{
  "format": "agent-article-v0",
  "article_id": "PM-GROK-CORE-03",
  "canonical_url": "/providers/pm-grok-core-03",
  "page_type": "reference",
  "claim_ids": [],
  "evidence_ids": [
    "PM-GROK-CORE-03-DOCUMENT-READBACK"
  ],
  "unknown_ids": [
    "PM-GROK-CORE-03-CURRENTNESS"
  ],
  "operational_object_ids": [],
  "provider": [
    "GROK"
  ],
  "models": [],
  "plans": [],
  "surfaces": [],
  "regions": [],
  "current_status": "UNKNOWN",
  "observation_status": "UNOBSERVED",
  "freshness": "UNKNOWN",
  "last_verified": null,
  "supersedes": [],
  "superseded_by": [],
  "provenance": [
    "DOC-083cde4caa627644-PM-GROK-CORE-03"
  ],
  "article_sources": [
    {
      "id": "PM-GROK-CORE-03-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-GROK-CORE-03"
      ],
      "supporting_locators": [
        "18WPldGemks2px0ukhKYINeBMrBwAmlAi",
        "083cde4caa627644036a2225f1f6ff29b8015bd902614fe1f7a00a997b4eb59c",
        "provider_mappings/PM-GROK-CORE-03.md",
        "ef0c2d16e46a1ebe868b8421cd7abf4d3f81f1a24dc78273dd97636036453401"
      ],
      "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 7f28ba7f8ce76abc662e1bfef0f6739a372820dc4e346edf353f8cd6fbb524ef"
    },
    {
      "heading": "Product Mapping — Grok / CORE-03 Orchestration, Lineage and Recovery",
      "text": "MAPPING_ID: PM-GROK-CORE-03-v0.1\n\nPRODUCT_MAPPING_ID: PM-GROK-CORE-03-v0.1\n\nPROVIDER: Grok / xAI\n\nCORE_PACKET: PKT-GEMINI-A-CORE-03-v0.1\n\nSTATUS: PUBLIC_MAPPING_AUDITED / MULTI_AGENT_AND_TOOL_LOOP_BOUNDED / LIVE_RUNTIME_AND_ORG_PENDING\n\nAS_OF: 2026-09-12\n\nLAYER: Product Mapping"
    },
    {
      "heading": "1. Boundary",
      "text": "This mapping binds xAI Responses, function/tool calling, Multi-Agent, Remote MCP, context compaction, WebSocket mode, deferred requests and Batch to the existing CORE-03 contract. Provider continuation, tool-loop completion, leader output, encrypted state or terminal request status is not treated as a canonical stage ledger, complete subagent lineage, accepted artifact, external transaction or business completion.\n\nNo xAI API key, team, Console, model, SDK process, WebSocket, MCP server, custom tool, deferred request, Batch job or external system was operated. Existing CORE-03 6/6 contract and LangGraph/SQLite 15/15 artifacts remain COMPILER_EXPERIMENT and are not Grok observations."
    },
    {
      "heading": "2. Surface separation",
      "text": "- Responses state and previous_response_id provide provider continuation; the application still owns the canonical task, stage, artifact, attempt and acceptance records.\n\n- Server-side tools execute within xAI's request loop. Client-side/custom tools pause the loop, return control to the application and require a new request after execution.\n\n- Function calling permits multiple tool calls in one response and enables parallel_tool_calls by default. Parallel proposal is not proof that shared writes or side effects are safe to execute concurrently.\n\n- Multi-Agent uses a leader with 4 or 16 subagents. Leader-visible tool calls and final output are not a complete record of subagent intermediate state, tool use or outputs.\n\n- Remote MCP is provider-managed connectivity. Tool exposure, upstream authorization, write approval and authoritative readback remain distinct controls.\n\n- Context compaction and WebSocket continuation are context carriers. They do not replace externalized versioned artifacts or durable recovery records."
