# OpenAI / GPT：状態を持つAIの運用契約

CURRENT_STATUS: UNKNOWN
OBSERVATION: UNOBSERVED
FRESHNESS: UNKNOWN
PAGE_TYPE: reference
CANONICAL_URL: /providers/pm-gpt-core-01
CLAIM_IDS: none
EVIDENCE_IDS: PM-GPT-CORE-01-DOCUMENT-READBACK
UNKNOWN_IDS: PM-GPT-CORE-01-CURRENTNESS
OPS_IDS: none

## Summary
This mapping binds the provider's response, conversation, background execution, webhook, Agents SDK session, result, and RunState surfaces to CORE-01. It does not treat provider persistence as the authoritative business task record, model output as artifact acceptance, or a provider terminal state as proof of an external commit.

## Current judgment
This mapping binds the provider's response, conversation, background execution, webhook, Agents SDK session, result, and RunState surfaces to CORE-01. It does not treat provider persistence as the authoritative business task record, model output as artifact acceptance, or a provider terminal state as proof of an external commit.

## Scope
This mapping binds the provider's response, conversation, background execution, webhook, Agents SDK session, result, and RunState surfaces to CORE-01. It does not treat provider persistence as the authoritative business task record, model output as artifact acceptance, or a provider terminal state as proof of an external commit.

No OpenAI API, tenant, model, webhook endpoint, session backend, tool, or external system was operated in this unit. Existing CORE-01 local tests remain COMPILER_EXPERIMENT; they are not GPT observations.

## Confirmed

## UNKNOWN
- PM-GPT-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。

## Evidence
- PM-GPT-CORE-01-DOCUMENT-READBACK (IMPL): 原資料との収録照合

## Practice
MAPPING_ID: PM-GPT-CORE-01-v0.1

PRODUCT_MAPPING_ID: PM-GPT-CORE-01-v0.1

PROVIDER: GPT / OpenAI

CORE_PACKET: PKT-GEMINI-A-CORE-01-v0.1

STATUS: PUBLIC_MAPPING_BOUNDED_COMPLETE / FIXED_V0.22_STATE_BOUNDARY / LIVE_ORG_PENDING

AS_OF: 2026-09-09
## 資料の識別情報
Source SHA-256 af28ba220bac3f73193a750b02e4e2a0ee47abfa985f88c6281813f1812599ff

## Product Mapping — GPT / CORE-01 Operating Contracts for Stateful AI
MAPPING_ID: PM-GPT-CORE-01-v0.1

PRODUCT_MAPPING_ID: PM-GPT-CORE-01-v0.1

PROVIDER: GPT / OpenAI

CORE_PACKET: PKT-GEMINI-A-CORE-01-v0.1

STATUS: PUBLIC_MAPPING_BOUNDED_COMPLETE / FIXED_V0.22_STATE_BOUNDARY / LIVE_ORG_PENDING

AS_OF: 2026-09-09

## 1. Boundary
This mapping binds the provider's response, conversation, background execution, webhook, Agents SDK session, result, and RunState surfaces to CORE-01. It does not treat provider persistence as the authoritative business task record, model output as artifact acceptance, or a provider terminal state as proof of an external commit.

No OpenAI API, tenant, model, webhook endpoint, session backend, tool, or external system was operated in this unit. Existing CORE-01 local tests remain COMPILER_EXPERIMENT; they are not GPT observations.

## 2. Claim mapping
1. MAPPING CLAIM: PM-C01-GPT-001; GPT / OPENAI SURFACE: Response, Conversation, Session, RunState; CORE-01 DECISION: Keep provider execution and dialogue identities separate from TASK_ID, CONTRACT_VERSION, artifact identity and business acceptance.; EVIDENCE / LIMIT: EXT-C01-021/024/025; provider state surfaces are partial carriers, not the complete Operating Contract.

2. MAPPING CLAIM: PM-C01-GPT-002; GPT / OPENAI SURFACE: Conversations API and previous_response_id; CORE-01 DECISION: A durable conversation identifier and item lineage can support continuity, but source snapshot, authority, permitted scope, completed/not-done and acceptance remain explicit contract fields.; EVIDENCE / LIMIT: EXT-C01-021; conversation item persistence is not source-of-truth proof.

3. MAPPING CLAIM: PM-C01-GPT-003; GPT / OPENAI SURFACE: Response storage, Conversation retention, background temporary storage; CORE-01 DECISION: Record the selected carrier, store policy, retention class and deletion/readback evidence separately; do not infer one retention rule across surfaces.; EVIDENCE / LIMIT: EXT-C01-021/022; public documentation does not establish tenant-specific purge completion.

4. MAPPING CLAIM: PM-C01-GPT-004; GPT / OPENAI SURFACE: Background Responses; CORE-01 DECISION: queued, in_progress and terminal response status are runtime states. Business completion still requires durable artifact or authoritative external-state readback.; EVIDENCE / LIMIT: EXT-C01-022; cancellation and stream resumption do not prove compensation or acceptance.

