# CLAUDE：モデル・ツール・情報源・状態の選択

CURRENT_STATUS: UNKNOWN
OBSERVATION: UNOBSERVED
FRESHNESS: UNKNOWN
PAGE_TYPE: reference
CANONICAL_URL: /providers/pm-claude-core-04
CLAIM_IDS: none
EVIDENCE_IDS: EVD-PM-CLAUDE-CORE-04-URL-9332683246DB, EVD-PM-CLAUDE-CORE-04-URL-82747C62F33E, EVD-PM-CLAUDE-CORE-04-URL-A2C7CB24DC72, EVD-PM-CLAUDE-CORE-04-URL-DBCB362197DB, EVD-PM-CLAUDE-CORE-04-URL-DF1A1BB212DB, EVD-PM-CLAUDE-CORE-04-URL-854B88AAD19B, EVD-PM-CLAUDE-CORE-04-URL-ED2EA3ED3DC1, EVD-PM-CLAUDE-CORE-04-URL-2AA41A0F518B, EVD-PM-CLAUDE-CORE-04-URL-2B13AE4354AB, EVD-PM-CLAUDE-CORE-04-URL-8A6DFE25191D, EVD-PM-CLAUDE-CORE-04-URL-76592DD9412F, PM-CLAUDE-CORE-04-DOCUMENT-READBACK
UNKNOWN_IDS: PM-CLAUDE-CORE-04-CURRENTNESS
OPS_IDS: none

## Summary
CORE-04のmodel / tool / source / state routing、fallback、batch、observability契約をClaude API、Message Batches、service tier、Claude Agent SDKへ対応づける。Claude.ai、Claude API、Agent SDK、Claude Code CLI、Bedrock、Google Cloud、Managed Agentsを同一routeとせず、公開仕様から閉じないtenant・価格・quota・live品質・内部routingはUNKNOWNに残す。

## Current judgment
CORE-04のmodel / tool / source / state routing、fallback、batch、observability契約をClaude API、Message Batches、service tier、Claude Agent SDKへ対応づける。Claude.ai、Claude API、Agent SDK、Claude Code CLI、Bedrock、Google Cloud、Managed Agentsを同一routeとせず、公開仕様から閉じないtenant・価格・quota・live品質・内部routingはUNKNOWNに残す。

## Scope
CORE-04のmodel / tool / source / state routing、fallback、batch、observability契約をClaude API、Message Batches、service tier、Claude Agent SDKへ対応づける。Claude.ai、Claude API、Agent SDK、Claude Code CLI、Bedrock、Google Cloud、Managed Agentsを同一routeとせず、公開仕様から閉じないtenant・価格・quota・live品質・内部routingはUNKNOWNに残す。

## Confirmed

## UNKNOWN
- PM-CLAUDE-CORE-04-CURRENTNESS: 原資料の確定版から収録しています。外部根拠の最新性を今回の取込日で確認済みと扱いません。本文の未解決事項・残余制約を保持します。

## Evidence
- EVD-PM-CLAUDE-CORE-04-URL-9332683246DB (RESEARCH): Anthropic model IDs and versioning: <[ — https://platform.claude.com/docs/en/about-claude/models/model-ids-and-versions
- EVD-PM-CLAUDE-CORE-04-URL-82747C62F33E (RESEARCH): Anthropic tool use overview/how tool use works: <[ — https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview
- EVD-PM-CLAUDE-CORE-04-URL-A2C7CB24DC72 (RESEARCH): Anthropic tool use overview/how tool use works: <[https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview](https://platform.claude.com/docs/en/agents-and-tools/tool- — https://platform.claude.com/docs/en/agents-and-tools/tool-use/how-tool-use-works
- EVD-PM-CLAUDE-CORE-04-URL-DBCB362197DB (RESEARCH): Anthropic Message Batches: <[ — https://platform.claude.com/docs/en/build-with-claude/batch-processing
- EVD-PM-CLAUDE-CORE-04-URL-DF1A1BB212DB (RESEARCH): Anthropic service tiers: <[ — https://platform.claude.com/docs/en/api/service-tiers
- EVD-PM-CLAUDE-CORE-04-URL-854B88AAD19B (RESEARCH): Anthropic API errors/rate limits: <[ — https://platform.claude.com/docs/en/api/errors
- EVD-PM-CLAUDE-CORE-04-URL-ED2EA3ED3DC1 (RESEARCH): Anthropic API errors/rate limits: <[https://platform.claude.com/docs/en/api/errors](https://platform.claude.com/docs/en/api/errors)> ; <[ — https://platform.claude.com/docs/en/api/rate-limits
- EVD-PM-CLAUDE-CORE-04-URL-2AA41A0F518B (RESEARCH): Claude Agent SDK Python/TypeScript references: <[ — https://code.claude.com/docs/en/agent-sdk/python
- EVD-PM-CLAUDE-CORE-04-URL-2B13AE4354AB (RESEARCH): Claude Agent SDK Python/TypeScript references: <[https://code.claude.com/docs/en/agent-sdk/python](https://code.claude.com/docs/en/agent-sdk/python)> ; <[ — https://code.claude.com/docs/en/agent-sdk/typescript
- EVD-PM-CLAUDE-CORE-04-URL-8A6DFE25191D (RESEARCH): Claude Agent SDK permissions/observability: <[ — https://code.claude.com/docs/en/agent-sdk/permissions
- EVD-PM-CLAUDE-CORE-04-URL-76592DD9412F (RESEARCH): Claude Agent SDK permissions/observability: <[https://code.claude.com/docs/en/agent-sdk/permissions](https://code.claude.com/docs/en/agent-sdk/permissions)> ; <[ — https://code.claude.com/docs/en/agent-sdk/observability
- PM-CLAUDE-CORE-04-DOCUMENT-READBACK (IMPL): 原資料との収録照合

