{
  "blocked_count": 0,
  "blockers_remaining": [
    "RUNTIME_SSE_EVAL_FAILED",
    "ANSWER_LADDER_EVAL_FAILED",
    "SOURCE_GROUNDING_EVAL_FAILED",
    "SOURCE_FAKE_GUARD_EVAL_FAILED",
    "FINAL_JUDGE_EVAL_FAILED"
  ],
  "dashboard_path": "config/capabilities/aw1_capability_factory_dashboard_phase3_runtime_source_memory.json",
  "eval_results": {
    "answer_ladder_eval": {
      "done_count": 0,
      "final_done_count": 0,
      "latency_ms": 8011.615,
      "missing_required_events": [
        "final_done",
        "judge_final",
        "quick_answer",
        "source_bound_deep_answer",
        "token_delta"
      ],
      "pass": false,
      "required_events_present": []
    },
    "citation_freshness_eval": {
      "checked_count": 10,
      "pass": true,
      "row_count": 10,
      "stale_reject_policy_present": true
    },
    "fact_pack_memory_eval": {
      "cache_hit_rate_observed": 1.0,
      "leak_count": 0,
      "missing_required_field_count": 0,
      "pass": true,
      "receipt_path": "aw1_source_fact_pack_memory_latest.jsonl",
      "row_count": 10,
      "source_bound_count": 7,
      "valid_schema_count": 10
    },
    "final_judge_eval": {
      "judge_final_count": 0,
      "latency_ms": 8011.615,
      "no_fake_pass_seen": false,
      "pass": false
    },
    "runtime_sse_eval": {
      "content_delta_count": 0,
      "done_count": 0,
      "event_counts": {
        "activity.started": 1,
        "analysis.summary": 1,
        "answer.quick.completed": 1,
        "answer.quick.delta": 1,
        "answer.quick.started": 1,
        "claim.created": 4,
        "evidence.gap_detected": 1,
        "group.started": 2,
        "research.plan.created": 1,
        "response.started": 1,
        "source.search_started": 1
      },
      "final_done_count": 0,
      "finish_reason_count": 0,
      "latency_ms": 8007.474,
      "missing_required_events": [
        "final_done",
        "judge_final",
        "quick_answer",
        "source_bound_deep_answer",
        "token_delta"
      ],
      "pass": false,
      "public_sse_metadata": [
        {
          "event": "response.started"
        },
        {
          "event": "group.started"
        },
        {
          "event": "activity.started"
        },
        {
          "event": "answer.quick.started"
        },
        {
          "event": "answer.quick.delta"
        },
        {
          "event": "answer.quick.completed"
        },
        {
          "event": "claim.created"
        },
        {
          "event": "claim.created"
        },
        {
          "event": "claim.created"
        },
        {
          "event": "claim.created"
        },
        {
          "event": "research.plan.created"
        },
        {
          "event": "group.started"
        },
        {
          "event": "source.search_started"
        },
        {
          "event": "evidence.gap_detected"
        },
        {
          "event": "analysis.summary"
        }
      ],
      "required_events_present": [],
      "returncode": 28,
      "stderr_tail_present": true
    },
    "source_fake_guard_eval": {
      "pass": false,
      "source_fake_count": 1
    },
    "source_grounding_eval": {
      "latency_ms": 8009.257,
      "live_stream_pass": false,
      "pass": false,
      "receipt_verdict": "PASS",
      "source_fake_counts_public": [],
      "tests_run": [
        "post_patch_chat_smoke_temp_log_timeout_180",
        "source_needed_binding_eval",
        "secret_internal_metadata_leak_scan"
      ]
    }
  },
  "implemented_count": 78,
  "no_fake_pass": true,
  "physical_delete_approved": false,
  "production_hot_path_changed": false,
  "production_ready_count": 11,
  "production_weights_change": false,
  "qwen_vllm_deleted": false,
  "registry_path": "config/capabilities/aw1_capability_factory_registry_phase3_runtime_source_memory.json",
  "schema_version": "aw1.capability_factory.phase3_runtime_source_memory.receipt.v1",
  "source_policy_decision": "S4_safe_runtime_source_memory_eval_no_production_mutation",
  "status_upgrades": [
    {
      "capability_id": 2,
      "eval_name": "fact_pack_memory_eval",
      "evidence_path": "/log/aw1_capability_factory_phase3_runtime_source_memory_latest.json",
      "evidence_receipt": "aw1_capability_factory_phase3_runtime_source_memory_latest.json",
      "fail_count": 0,
      "latency_if_available": null,
      "new_status": "production_ready",
      "old_status": "tested",
      "pass_count": 1,
      "remaining_gap": "Continue rolling production regression; no weight/tokenizer/Qwen delete in this phase."
    },
    {
      "capability_id": 8,
      "eval_name": "citation_freshness_eval",
      "evidence_path": "/log/aw1_capability_factory_phase3_runtime_source_memory_latest.json",
      "evidence_receipt": "aw1_capability_factory_phase3_runtime_source_memory_latest.json",
      "fail_count": 0,
      "latency_if_available": null,
      "new_status": "production_ready",
      "old_status": "tested",
      "pass_count": 1,
      "remaining_gap": "Continue rolling production regression; no weight/tokenizer/Qwen delete in this phase."
