# GROK：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-grok-core-01
CLAIM_IDS: none
EVIDENCE_IDS: PM-GROK-CORE-01-DOCUMENT-READBACK
UNKNOWN_IDS: PM-GROK-CORE-01-CURRENTNESS
OPS_IDS: none

## Summary
This mapping binds xAI Responses, stored continuation, deferred Chat Completions, Batch API, function calling, ZDR and the official Python SDK to CORE-01. Provider response, request, batch and tool-call states are runtime carriers; they are not the canonical business task record, artifact acceptance, an external commit receipt or proof of the intended final state.

## Current judgment
This mapping binds xAI Responses, stored continuation, deferred Chat Completions, Batch API, function calling, ZDR and the official Python SDK to CORE-01. Provider response, request, batch and tool-call states are runtime carriers; they are not the canonical business task record, artifact acceptance, an external commit receipt or proof of the intended final state.

## Scope
This mapping binds xAI Responses, stored continuation, deferred Chat Completions, Batch API, function calling, ZDR and the official Python SDK to CORE-01. Provider response, request, batch and tool-call states are runtime carriers; they are not the canonical business task record, artifact acceptance, an external commit receipt or proof of the intended final state.

No xAI API key, team, Console, model, SDK process, tool, batch, deferred request or external system was operated. Existing CORE-01 6/6 and 12/12 local results remain first-party COMPILER_EXPERIMENT evidence and are not Grok observations.

## Confirmed

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

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

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

PROVIDER: Grok / xAI

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

STATUS: PUBLIC_MAPPING_BOUNDED_COMPLETE / ROLLING_SDK_UNPINNED / LIVE_ORG_PENDING

AS_OF: 2026-09-12
## 資料の識別情報
Source SHA-256 641281bb8f1820e50f2b754f72230a6f011e14d1566695410e73a4a627e4cc29

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

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

PROVIDER: Grok / xAI

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

STATUS: PUBLIC_MAPPING_BOUNDED_COMPLETE / ROLLING_SDK_UNPINNED / LIVE_ORG_PENDING

AS_OF: 2026-09-12

## 1. Boundary
This mapping binds xAI Responses, stored continuation, deferred Chat Completions, Batch API, function calling, ZDR and the official Python SDK to CORE-01. Provider response, request, batch and tool-call states are runtime carriers; they are not the canonical business task record, artifact acceptance, an external commit receipt or proof of the intended final state.

No xAI API key, team, Console, model, SDK process, tool, batch, deferred request or external system was operated. Existing CORE-01 6/6 and 12/12 local results remain first-party COMPILER_EXPERIMENT evidence and are not Grok observations.

## 2. Claim mapping
| MAPPING CLAIM | GROK / xAI SURFACE | CORE-01 DECISION | EVIDENCE / LIMIT |
| --- | --- | --- | --- |
| PM-C01-GROK-001 | Responses API stored continuation | Server-side storage is on by default. Record response ID separately from TASK_ID, source snapshot, artifact and acceptance state. | EXT-G01-001; responses are documented as retrievable/continuable for 30 days, then removed. Provider retention is not a business checkpoint. |
| PM-C01-GROK-002 | store:false and client-held history | When server storage is disabled, the client owns complete history and encrypted reasoning/tool state needed for continuation. | EXT-G01-001/006; ZDR disables stateful Responses and requires client-side history. No public evidence proves application persistence or recovery quality. |
| PM-C01-GROK-003 | Function calling and previous_response_id | Separate model proposal, local authorization/execution, function output, continuation and authoritative readback. | EXT-G01-002; provider continuation does not supply business authorization, idempotency, receipt or compensation. |
| PM-C01-GROK-004 | Deferred Chat Completions | Track deferred request_id, readiness and consumption separately. A 202 is pending, not failure or completion. | EXT-G01-004; result is documented as retrievable exactly once within 24 hours, then discarded. This is a delivery lifecycle, not artifact acceptance. |
| PM-C01-GROK-005 | Batch API | Track batch ID and each batch_request_id; preserve pending/success/error independently and join results by ID. | EXT-G01-005; num_pending=0 includes success and error. Cancellation can leave completed results; batches expire. |
| PM-C01-GROK-006 | API retention and ZDR | Materialize effective team retention mode and verify the x-zero-data-retention response header for every run. | EXT-G01-006; default 30-day encrypted audit retention, team-wide ZDR and negotiated Enterprise terms are different regimes. |
| PM-C01-GROK-007 | API/Console audit surfaces | Provider request logs and admin audit events are evidence sources, not immutable business receipts. | EXT-G01-006; content logging changes under ZDR and public docs do not establish organization retention/export/immutability. |
| PM-C01-GROK-008 | Official Python SDK | Pin installed package/version and transport behavior before treating SDK state as reproducible. | EXT-G01-007; official gRPC SDK supports sync/async clients and states SemVer intent, while warning that internal breaking changes may occur in minor/patch releases. Rolling main is not a fixed runtime anchor. |
| PM-C01-GROK-009 | Model/product identity | Record requested model, observed response model/fingerprint/headers, route and organization policy separately. | EXT-G01-003/004; public grok-4.6 metadata and response examples do not prove the actual tenant route or configuration. |

