No English page yet. Showing the Japanese source.
CLAUDE:状態を持つAIの運用契約
記事状態: 公開 · 最終確認: 未確認
要点
This mapping binds Anthropic Messages, Message Batches and Claude Agent SDK session, resume, fork, hook and file-checkpoint surfaces to CORE-01. Provider dialogue/session state is not the canonical business task record. A successful message, ended batch, result message, hook event or file rewind is not artifact acceptance or proof that an external system reached the intended final state.
現在の判断
This mapping binds Anthropic Messages, Message Batches and Claude Agent SDK session, resume, fork, hook and file-checkpoint surfaces to CORE-01. Provider dialogue/session state is not the canonical business task record. A successful message, ended batch, result message, hook event or file rewind is not artifact acceptance or proof that an external system reached the intended final state.
どこまで当てはまるか
This mapping binds Anthropic Messages, Message Batches and Claude Agent SDK session, resume, fork, hook and file-checkpoint surfaces to CORE-01. Provider dialogue/session state is not the canonical business task record. A successful message, ended batch, result message, hook event or file rewind is not artifact acceptance or proof that an external system reached the intended final state.
No Anthropic API, organization, workspace, model, Agent SDK process, tool, file checkpoint or external system was operated. Existing CORE-01 6/6 and 12/12 local results remain first-party COMPILER_EXPERIMENT evidence and are not Claude observations.
詳しく理解する
Product Mapping — Claude / CORE-01 Operating Contracts for Stateful AI
MAPPING_ID: PM-CLAUDE-CORE-01-v0.1
PRODUCT_MAPPING_ID: PM-CLAUDE-CORE-01-v0.1
PROVIDER: Claude / Anthropic
CORE_PACKET: PKT-GEMINI-A-CORE-01-v0.1
STATUS: PUBLIC_MAPPING_BOUNDED_COMPLETE / FIXED_V0.2.152_STATE_BOUNDARY / LIVE_ORG_PENDING
AS_OF: 2026-09-09
1. Boundary
This mapping binds Anthropic Messages, Message Batches and Claude Agent SDK session, resume, fork, hook and file-checkpoint surfaces to CORE-01. Provider dialogue/session state is not the canonical business task record. A successful message, ended batch, result message, hook event or file rewind is not artifact acceptance or proof that an external system reached the intended final state.
No Anthropic API, organization, workspace, model, Agent SDK process, tool, file checkpoint or external system was operated. Existing CORE-01 6/6 and 12/12 local results remain first-party COMPILER_EXPERIMENT evidence and are not Claude observations.
2. Claim mapping
- MAPPING CLAIM: PM-C01-CLAUDE-001; CLAUDE / ANTHROPIC SURFACE: Messages API; CORE-01 DECISION: The API is stateless and the client sends conversation history. Preserve TASK_ID, contract version, source snapshot, work/acceptance state and artifact identity outside the message list.; EVIDENCE / LIMIT: EXT-C01-027; synthetic assistant messages may be supplied, so history is not origin/provenance proof.
- MAPPING CLAIM: PM-C01-CLAUDE-002; CLAUDE / ANTHROPIC SURFACE: Message ID, stop reason and message content; CORE-01 DECISION: Treat provider response identity and model termination as runtime evidence only. Require artifact or authoritative external-state readback before business acceptance.; EVIDENCE / LIMIT: EXT-C01-027; public response shape does not encode the complete Operating Contract.
- MAPPING CLAIM: PM-C01-CLAUDE-003; CLAUDE / ANTHROPIC SURFACE: Message Batches; CORE-01 DECISION: Track batch ID and per-request custom_id separately from task/artifact IDs. Join results by custom_id, not position, and retain every per-request result class.; EVIDENCE / LIMIT: EXT-C01-028; results may be out of order and succeeded/errored/canceled/expired; ended may include partial results.
- MAPPING CLAIM: PM-C01-CLAUDE-004; CLAUDE / ANTHROPIC SURFACE: Agent SDK continue/resume; CORE-01 DECISION: continue selects the most recent session in a directory while resume uses a tracked session ID. Neither selection rule establishes exclusive task ownership or source freshness.; EVIDENCE / LIMIT: EXT-C01-029; cross-host resume needs transcript movement or a SessionStore; application state is recommended as a robust alternative.
