{
  "advisory_priors": [
    "Client acceptance proves semantic correctness, tool continuation, selected route, and healthy completion; it is not a 13-of-13 strict-format claim.",
    "Existing conversations were not reloaded, no secondary-text readiness claim is made, and the OpenClaw compaction reserve is distinct from the router output ceiling.",
    "Public copies must sanitize private host paths, user identifiers, and the real WebUI domain before publication."
  ],
  "artifacts": {
    "client_acceptance_summary": {
      "path": "client-acceptance-summary.json",
      "sha256": "3622d7896ea19a347de6d0bb0f4c0a4d24286d36318295494b47c71426f91800"
    },
    "controller_mount_live_verification": {
      "path": "controller-mount-live-verification.json",
      "sha256": "82d3152bf0249f6a71a9cea4cc3587d2c768da15f9d7e61942acd9b1c0f6ec38"
    },
    "findings_summary": [
      {
        "classification": "confirmed",
        "severity": "none",
        "summary": "All 13 declared semantic client paths completed with the expected value or explicitly scoped auxiliary-vision result; native finishes were healthy and no fallback or model substitution was observed."
      },
      {
        "classification": "confirmed",
        "severity": "none",
        "summary": "Fleet reconciliation repeated with changed=0 and failed=0 on Primary Node, Companion Node, Laptop Node, and localhost."
      },
      {
        "classification": "confirmed",
        "severity": "none",
        "summary": "The controller exposes all seven retained r9/r10/r11 manifest dependencies as exact read-only mounts, remained healthy after its managed recreation, and the router and GLM container identities remained unchanged."
      },
      {
        "classification": "confirmed_limit",
        "severity": "low",
        "summary": "Strict return-only-JSON formatting did not pass on five Pi paths, Pi Web, two auxiliary-vision Hermes paths, or Mac OpenClaw. Semantic values, tool continuation, route selection, and termination passed. Mac OpenClaw additionally prepended explanatory prose."
      },
      {
        "classification": "confirmed_limit",
        "severity": "low",
        "summary": "Pi Web get_state reported messageCount=0, while its retained native six-message session transcript independently proves two tool results and a stop completion."
      },
      {
        "classification": "confirmed_limit",
        "severity": "low",
        "summary": "OpenClaw preserves its existing 20000-token compaction reserve. The router catalog still advertises the configured 65536-token output ceiling, but this evidence does not establish 65536-token usable output through OpenClaw."
      },
      {
        "classification": "confirmed_limit",
        "severity": "low",
        "summary": "Legacy Hermes and OpenClaw lack opaque request-header correlation. Their evidence is bounded to terminal requested/effective model receipts, no-fallback traces, and the same-config router decision window; no fabricated exact request join is claimed."
      }
    ],
    "fleet_repeat": {
      "path": "harness-r11-repeat.log",
      "sha256": "d8ba143f7c57c4ba28ec7cf0eced5831c02dc2ad3ef402053a8a9de715bf3c07"
    }
  },
  "breakage_probes": [
    {
      "axis": "fail_closed_unreadable_state",
      "classification": "refuted",
      "evidence": "All cited artifacts are readable and hash-bound.",
      "input_state": "Any client terminal receipt, fleet repeat, controller health, or exact mount observation is missing or unreadable",
      "wrong_result": "Accept convergence from an incomplete summary"
    },
    {
      "axis": "malformed_or_boundary_output",
      "classification": "confirmed",
      "evidence": "Per-path format_only_json and visible_format_caveat fields preserve the failures.",
      "input_state": "Correct semantic values are wrapped in Markdown or preceded by prose despite a return-only-JSON instruction",
      "wrong_result": "Report strict-format compliance"
    },
    {
      "axis": "resource_or_budget_boundary",
      "classification": "refuted",
      "evidence": "The artifact separates compaction reserve from the configured catalog output ceiling and makes no long-output client claim.",
      "input_state": "Treat OpenClaw reserveTokens=20000 as either a 65536 reserve or proof that the advertised output ceiling is unusable",
      "wrong_result": "Misstate client capacity"
    },
    {
      "axis": "state_drift_across_clients_router_controller",
      "classification": "refuted",
      "evidence": "Fleet repeat changed=0; controller health and seven mounts pass; router and GLM identities are unchanged.",
      "input_state": "Clients retain stale catalog state or the controller-only recreation changes router/GLM ownership",
      "wrong_result": "Close promotion with stale clients or changed serving identity"
    }
  ],
  "gate_state": "accepted_with_caveats",
  "human_gate_required": false,
  "promoted": false,
  "recommendation": "accept_configuration_promotion_and_client_control_plane_convergence_with_documented_caveats; publication_and_live_readback_remain_required_for_campaign_closure",
  "request": "Independently review final native-client acceptance, fleet convergence, and controller mount closure for the GLM DCP1 production configuration.",
  "schema_version": "operator-workflow/v1",
  "targets": [
    {
      "identity": "glm53-flash-exl3-4bpw-v84-fp8-tp2-dcp1-c4-327k-r11-vision8-apc-nospec",
      "kind": "evaluated_model"
    },
    {
      "kind": "implementation_author",
      "model": "GPT-6 Astra"
    },
    {
      "kind": "independent_reviewer",
      "model": "GPT-5.6 Sol"
    },
    {
      "count": 13,
      "kind": "client_paths"
    },
    {
      "count": 4,
      "kind": "fleet_targets"
    },
    {
      "count": 7,
      "kind": "controller_mounts"
    }
  ],
  "tools_used": [
    "read-only artifact inspection",
    "bounded raw-output sampling",
    "SHA-256 verification"
  ]
}