## Practice
PM-CLAUDE-CORE-04は公開仕様、固定SDK release anchor、既存provider-neutral反証の範囲でbounded closeとする。Claude固有機能をCoreへ昇格せず、既存local fixtureをClaude実測へ付け替えない。fallback trigger、live model/tier quality、quota/price、permission/source全経路、cross-surface state、retry/idempotency、telemetry retention/SLO、organization loss budgetはUNKNOWNとして残す。Core claimは変更していないためGemini/GPT Product Mappingの遡及監査は不要である。
## 資料の識別情報
Source SHA-256 c7a7f2451670a3181fceb778923909693cef097f921bc7a6730abf33d602232f

## Product Mapping — Claude / CORE-04
PRODUCT_MAPPING_ID: PM-CLAUDE-CORE-04-v0.1

PROVIDER: CLAUDE / ANTHROPIC

CORE_PACKET: PKT-GEMINI-A-CORE-04-v0.1

AS_OF: 2026-09-09 21:00 JST

STATUS: PUBLIC_SURFACES_MAPPED / CORE_EVIDENCE_REUSED / LIVE_TENANT_VALUES_UNKNOWN

LAYER: PRODUCT_MAPPING (not Enterprise AI Core; not Organization Overlay)

## Scope
CORE-04のmodel / tool / source / state routing、fallback、batch、observability契約をClaude API、Message Batches、service tier、Claude Agent SDKへ対応づける。Claude.ai、Claude API、Agent SDK、Claude Code CLI、Bedrock、Google Cloud、Managed Agentsを同一routeとせず、公開仕様から閉じないtenant・価格・quota・live品質・内部routingはUNKNOWNに残す。

## Core → Claude mapping rows
1. ROW: PM-C04-CLAUDE-001; Core contract: model route identity; Claude / Anthropic surface and current evidence: Claude APIの4.6世代以降のdateless model IDは固定snapshotで、旧世代の短縮aliasとは意味が異なる。BedrockとGoogle CloudではID表記も異なる。model weightsが固定でもrequest router、安全分類器、sampling等のserving infrastructureは変わり得る; Mapping decision: requested_model / resolved_model_id / platform / API_version / beta_headers / SDK_version / serving_cohortを別記録し、文字列がdatelessという理由だけでalias扱いしない; Required verification / UNKNOWN: tenant availability、実際のserving cohort、infrastructure change、cross-platform equivalenceはUNKNOWN

2. ROW: PM-C04-CLAUDE-002; Core contract: SDK primary/fallback route; Claude / Anthropic surface and current evidence: Agent SDK Pythonはmodelとfallback_modelを別optionとして公開し、TypeScriptはsession中のmodel override/解除を持つ。固定release anchorはPython v0.2.152／commit a8b1e28／bundled CLI 2.1.259; Mapping decision: primaryとfallback、session中の切替、project/user defaultへの復帰を別route eventにする。fallback採用前にtool、context、output schema、risk、stateのlossを検査する; Required verification / UNKNOWN: 「primary fails」の判定条件、retry順、全optionの移送、CLI/model rollout、live fallback品質は公開referenceだけでは閉じない