- MAPPING CLAIM: PM-C01-CLAUDE-005; CLAUDE / ANTHROPIC SURFACE: Agent SDK fork; CORE-01 DECISION: Fork creates a new session ID with copied prior history while leaving the original history unchanged. Record parent session, fork point, source snapshot and decision lineage explicitly.; EVIDENCE / LIMIT: EXT-C01-029; separate session histories do not themselves provide merge, acceptance or conflict resolution.
- MAPPING CLAIM: PM-C01-CLAUDE-006; CLAUDE / ANTHROPIC SURFACE: Result/session ID; CORE-01 DECISION: A result message exposes a session ID on success or error, but connection/process failures may yield no result message. Persist the canonical task state before relying on the provider identifier.; EVIDENCE / LIMIT: EXT-C01-029; session availability and transcript storage are deployment dependent.
- MAPPING CLAIM: PM-C01-CLAUDE-007; CLAUDE / ANTHROPIC SURFACE: File checkpointing and rewind; CORE-01 DECISION: A checkpoint is a scoped local-file recovery carrier, not general rollback. Record covered tool/path set and skipped paths, and separately compensate external effects.; EVIDENCE / LIMIT: EXT-C01-030; Bash, most subagent edits, directory operations and remote/network files are outside coverage; rewind does not rewind conversation.
- MAPPING CLAIM: PM-C01-CLAUDE-008; CLAUDE / ANTHROPIC SURFACE: Agent SDK hooks; CORE-01 DECISION: Pre/post/session hook events can enforce or observe parts of a trajectory, but hook completion/logging is not an authoritative operation receipt or final-state readback.; EVIDENCE / LIMIT: EXT-C01-031; callback timeout can discard output and permit the session to continue depending on event; Python/TypeScript event support differs.
- MAPPING CLAIM: PM-C01-CLAUDE-009; CLAUDE / ANTHROPIC SURFACE: Python Agent SDK release and issue history; CORE-01 DECISION: Pin package and bundled CLI together. Historical capability gaps remain regression intent and are not promoted to a current defect without reproduction.; EVIDENCE / LIMIT: EXT-C01-032/033; v0.2.152 / commit a8b1e28, bundled CLI 2.1.259; issue #180 is closed and bounded.
確認・評価方法
Use synthetic non-personal fixtures in a disposable local directory. Do not enable production tools or external writes.
- Assign distinct task, contract, source snapshot, message, batch, custom_id, session, parent/fork, checkpoint and artifact IDs.
- Reconstruct a stateless Messages turn from the canonical task record and prove that synthetic assistant history cannot be mistaken for provenance.
- Submit a dry-run-equivalent batch fixture containing success, error, cancel and expiry records in shuffled order; join only by custom_id and reject completion while any required result is absent.
- Resume a captured session ID and separately fork it. Verify exclusive ownership, distinct fork ID, unchanged parent history and explicit parent/fork lineage.
- Simulate a missing result message or unavailable transcript and recover from application state rather than guessing the last completed work.
- Checkpoint edits made through tracked file tools, then include Bash-created content, a directory operation and a remote-effect fixture. Rewind must only clear covered file changes; uncovered/external effects stay open for compensation and readback.
- Record hook event, intended decision, actual tool result and authoritative readback as separate evidence. A hook timeout or SessionEnd event cannot set BUSINESS_ACCEPTED.
- Accept only after artifact existence, digest and declared external final state pass the CORE-01 acceptance contract.
This profile transfers the intent of the existing CORE-01 fixtures. It is not a record of executing Claude or the Agent SDK.
4. Organization Overlay routing
The Organization Overlay must supply the adopted API/model/SDK and bundled CLI pins; message/batch retention and workspace permissions; session directory or SessionStore, ownership, locking, encryption, cleanup and cross-host recovery; hook policy and timeout handling; checkpoint scope and backup retention; authoritative artifact/external system; batch SLA; RTO/RPO; retry/compensation owner; and acceptance authority. Public documentation cannot close these values.