    },
    {
      "capability_id": 50,
      "eval_name": "fact_pack_memory_eval",
      "evidence_path": "/log/aw1_capability_factory_phase3_runtime_source_memory_latest.json",
      "evidence_receipt": "aw1_capability_factory_phase3_runtime_source_memory_latest.json",
      "fail_count": 0,
      "latency_if_available": null,
      "new_status": "production_ready",
      "old_status": "tested",
      "pass_count": 1,
      "remaining_gap": "Continue rolling production regression; no weight/tokenizer/Qwen delete in this phase."
    },
    {
      "capability_id": 74,
      "eval_name": "citation_freshness_eval",
      "evidence_path": "/log/aw1_capability_factory_phase3_runtime_source_memory_latest.json",
      "evidence_receipt": "aw1_capability_factory_phase3_runtime_source_memory_latest.json",
      "fail_count": 0,
      "latency_if_available": null,
      "new_status": "production_ready",
      "old_status": "tested",
      "pass_count": 1,
      "remaining_gap": "Continue rolling production regression; no weight/tokenizer/Qwen delete in this phase."
    },
    {
      "capability_id": 83,
      "eval_name": "citation_freshness_eval",
      "evidence_path": "/log/aw1_capability_factory_phase3_runtime_source_memory_latest.json",
      "evidence_receipt": "aw1_capability_factory_phase3_runtime_source_memory_latest.json",
      "fail_count": 0,
      "latency_if_available": null,
      "new_status": "production_ready",
      "old_status": "tested",
      "pass_count": 1,
      "remaining_gap": "Continue rolling production regression; no weight/tokenizer/Qwen delete in this phase."
    },
    {
      "capability_id": 96,
      "eval_name": "fact_pack_memory_eval",
      "evidence_path": "/log/aw1_capability_factory_phase3_runtime_source_memory_latest.json",
      "evidence_receipt": "aw1_capability_factory_phase3_runtime_source_memory_latest.json",
      "fail_count": 0,
      "latency_if_available": null,
      "new_status": "production_ready",
      "old_status": "tested",
      "pass_count": 1,
      "remaining_gap": "Continue rolling production regression; no weight/tokenizer/Qwen delete in this phase."
    },
    {
      "capability_id": 97,
      "eval_name": "fact_pack_memory_eval",
      "evidence_path": "/log/aw1_capability_factory_phase3_runtime_source_memory_latest.json",
      "evidence_receipt": "aw1_capability_factory_phase3_runtime_source_memory_latest.json",
      "fail_count": 0,
      "latency_if_available": null,
      "new_status": "production_ready",
      "old_status": "tested",
      "pass_count": 1,
      "remaining_gap": "Continue rolling production regression; no weight/tokenizer/Qwen delete in this phase."
    },
    {
      "capability_id": 98,
      "eval_name": "fact_pack_memory_eval",
      "evidence_path": "/log/aw1_capability_factory_phase3_runtime_source_memory_latest.json",
      "evidence_receipt": "aw1_capability_factory_phase3_runtime_source_memory_latest.json",
      "fail_count": 0,
      "latency_if_available": null,
      "new_status": "production_ready",
      "old_status": "tested",
      "pass_count": 1,
      "remaining_gap": "Continue rolling production regression; no weight/tokenizer/Qwen delete in this phase."
    },
    {
      "capability_id": 101,
      "eval_name": "fact_pack_memory_eval",
      "evidence_path": "/log/aw1_capability_factory_phase3_runtime_source_memory_latest.json",
      "evidence_receipt": "aw1_capability_factory_phase3_runtime_source_memory_latest.json",
      "fail_count": 0,
      "latency_if_available": null,
      "new_status": "production_ready",
      "old_status": "tested",
      "pass_count": 1,
      "remaining_gap": "Continue rolling production regression; no weight/tokenizer/Qwen delete in this phase."
    },
    {
      "capability_id": 102,
      "eval_name": "fact_pack_memory_eval",
      "evidence_path": "/log/aw1_capability_factory_phase3_runtime_source_memory_latest.json",
      "evidence_receipt": "aw1_capability_factory_phase3_runtime_source_memory_latest.json",
      "fail_count": 0,
      "latency_if_available": null,
      "new_status": "production_ready",
      "old_status": "tested",
      "pass_count": 1,
      "remaining_gap": "Continue rolling production regression; no weight/tokenizer/Qwen delete in this phase."
    },
    {
      "capability_id": 107,
      "eval_name": "fact_pack_memory_eval",
      "evidence_path": "/log/aw1_capability_factory_phase3_runtime_source_memory_latest.json",
      "evidence_receipt": "aw1_capability_factory_phase3_runtime_source_memory_latest.json",
      "fail_count": 0,
      "latency_if_available": null,
      "new_status": "production_ready",
      "old_status": "tested",
      "pass_count": 1,
      "remaining_gap": "Continue rolling production regression; no weight/tokenizer/Qwen delete in this phase."
    }
  ],
  "tested_count": 78,
  "timestamp_utc": "2026-08-03T13:56:24Z",
  "tokenizer_changed": false,
  "total_capabilities": 156,
  "upgrade_count": 11,
  "verdict": "BLOCKED_WITH_REASON"
}