5. MAPPING CLAIM: PM-C01-GPT-005; GPT / OPENAI SURFACE: Webhook response.completed; CORE-01 DECISION: Verify signature, deduplicate by webhook ID, retrieve the response by ID, then apply the separate acceptance contract and external readback.; EVIDENCE / LIMIT: EXT-C01-023; webhook delivery/completion is notification evidence, not business final state.

6. MAPPING CLAIM: PM-C01-GPT-006; GPT / OPENAI SURFACE: Agents SDK Sessions and backend adapters; CORE-01 DECISION: Session continuity depends on backend durability, TTL, consistency, ownership and concurrency controls. Session history must not be the only copy of critical task state.; EVIDENCE / LIMIT: EXT-C01-024; rolling documentation plus fixed v0.22.0 code boundary; live backend values UNKNOWN.

7. MAPPING CLAIM: PM-C01-GPT-007; GPT / OPENAI SURFACE: RunResult.final_output, RunResult.to_state(), RunState; CORE-01 DECISION: Final agent output, paused state and serializable resume state are distinct from artifact existence, acceptance and external effect.; EVIDENCE / LIMIT: EXT-C01-025; fixed-release shape does not guarantee application context serialization or safe concurrent resume.

8. MAPPING CLAIM: PM-C01-GPT-008; GPT / OPENAI SURFACE: Handoff plus RunState restore; CORE-01 DECISION: Handoff payload must carry/refer to task, contract, source, artifacts, open work, authority and stop/readback conditions; restore must bind the original backend/session identity and exclusive ownership.; EVIDENCE / LIMIT: EXT-C01-011 reused; EXT-C01-024/025; the additional fields are CORE requirements, not provider guarantees.

9. MAPPING CLAIM: PM-C01-GPT-009; GPT / OPENAI SURFACE: Session/conversation regression signals; CORE-01 DECISION: Concurrent first use and model/session-history changes belong in regression tests even when reported defects were fixed. They are not evidence of a current v0.22.0 defect.; EVIDENCE / LIMIT: EXT-C01-026; issues #3820/#3123 are historical, closed, version-bounded signals.

## 3. Safe test profile
Use synthetic, non-personal fixtures and a disposable application namespace. Do not execute production tools or external writes.

- Assign separate TASK_ID, CONTRACT_VERSION, provider response ID, conversation ID, session backend/ID, RunState digest and artifact ID.

- Exercise previous_response_id, Conversation, SDK Session and serialized RunState as separate carriers; verify that the canonical task record can resume without hidden chat history.

- Record store, retention class and deletion/readback result per carrier. Keep ZDR/background temporary-storage behavior separate from ordinary response and Conversation persistence.

- Drive a background request through queued/in-progress/terminal and cancellation paths. Treat terminal as runtime-only until an acceptance fixture reads the artifact.

- Replay a signed webhook fixture twice, deduplicate by webhook ID, retrieve by response ID, and ensure duplicate notification cannot duplicate business acceptance.

- Inject concurrent first-use of the same session ID and reject split or ambiguous conversation identity. Serialize/restore RunState only with the expected backend, session ID and exclusive resume owner.

- Require artifact digest/readback or external authoritative receipt before BUSINESS_ACCEPTED; keep final_output, response completion and human approval as separate evidence.

This profile reuses the intent of CORE-01's 6/6 mock integration and 12/12 requiredness tests. Those prior results remain first-party compiler evidence and are not rerun or relabeled here.

## 4. Organization Overlay routing
The Organization Overlay must supply adopted API/SDK surfaces, session backend and owner, persistence/TTL/deletion policy, ZDR/MAM configuration, concurrency lock, webhook dedup retention, polling and incident SLO, authoritative artifact/source system, acceptance authority, retry/repair owner, and live pass thresholds. Public documentation cannot close these values.

## 5. Bounded questions
| QUESTION_ID | QUESTION | STATUS |
| --- | --- | --- |
| Q-PM-C01-GPT-001 | Which response/conversation/session/task identifiers and retention policies are actually adopted? | BOUNDED_PUBLIC_MAPPING / TENANT_CONFIGURATION_UNKNOWN |
| Q-PM-C01-GPT-002 | How are background terminal events and duplicate webhooks joined to artifact/external-state readback? | BOUNDED_SAFE_PROFILE / LIVE_ACCEPTANCE_UNKNOWN |
| Q-PM-C01-GPT-003 | Which session backend, lock, RunState serializer and exclusive-resume rule are used? | BOUNDED_FIXED_CODE / DEPLOYMENT_CONCURRENCY_UNKNOWN |
| Q-PM-C01-GPT-004 | What are the organization retention, RTO/RPO, acceptance, retry and incident-owner values? | ORGANIZATION_OVERLAY_ONLY / UNKNOWN |

## 6. Disposition
- CORE-01 claims remain unchanged; no Gemini backprop is required.

- AUD-C01-001 is bounded for the public GPT mapping but remains open for live provider/tenant execution.

- AUD-C01-007 remains open as a cross-provider falsification target; GPT surfaces do not prove a universal minimum field set.

- Auxiliary evidence communication reported no new evidence gap for this unit; no AUX source was promoted or duplicated.

- Next provider cycle: Claude, beginning with CORE-01 against the same test intent.

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