5. Bounded questions
| QUESTION_ID | QUESTION | STATUS |
| Q-PM-C01-CLAUDE-001 | Which message, batch, session and task identifiers and retention policies are actually adopted? | BOUNDED_PUBLIC_MAPPING / DEPLOYMENT_CONFIGURATION_UNKNOWN |
| Q-PM-C01-CLAUDE-002 | How are shuffled, partial, canceled and expired batch results joined to artifact/external-state readback? | BOUNDED_SAFE_PROFILE / LIVE_ACCEPTANCE_UNKNOWN |
| Q-PM-C01-CLAUDE-003 | Which SessionStore, lock, parent/fork lineage, checkpoint coverage and failure-recovery rules are used? | BOUNDED_FIXED_CODE / DEPLOYMENT_CONCURRENCY_UNKNOWN |
| Q-PM-C01-CLAUDE-004 | What are the organization retention, RTO/RPO, compensation, acceptance and incident-owner values? | ORGANIZATION_OVERLAY_ONLY / UNKNOWN |
6. Disposition
- CORE-01 claims remain unchanged; Gemini and GPT backprop are not required.
- AUD-C01-001 is open for live adapter execution and AUD-C01-007 for cross-provider falsification of the provisional requiredness profile.
- Public evidence is bounded to 2026-09-09 and Python Agent SDK v0.2.152 / release commit a8b1e28 with bundled Claude CLI 2.1.259.
- No auxiliary evidence ID was promoted because the communication folder contained no newer evidence gap for this unit.
- Next provider unit: Claude CORE-02 against the existing permission-aware retrieval, memory lifecycle and bounded long-context test intent.
実務での使い方
MAPPING_ID: PM-CLAUDE-CORE-01-v0.1
PRODUCT_MAPPING_ID: PM-CLAUDE-CORE-01-v0.1
PROVIDER: Claude / Anthropic
CORE_PACKET: PKT-GEMINI-A-CORE-01-v0.1
STATUS: PUBLIC_MAPPING_BOUNDED_COMPLETE / FIXED_V0.2.152_STATE_BOUNDARY / LIVE_ORG_PENDING
AS_OF: 2026-09-09
まだ確認できていないこと
- PM-CLAUDE-CORE-01-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。
根拠資料
この記事には、直接開ける外部出典が登録されていません。出典IDや検証情報は末尾で確認できます。
出典ID・検証情報・機械可読データ
根拠の種類と情報の鮮度は、ラベルとアイコンで確認できます。
- 記事ID
- PM-CLAUDE-CORE-01
- 最終確認日
- —
- 提供元・モデル・プラン・利用環境
- CLAUDE
- 対象地域
- —
元資料の全文と収録照合記録
資料の識別情報
Source SHA-256 63a3ab32b94ddafcbe991110306bced90ad3e5d6f069f496ac06bb1f7c61d923
Product Mapping — Claude / CORE-01 Operating Contracts for Stateful AI
MAPPING_ID: PM-CLAUDE-CORE-01-v0.1
PRODUCT_MAPPING_ID: PM-CLAUDE-CORE-01-v0.1
PROVIDER: Claude / Anthropic
CORE_PACKET: PKT-GEMINI-A-CORE-01-v0.1
STATUS: PUBLIC_MAPPING_BOUNDED_COMPLETE / FIXED_V0.2.152_STATE_BOUNDARY / LIVE_ORG_PENDING
AS_OF: 2026-09-09
1. Boundary
This mapping binds Anthropic Messages, Message Batches and Claude Agent SDK session, resume, fork, hook and file-checkpoint surfaces to CORE-01. Provider dialogue/session state is not the canonical business task record. A successful message, ended batch, result message, hook event or file rewind is not artifact acceptance or proof that an external system reached the intended final state.
No Anthropic API, organization, workspace, model, Agent SDK process, tool, file checkpoint or external system was operated. Existing CORE-01 6/6 and 12/12 local results remain first-party COMPILER_EXPERIMENT evidence and are not Claude observations.
2. Claim mapping
- MAPPING CLAIM: PM-C01-CLAUDE-001; CLAUDE / ANTHROPIC SURFACE: Messages API; CORE-01 DECISION: The API is stateless and the client sends conversation history. Preserve TASK_ID, contract version, source snapshot, work/acceptance state and artifact identity outside the message list.; EVIDENCE / LIMIT: EXT-C01-027; synthetic assistant messages may be supplied, so history is not origin/provenance proof.