3. ROW: PM-C04-CLAUDE-003; Core contract: tool proposal / execution / authorization; Claude / Anthropic surface and current evidence: Claude APIはclient toolとserver toolを分ける。client toolはClaudeがstructured tool_useを返しapplicationが実行、server toolはAnthropic側で実行する。Agent SDKはagent loop内でtoolを実行するが、PreToolUse hook等で介入できる; Mapping decision: tool_declared / selected / execution_location / permission_decision / executed / tool_use_id / tool_result / external_receipt / readback / acceptedを分離する。model選択やtool_use blockを業務認可・commit証拠にしない; Required verification / UNKNOWN: server tool内部dispatch、client側IAM、外部transaction、同一target並列安全性はUNKNOWN

4. ROW: PM-C04-CLAUDE-004; Core contract: permission-aware tool route; Claude / Anthropic surface and current evidence: Agent SDKのpermission評価はhook、deny、ask、mode、allow、canUseToolの順序を持つ。allow相当の設定によりcallbackがconsultされない場合があり、全callをgateする用途はPreToolUse hookへ送られる; Mapping decision: route descriptorへpermission mode、allow/deny rule digest、hook version、callback consultation、target digestを含める。permission自動許可とsource object ACL・業務権限を分離する; Required verification / UNKNOWN: organization settings、connector/source ACL、runtime policy merge、管理者制御の実効値はOrganization OverlayでUNKNOWN

5. ROW: PM-C04-CLAUDE-005; Core contract: service-capacity route; Claude / Anthropic surface and current evidence: APIのservice_tier=autoは利用可能ならPriorityを使い、capacity不足時は他capacityへ進む。response usageは実際のstandard / priority / batchを返す。Standardはbest-effort、既存Priority契約の範囲もmodel/capacity依存; Mapping decision: requested tierとobserved tierを分離し、tier fallbackをmodel fallbackや品質同値とみなさない。request/response、rate-limit header、latency、error、acceptanceをjoinする; Required verification / UNKNOWN: 契約capacity、価格、quota、SLO、overload時挙動、organizationの許容tierはlive/contract依存

6. ROW: PM-C04-CLAUDE-006; Core contract: online → batch fallback; Claude / Anthropic surface and current evidence: Message Batchesは非同期・独立requestで、結果順は保証されずcustom_id結合が必要。各結果はsucceeded/errored/canceled/expired、他requestの失敗から独立し、最大24時間処理・作成後29日結果保持・Workspace isolationを持つ; Mapping decision: onlineとBatchを別routeとし、deadline、streaming、interaction、ordering、partial result、retention、Workspace、receiptのlossを事前計算する。batch終了を業務完了にしない; Required verification / UNKNOWN: 実completion time、cancel race、tenant retention/ZDR、再送dedupe、業務deadlineはUNKNOWN

7. ROW: PM-C04-CLAUDE-007; Core contract: retry / failure envelope; Claude / Anthropic surface and current evidence: API errorはtype/message/request_idを返し、overloaded_error等を区別する。rate-limit headerはlimit/remaining/resetを返す。Coreで採用済みのLiteLLM循環fallback事故はClaude固有欠陥ではないが、retry amplificationを防ぐ反証意図として再利用する; Mapping decision: request_id / attempt / error_type / error_hash / retry_after / response_started / visited_route / wall_clock / capability_lossを記録し、同一error循環・budget超過・応答開始後のblind replayを停止する; Required verification / UNKNOWN: SDK/APIの実retry policy、network timeout、stream中断、idempotency、cross-model/provider fallbackは固定live試験なしではUNKNOWN

8. ROW: PM-C04-CLAUDE-008; Core contract: route observability / final acceptance; Claude / Anthropic surface and current evidence: Agent SDKのOTelはCLI child processがmodel request/tool execution span、token/cost metric、prompt/tool-result logをexportし、SDK自身はtelemetryを生成しない。Agent loopのResultMessageはfinal text、usage、cost、session IDを持つ; Mapping decision: CLI/SDK/model/config/collectorを同じrelease recordへ固定し、route decision、provider/tool result、artifact/external readback、verifier、final acceptanceを別eventにする。trace欠落時はapplication ledgerを要求する; Required verification / UNKNOWN: telemetry enablement、sampling、redaction、delivery/retention、cost完全性、business acceptanceはOrganization OverlayでUNKNOWN