    },
    {
      "heading": "3. Claim mapping",
      "text": "1. MAPPING CLAIM: PM-C03-GROK-001; GROK / xAI SURFACE: Responses status and continuation; CORE-03 DECISION: Record response ID/status separately from stage completion, artifact acceptance and business completion.; EVIDENCE / LIMIT: EXT-G03-007; provider states include in-progress/incomplete/completed, but those states do not verify downstream effects.\n\n2. MAPPING CLAIM: PM-C03-GROK-002; GROK / xAI SURFACE: Server-side versus client-side tools; CORE-03 DECISION: Treat a client-tool pause as a handoff: persist call ID, input digest, attempt, owner, result and authoritative readback before continuation.; EVIDENCE / LIMIT: EXT-G03-001/002; the application executes custom tools and supplies results.\n\n3. MAPPING CLAIM: PM-C03-GROK-003; GROK / xAI SURFACE: Default parallel function calls; CORE-03 DECISION: Execute only read-only, disjoint-target or deterministically mergeable calls concurrently; serialize/lock shared writes and side effects.; EVIDENCE / LIMIT: EXT-G03-001; multiple calls may be requested and parallel_tool_calls is enabled by default. Safety of external concurrency is not guaranteed.\n\n4. MAPPING CLAIM: PM-C03-GROK-004; GROK / xAI SURFACE: max_turns; CORE-03 DECISION: Enforce an application-global step/depth/time/token/cost budget across requests.; EVIDENCE / LIMIT: EXT-G03-002; the documented limit applies within a request and resets after a client-side tool pause followed by a new request.\n\n5. MAPPING CLAIM: PM-C03-GROK-005; GROK / xAI SURFACE: Multi-Agent leader/subagents; CORE-03 DECISION: Preserve leader request, agent count, tool policy, encrypted carrier, accepted output and verification separately. Do not infer complete lineage from leader output.; EVIDENCE / LIMIT: EXT-G03-003; 4/16-agent modes and leader-visible output are documented; subagent intermediate state is hidden by default and optionally returned encrypted.\n\n6. MAPPING CLAIM: PM-C03-GROK-006; GROK / xAI SURFACE: Multi-Agent compatibility; CORE-03 DECISION: Treat Multi-Agent as a restricted product surface, not a generic substitute for application orchestration.; EVIDENCE / LIMIT: EXT-G03-003; only built-in tools and Remote MCP are supported, not client/custom tools; Responses/xAI SDK only.\n\n7. MAPPING CLAIM: PM-C03-GROK-007; GROK / xAI SURFACE: Remote MCP; CORE-03 DECISION: Use explicit allowed_tools, least privilege, upstream authorization and external receipt/readback. Do not treat MCP reachability as approval.; EVIDENCE / LIMIT: EXT-G03-004; omitting allowed_tools exposes all server tools, while Responses-style require_approval is unsupported.\n\n8. MAPPING CLAIM: PM-C03-GROK-008; GROK / xAI SURFACE: Context compaction; CORE-03 DECISION: Externalize stage evidence before compaction and bind the opaque item plus dropped-message count to the task ledger.; EVIDENCE / LIMIT: EXT-G03-005; compaction can drop verbose tool output and yields an opaque item that cannot be parsed, edited or merged. Continuity is not audit completeness.\n\n9. MAPPING CLAIM: PM-C03-GROK-009; GROK / xAI SURFACE: WebSocket mode; CORE-03 DECISION: Record connection lineage and recovery branch; do not assume a failed continuation can resume. Use separate connections for intentional parallelism.; EVIDENCE / LIMIT: EXT-G03-006; one connection is serial, sessions end after 25 minutes, and some errors evict prior response state. Under ZDR/store:false, missing in-memory state requires a fresh chain.\n\n10. MAPPING CLAIM: PM-C03-GROK-010; GROK / xAI SURFACE: Deferred and Batch; CORE-03 DECISION: Join provider lifecycle/results to artifact validation and authoritative external-state acceptance.; EVIDENCE / LIMIT: EXT-G01-004/005 reused; terminal or drained provider work can still contain expiry, cancellation, error or partial results."
