\n- candidate、rendered input、tool/source/policy、dataset split/version、scorer/judge、environment、run、release decisionを一つのevaluation contractへ参照結合する。\n\n- exact/schema/rule checksを先行し、grader exception、timeout、invalid score、missing sliceをcandidate failureと別分類する。\n\n- model judgeではcandidate順序をswap/permutationし、judge version・samplingを固定して反復し、人手labelとのdisagreementをslice別に測る。\n\n- tuning/optimizerへ触れたcaseをsealed holdoutから失効させ、public benchmark、rephrase、fresh challenge、failure corpusのscore gapを記録する。\n\n- overall passでもcritical slice、known regression、Japanese、safety、over-refusal、tail、varianceのいずれかがfailならreleaseをblockする。\n\n- hosted export/dual-run後にrun count、per-item result、grader definition、artifact digest、retention/readabilityをread backし、移行不能ならmanual release stopへ切り替える。\n\n既存EXP-C07-001の12/12は本研究のCOMPILER_EXPERIMENTであり、GPT/OpenAI、hosted Evals、grader、model judgeのlive試験へ再分類しない。"
    },
    {
      "heading": "Organization Overlay handoff",
      "text": "- adopted evaluation surfaces、workspace/tenant、dataset owner、access group、retention、export location\n\n- dataset provenance、PII/security review、language/risk slice、sealed holdout custodian、leak response\n\n- grader/judge allowlist、model snapshot、rubric owner、human calibration sample、disagreement threshold\n\n- release threshold、critical slice、variance budget、zero-tolerance failure、exception approver\n\n- Evals/Graders migration owner、2026-10-31/11-30 cutover、dual-run、rollback、evidence retention\n\n- production telemetry、incident/failure corpus intake、continuous-eval cadence、release/rollback owner"
    },
    {
      "heading": "Provenance",
      "text": "- OpenAI, Evaluation best practices: <[https://developers.openai.com/api/docs/guides/evaluation-best-practices](https://developers.openai.com/api/docs/guides/evaluation-best-practices)> (accessed 2026-09-09)\n\n- OpenAI, Working with evals: <[https://developers.openai.com/api/docs/guides/evals](https://developers.openai.com/api/docs/guides/evals)> (accessed 2026-09-09)\n\n- OpenAI, Datasets / Graders / Agent evals: <[https://developers.openai.com/api/docs/guides/evaluation-getting-started](https://developers.openai.com/api/docs/guides/evaluation-getting-started)>; <[https://developers.openai.com/api/docs/guides/graders](https://developers.openai.com/api/docs/guides/graders)>; <[https://developers.openai.com/api/docs/guides/agent-evals](https://developers.openai.com/api/docs/guides/agent-evals)> (accessed 2026-09-09)\n\n- OpenAI, Prompt optimizer / Fine-tuning best practices: <[https://developers.openai.com/api/docs/guides/prompt-optimizer](https://developers.openai.com/api/docs/guides/prompt-optimizer)>; <[https://developers.openai.com/api/docs/guides/fine-tuning-best-practices](https://developers.openai.com/api/docs/guides/fine-tuning-best-practices)> (accessed 2026-09-09)\n\n- OpenAI, Deprecations: <[https://developers.openai.com/api/docs/deprecations](https://developers.openai.com/api/docs/deprecations)> (accessed 2026-09-09)\n\n- OpenAI, openai/evals: <[https://github.com/openai/evals](https://github.com/openai/evals)>; commit <[https://github.com/openai/evals/commit/8eac7a7de5215c907fbddc30efdaf316913eccdd](https://github.com/openai/evals/commit/8eac7a7de5215c907fbddc30efdaf316913eccdd)>\n\n- Shi et al., Judging the Judges, ACL-IJCNLP 2025: <[https://aclanthology.org/2025.ijcnlp-long.18/](https://aclanthology.org/2025.ijcnlp-long.18/)>\n\n- Dekoninck et al., ConStat: <[https://arxiv.org/abs/2405.16281](https://arxiv.org/abs/2405.16281)>"
    },
    {
      "heading": "Exit condition",
      "text": "OpenAI固有surface、hosted Evals/Gradersの停止予定、judge bias、optimizer/holdout漏洩、hosted/OSS境界をCORE-07へ対応づけた。公開文書内の「DatasetからEvalへexport」という現行案と、Evals停止予定の時間的不一致は移行残差として保持し、公開されていない完全な後継・export/retention保証を推測しない。Core claimは変更しないためGemini遡及監査は不要である。"
    }
  ],
  "human_url": "/providers/pm-gpt-core-07",
  "markdown_url": "/agent/article.md?path=%2Fproviders%2Fpm-gpt-core-07",
  "json_url": "/agent/article.json?path=%2Fproviders%2Fpm-gpt-core-07"
}