## Fallback loss profile
| FROM | TO | MUST COMPARE BEFORE ROUTE | STOP CONDITION |
| --- | --- | --- | --- |
| fixed Claude model ID | older short alias / another model | snapshot semantics、platform ID、tools、context、output contract、risk | approved identityまたはrequired capabilityを失う |
| API direct | Agent SDK / Claude Code CLI | loop、permissions、config loading、session、tool execution、telemetry owner | effective configまたはreceipt chainを再構成できない |
| primary model | fallback_model | trigger、attempt、model capability、tool/schema/state、loss budget | failure条件不明、cycle、required capability loss |
| Priority | Standard | observed service tier、capacity、latency/SLO、deadline、cost | contractual/SLO条件またはdeadlineを満たさない |
| online Messages | Message Batches | interactivity、ordering、partial/expiry、Workspace、29-day result window | interactive tool/state loopまたは受入deadlineを失う |
| source/tool-bound | model-only | source authority、ACL、citation、freshness、tool receipt | source/tool requirementまたはloss budget超過 |

## Safe verification profile
許可された非業務環境と非機密synthetic fixtureだけで実施する。今回はClaude API、organization、model、Agent SDK process、tool、source、external systemを操作していない。

- fixed ID、旧alias、platform-specific ID、SDK default/override/fallbackをroute recordで区別する。

- primaryを制御されたretryable failureへ送り、attempt、fallback trigger、visited route、capability loss、ResultMessageを記録する。

- client read-only toolとserver toolを分け、proposal、permission、execution location、tool result、readbackを照合する。

- permission rule/hook/callbackの組合せで、どのgateが実際にconsultされたかを検査する。

- service_tier=autoのrequested/observed tierを記録し、tier変化をmodel変化と混同しない。

- onlineと小規模Batchへ同じsynthetic operation IDを送り、unordered/partial/expired/canceled joinを検査する。

7.同一error hash、attempt/wall-clock budget、response-startedを用いてretry/fallback loopを停止する。

- OTel traceとapplication ledger、artifact/external readback、final acceptanceを別々に突合する。

既存CORE-04のprovider-neutral local fixtures（10/10、16/16）はCOMPILER_EXPERIMENTとして再利用するが、Claude/Anthropicのlive性能・可用性・fallback挙動の実測には再分類しない。

## Evidence and provenance
- Anthropic model IDs and versioning: <[https://platform.claude.com/docs/en/about-claude/models/model-ids-and-versions](https://platform.claude.com/docs/en/about-claude/models/model-ids-and-versions)> (accessed 2026-09-09)

- Anthropic tool use overview/how tool use works: <[https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview](https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview)> ; <[https://platform.claude.com/docs/en/agents-and-tools/tool-use/how-tool-use-works](https://platform.claude.com/docs/en/agents-and-tools/tool-use/how-tool-use-works)> (accessed 2026-09-09)

- Anthropic Message Batches: <[https://platform.claude.com/docs/en/build-with-claude/batch-processing](https://platform.claude.com/docs/en/build-with-claude/batch-processing)> (accessed 2026-09-09)

- Anthropic service tiers: <[https://platform.claude.com/docs/en/api/service-tiers](https://platform.claude.com/docs/en/api/service-tiers)> (accessed 2026-09-09)

- Anthropic API errors/rate limits: <[https://platform.claude.com/docs/en/api/errors](https://platform.claude.com/docs/en/api/errors)> ; <[https://platform.claude.com/docs/en/api/rate-limits](https://platform.claude.com/docs/en/api/rate-limits)> (accessed 2026-09-09)

- Claude Agent SDK Python/TypeScript references: <[https://code.claude.com/docs/en/agent-sdk/python](https://code.claude.com/docs/en/agent-sdk/python)> ; <[https://code.claude.com/docs/en/agent-sdk/typescript](https://code.claude.com/docs/en/agent-sdk/typescript)> (accessed 2026-09-09)

- Claude Agent SDK permissions/observability: <[https://code.claude.com/docs/en/agent-sdk/permissions](https://code.claude.com/docs/en/agent-sdk/permissions)> ; <[https://code.claude.com/docs/en/agent-sdk/observability](https://code.claude.com/docs/en/agent-sdk/observability)> (accessed 2026-09-09)

- fixed release anchor reused: EXT-C01-032

- provider-neutral research/incident evidence remains in CORE-04; no Claude-specific defect is inferred from it.

## Boundary decision
PM-CLAUDE-CORE-04は公開仕様、固定SDK release anchor、既存provider-neutral反証の範囲でbounded closeとする。Claude固有機能をCoreへ昇格せず、既存local fixtureをClaude実測へ付け替えない。fallback trigger、live model/tier quality、quota/price、permission/source全経路、cross-surface state、retry/idempotency、telemetry retention/SLO、organization loss budgetはUNKNOWNとして残す。Core claimは変更していないためGemini/GPT Product Mappingの遡及監査は不要である。

NEXT: /agent/start.json
STOP: This is a knowledge projection, not an execution order.