    },
    {
      "heading": "4. Safe live-test profile",
      "text": "Use synthetic inputs, disposable credentials and mock tools/MCP endpoints with no production data or external side effects.\n\n- Compare a single-stage baseline with explicit client-tool handoff and verify call/result/attempt/artifact/readback joins.\n\n- Return two parallel calls for read-only targets, disjoint writes and a shared write. Permit only the first two without an explicit lock/merge contract.\n\n- Cross a client-tool pause repeatedly and confirm the application-global budget stops the workflow even though request-local max_turns resets.\n\n- Compare 4-agent and 16-agent Multi-Agent runs; record leader-visible evidence, encrypted carrier availability, usage/latency and missing subagent lineage. Do not accept leader summary without artifact verification.\n\n- Expose an MCP server with read and synthetic write tools. Confirm explicit allowlisting, upstream identity, application approval and post-write readback; omission of any control must fail closed.\n\n- Compact after writing a stage-evidence digest. Verify that the external ledger remains reconstructable even when verbose tool output is dropped and the compacted item is opaque.\n\n- Exercise WebSocket disconnect, 4xx/5xx eviction, 25-minute boundary simulation and ZDR/store:false missing-state recovery. Resume only with a verified chain; otherwise start a new attempt linked to the prior failure.\n\n- Treat deferred/Batch completion as input to a separate artifact and business-acceptance gate.\n\nThis is a test design, not an xAI runtime observation."
    },
    {
      "heading": "5. Organization Overlay routing",
      "text": "The Organization Overlay must supply adopted model/API/SDK release, allowed agent count, tool and MCP allowlists, authentication owner, approval owner, global step/token/cost/time limits, concurrency and lock policy, artifact/session stores, retention/ZDR regime, retry/idempotency/compensation policy, verifier, business-acceptance owner, audit export, RTO/RPO and incident owner. Public defaults cannot populate these values."
    },
    {
      "heading": "6. Bounded questions",
      "text": "| QUESTION_ID | QUESTION | STATUS |\n| --- | --- | --- |\n| Q-PM-C03-GROK-001 | Across request-local max_turns, client-tool pauses and default parallel calls, what application-global bound and shared-write policy is enforced? | PUBLIC_LOOP_BOUNDARY_BOUNDED / LIVE_RUNTIME_UNKNOWN |\n| Q-PM-C03-GROK-002 | In Multi-Agent mode, how are leader output, hidden/encrypted subagent work, artifacts, fan-in verification and acceptance joined? | PUBLIC_VISIBILITY_BOUNDARY_BOUNDED / COMPLETE_LINEAGE_UNKNOWN |\n| Q-PM-C03-GROK-003 | After compaction, WebSocket failure or ZDR/store:false state loss, what evidence survives and what verified recovery branch is used? | PUBLIC_RECOVERY_BOUNDARY_BOUNDED / LIVE_DURABILITY_UNKNOWN |\n| Q-PM-C03-GROK-004 | Which MCP tools, upstream identities, approval owners, receipts and authoritative readbacks are effective in the adopted organization? | PUBLIC_CONTROL_GAP_BOUNDED / ORGANIZATION_VALUES_UNKNOWN |"
    },
    {
      "heading": "7. Disposition",
      "text": "- CORE-03 claims remain unchanged; Gemini, GPT and Claude backprop are not required.\n\n- Public xAI evidence establishes product carriers and important failure boundaries, not complete orchestration lineage, external atomicity or business acceptance.\n\n- No reliable independent xAI production report, fixed author-maintained SDK reproduction or authorized live result was promoted in this unit. The gap remains explicit.\n\n- Next provider unit: Grok CORE-04. REQ-NLM-20260909-001 remains a mandatory pre-integration supplement."
    }
  ],
  "human_url": "/providers/pm-grok-core-03",
  "markdown_url": "/agent/article.md?path=%2Fproviders%2Fpm-grok-core-03",
  "json_url": "/agent/article.json?path=%2Fproviders%2Fpm-grok-core-03"
}