- MAPPING CLAIM: PM-C01-CLAUDE-002; CLAUDE / ANTHROPIC SURFACE: Message ID, stop reason and message content; CORE-01 DECISION: Treat provider response identity and model termination as runtime evidence only. Require artifact or authoritative external-state readback before business acceptance.; EVIDENCE / LIMIT: EXT-C01-027; public response shape does not encode the complete Operating Contract.
- MAPPING CLAIM: PM-C01-CLAUDE-003; CLAUDE / ANTHROPIC SURFACE: Message Batches; CORE-01 DECISION: Track batch ID and per-request custom_id separately from task/artifact IDs. Join results by custom_id, not position, and retain every per-request result class.; EVIDENCE / LIMIT: EXT-C01-028; results may be out of order and succeeded/errored/canceled/expired; ended may include partial results.
- MAPPING CLAIM: PM-C01-CLAUDE-004; CLAUDE / ANTHROPIC SURFACE: Agent SDK continue/resume; CORE-01 DECISION: continue selects the most recent session in a directory while resume uses a tracked session ID. Neither selection rule establishes exclusive task ownership or source freshness.; EVIDENCE / LIMIT: EXT-C01-029; cross-host resume needs transcript movement or a SessionStore; application state is recommended as a robust alternative.
- MAPPING CLAIM: PM-C01-CLAUDE-005; CLAUDE / ANTHROPIC SURFACE: Agent SDK fork; CORE-01 DECISION: Fork creates a new session ID with copied prior history while leaving the original history unchanged. Record parent session, fork point, source snapshot and decision lineage explicitly.; EVIDENCE / LIMIT: EXT-C01-029; separate session histories do not themselves provide merge, acceptance or conflict resolution.
- MAPPING CLAIM: PM-C01-CLAUDE-006; CLAUDE / ANTHROPIC SURFACE: Result/session ID; CORE-01 DECISION: A result message exposes a session ID on success or error, but connection/process failures may yield no result message. Persist the canonical task state before relying on the provider identifier.; EVIDENCE / LIMIT: EXT-C01-029; session availability and transcript storage are deployment dependent.
- MAPPING CLAIM: PM-C01-CLAUDE-007; CLAUDE / ANTHROPIC SURFACE: File checkpointing and rewind; CORE-01 DECISION: A checkpoint is a scoped local-file recovery carrier, not general rollback. Record covered tool/path set and skipped paths, and separately compensate external effects.; EVIDENCE / LIMIT: EXT-C01-030; Bash, most subagent edits, directory operations and remote/network files are outside coverage; rewind does not rewind conversation.
- MAPPING CLAIM: PM-C01-CLAUDE-008; CLAUDE / ANTHROPIC SURFACE: Agent SDK hooks; CORE-01 DECISION: Pre/post/session hook events can enforce or observe parts of a trajectory, but hook completion/logging is not an authoritative operation receipt or final-state readback.; EVIDENCE / LIMIT: EXT-C01-031; callback timeout can discard output and permit the session to continue depending on event; Python/TypeScript event support differs.
- MAPPING CLAIM: PM-C01-CLAUDE-009; CLAUDE / ANTHROPIC SURFACE: Python Agent SDK release and issue history; CORE-01 DECISION: Pin package and bundled CLI together. Historical capability gaps remain regression intent and are not promoted to a current defect without reproduction.; EVIDENCE / LIMIT: EXT-C01-032/033; v0.2.152 / commit a8b1e28, bundled CLI 2.1.259; issue #180 is closed and bounded.
3. Safe test profile
Use synthetic non-personal fixtures in a disposable local directory. Do not enable production tools or external writes.
- Assign distinct task, contract, source snapshot, message, batch, custom_id, session, parent/fork, checkpoint and artifact IDs.
- Reconstruct a stateless Messages turn from the canonical task record and prove that synthetic assistant history cannot be mistaken for provenance.
- Submit a dry-run-equivalent batch fixture containing success, error, cancel and expiry records in shuffled order; join only by custom_id and reject completion while any required result is absent.