## 3. Safe test profile
Use synthetic non-personal inputs and a disposable mock sink. Do not connect production tools or data.

- Assign distinct task, contract, source snapshot, response, deferred request, batch, batch-request, tool-call, artifact and acceptance IDs.

- Run equivalent local fixtures for server-stored and store:false paths; prove that loss of client-held history fails closed rather than inventing prior work.

- Continue a stored response before and after a synthetic expiry boundary; do not infer real retention without a permitted tenant test.

- Inject pending (202), success, consumed/missing and expired deferred-result records. A deferred result may be accepted only once and only after artifact validation.

- Shuffle batch results and mix success/error/cancel. Join by batch_request_id; num_pending=0 cannot set BUSINESS_ACCEPTED while a required item errored or is absent.

- Replay the same function call against an idempotent mock sink. Record proposal, authorization, operation digest, provider continuation, sink receipt and readback independently.

- Compare declared retention mode with observed response header; reject a run whose effective mode is missing or mismatched.

- Accept only after artifact existence/digest and authoritative external-state readback satisfy CORE-01.

This is a test design, not a record of running xAI services or the SDK.

## 4. Organization Overlay routing
The Organization Overlay must supply adopted API/model/SDK pins; team and key ownership; default-retention/ZDR/negotiated-term choice; request-log and audit-log retention/export; client-side encrypted history store; deferred/batch polling, expiry and recovery policy; idempotency key and compensation owner; authoritative artifact/external system; RTO/RPO; incident owner; and business acceptance authority. Public documentation cannot close these values.

## 5. Bounded questions
| QUESTION_ID | QUESTION | STATUS |
| --- | --- | --- |
| Q-PM-C01-GROK-001 | Which stored-response, store:false, ZDR or negotiated-retention regime is actually adopted, and how is expiry handled? | PUBLIC_DEFAULTS_CLOSED / EFFECTIVE_TEAM_CONFIGURATION_UNKNOWN |
| Q-PM-C01-GROK-002 | Can repeated tool continuation or retry duplicate an external effect, and what receipt/readback/compensation contract is used? | BOUNDED_SAFE_PROFILE / LIVE_EXTERNAL_EFFECT_UNKNOWN |
| Q-PM-C01-GROK-003 | Which model, SDK build, route, fingerprint, headers and logs identify the actual execution? | ROLLING_PUBLIC_SURFACE / FIXED_LIVE_IDENTITY_UNKNOWN |
| Q-PM-C01-GROK-004 | How are deferred and batch terminal states joined to artifact validation and business acceptance? | PUBLIC_LIFECYCLE_BOUNDED / LIVE_ACCEPTANCE_UNKNOWN |

## 6. Disposition
- CORE-01 claims remain unchanged; Gemini, GPT and Claude backprop are not required.

- AUD-C01-001 remains open for live adapter execution and AUD-C01-007 for cross-provider falsification of the provisional requiredness profile.

- Public evidence is bounded to 2026-09-12. The SDK repository is author-maintained but not fixed to a locally executed package/version/commit.

- No independent production operation, xAI issue reproduction or live tenant test was available in this bounded unit; absence is retained as evidence gap, not converted to safety evidence.

- Next provider unit: Grok CORE-02 against the existing permission-aware retrieval, memory lifecycle and bounded long-context test intent.

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