- Resume a captured session ID and separately fork it. Verify exclusive ownership, distinct fork ID, unchanged parent history and explicit parent/fork lineage.
- Simulate a missing result message or unavailable transcript and recover from application state rather than guessing the last completed work.
- Checkpoint edits made through tracked file tools, then include Bash-created content, a directory operation and a remote-effect fixture. Rewind must only clear covered file changes; uncovered/external effects stay open for compensation and readback.
- Record hook event, intended decision, actual tool result and authoritative readback as separate evidence. A hook timeout or SessionEnd event cannot set BUSINESS_ACCEPTED.
- Accept only after artifact existence, digest and declared external final state pass the CORE-01 acceptance contract.
This profile transfers the intent of the existing CORE-01 fixtures. It is not a record of executing Claude or the Agent SDK.
4. Organization Overlay routing
The Organization Overlay must supply the adopted API/model/SDK and bundled CLI pins; message/batch retention and workspace permissions; session directory or SessionStore, ownership, locking, encryption, cleanup and cross-host recovery; hook policy and timeout handling; checkpoint scope and backup retention; authoritative artifact/external system; batch SLA; RTO/RPO; retry/compensation owner; and acceptance authority. Public documentation cannot close these values.
5. Bounded questions
| QUESTION_ID | QUESTION | STATUS |
| Q-PM-C01-CLAUDE-001 | Which message, batch, session and task identifiers and retention policies are actually adopted? | BOUNDED_PUBLIC_MAPPING / DEPLOYMENT_CONFIGURATION_UNKNOWN |
| Q-PM-C01-CLAUDE-002 | How are shuffled, partial, canceled and expired batch results joined to artifact/external-state readback? | BOUNDED_SAFE_PROFILE / LIVE_ACCEPTANCE_UNKNOWN |
| Q-PM-C01-CLAUDE-003 | Which SessionStore, lock, parent/fork lineage, checkpoint coverage and failure-recovery rules are used? | BOUNDED_FIXED_CODE / DEPLOYMENT_CONCURRENCY_UNKNOWN |
| Q-PM-C01-CLAUDE-004 | What are the organization retention, RTO/RPO, compensation, acceptance and incident-owner values? | ORGANIZATION_OVERLAY_ONLY / UNKNOWN |
6. Disposition
- CORE-01 claims remain unchanged; Gemini and GPT backprop are not required.
- AUD-C01-001 is open for live adapter execution and AUD-C01-007 for cross-provider falsification of the provisional requiredness profile.
- Public evidence is bounded to 2026-09-09 and Python Agent SDK v0.2.152 / release commit a8b1e28 with bundled Claude CLI 2.1.259.
- No auxiliary evidence ID was promoted because the communication folder contained no newer evidence gap for this unit.
- Next provider unit: Claude CORE-02 against the existing permission-aware retrieval, memory lifecycle and bounded long-context test intent.
PM-CLAUDE-CORE-01-DOCUMENT-READBACK実装資料実装資料根拠の出所: DOCUMENT_INGESTION_READBACK原資料との収録照合
- 裏付ける内容
- 資料単位の本文ハッシュ、段落数、表数を照合しています。
- 確認日
- —
- 出典・実験ID
- DOC-083cde4caa627644-PM-CLAUDE-CORE-01
- 検証箇所
- 18WPldGemks2px0ukhKYINeBMrBwAmlAi · 083cde4caa627644036a2225f1f6ff29b8015bd902614fe1f7a00a997b4eb59c · provider_mappings/PM-CLAUDE-CORE-01.md · c7b48646742473769deefbfc43e08b0cb9028a3bed50ce458cf9cfcd22299c82
- 適用範囲
- 公開指定を受けた資料単位の全文・表・出典。
- 出典・来歴
- SOURCE_IDENTITY_VERIFIED
- 限界
- 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。
- 出典ID
- DOC-083cde4caa627644-PM-CLAUDE-CORE-01
- 追跡中のUNKNOWN ID
- PM-CLAUDE-CORE-01-CURRENTNESS
- 再確認のきっかけ
- 原資料の更新 · 参照資料の変更 · 新しい反証・検証記録
機械検査: 通過 — Automated validation does not guarantee accuracy or authorize execution.
同じ記事IDの機械可読データです。