{
  "schemaVersion": "agentcert.semantic_adapter_matrix.v0.1",
  "generatedAt": "2026-07-20T19:46:59.986Z",
  "dataset": {
    "schemaVersion": "agentcert.semantic_golden_dataset.v0.1",
    "version": "0.1.0",
    "createdAt": "2026-07-20T00:00:00.000Z"
  },
  "status": "passed",
  "metrics": {
    "total": 33,
    "expectedKnown": 28,
    "expectedUnknown": 5,
    "exact": 33,
    "falseUnknown": 0,
    "misclassified": 0,
    "falseKnown": 0,
    "exactMatchRate": 100,
    "falseUnknownRate": 0
  },
  "adapters": [
    {
      "id": "browser-use",
      "category": "browser",
      "project": "Browser Use",
      "repository": "https://github.com/browser-use/browser-use",
      "sourceRef": "2be09b6c5eb702a9287684b42b27e7042a1aba29",
      "sourceUrl": "https://github.com/browser-use/browser-use/blob/2be09b6c5eb702a9287684b42b27e7042a1aba29/browser_use/agent/views.py",
      "framework": "browser-use",
      "calibrationMode": "recorded_run_and_source_contract",
      "cases": [
        {
          "id": "browser-use-navigate",
          "observedName": "navigate",
          "eventType": "browser_use.step.completed",
          "expectedCapabilityId": "browser.navigate",
          "rationale": "Browser Use records navigation as a named browser step."
        },
        {
          "id": "browser-use-search",
          "observedName": "search",
          "eventType": "browser_use.step.completed",
          "expectedCapabilityId": "browser.navigate",
          "rationale": "Search changes the page context while remaining a read-oriented navigation operation."
        },
        {
          "id": "browser-use-click",
          "observedName": "click",
          "eventType": "browser_use.step.completed",
          "expectedCapabilityId": "browser.interact",
          "rationale": "Click interacts with a page element and may mutate page-local state."
        },
        {
          "id": "browser-use-input",
          "observedName": "input",
          "eventType": "browser_use.step.completed",
          "expectedCapabilityId": "browser.interact",
          "rationale": "Input writes into a page element but does not itself prove an external submission."
        },
        {
          "id": "browser-use-scroll",
          "observedName": "scroll",
          "eventType": "browser_use.step.completed",
          "expectedCapabilityId": "browser.interact",
          "rationale": "Scroll is a page interaction rather than a navigation or submission."
        }
      ],
      "status": "passed",
      "metrics": {
        "total": 5,
        "expectedKnown": 5,
        "expectedUnknown": 0,
        "exact": 5,
        "falseUnknown": 0,
        "misclassified": 0,
        "falseKnown": 0,
        "exactMatchRate": 100,
        "falseUnknownRate": 0
      }
    },
    {
      "id": "openhands",
      "category": "coding",
      "project": "OpenHands",
      "repository": "https://github.com/OpenHands/OpenHands",
      "sourceRef": "fad3228f0cfe2422ed4f061d2606e58e74b3e2b5",
      "sourceUrl": "https://github.com/OpenHands/OpenHands/tree/fad3228f0cfe2422ed4f061d2606e58e74b3e2b5",
      "framework": "OpenHands",
      "calibrationMode": "source_pinned_contract",
      "cases": [
        {
          "id": "openhands-bash",
          "observedName": "execute_bash",
          "eventType": "tool.completed",
          "expectedCapabilityId": "coding.execute",
          "rationale": "The OpenHands Bash tool executes a process in the workspace."
        },
        {
          "id": "openhands-ipython",
          "observedName": "execute_ipython_cell",
          "eventType": "tool.completed",
          "expectedCapabilityId": "coding.execute",
          "rationale": "The OpenHands IPython tool executes code."
        },
        {
          "id": "openhands-editor-view",
          "observedName": "str_replace_editor",
          "eventType": "tool.completed",
          "payload": {
            "tool_input": {
              "command": "view",
              "path": "/workspace/app.ts"
            }
          },
          "expectedCapabilityId": "coding.read",
          "rationale": "The view command reads a source file without modifying it."
        },
        {
          "id": "openhands-editor-create",
          "observedName": "str_replace_editor",
          "eventType": "tool.completed",
          "payload": {
            "tool_input": {
              "command": "create",
              "path": "/workspace/app.ts"
            }
          },
          "expectedCapabilityId": "coding.write",
          "rationale": "The create command writes a source file."
        },
        {
          "id": "openhands-editor-replace",
          "observedName": "str_replace_editor",
          "eventType": "tool.completed",
          "payload": {
            "tool_input": {
              "command": "str_replace",
              "path": "/workspace/app.ts"
            }
          },
          "expectedCapabilityId": "coding.write",
          "rationale": "The str_replace command mutates an existing source file."
        }
      ],
      "status": "passed",
      "metrics": {
        "total": 5,
        "expectedKnown": 5,
        "expectedUnknown": 0,
        "exact": 5,
        "falseUnknown": 0,
        "misclassified": 0,
        "falseKnown": 0,
        "exactMatchRate": 100,
        "falseUnknownRate": 0
      }
    },
    {
      "id": "langchain-sql-toolkit",
      "category": "data",
      "project": "LangChain Community SQLDatabaseToolkit",
      "repository": "https://github.com/langchain-ai/langchain-community",
      "sourceRef": "f425a3ed1933173fb3694b81359d1519c4f82d36",
      "sourceUrl": "https://github.com/langchain-ai/langchain-community/blob/f425a3ed1933173fb3694b81359d1519c4f82d36/libs/community/langchain_community/tools/sql_database/tool.py",
      "framework": "langchain-sql-agent",
      "calibrationMode": "source_pinned_contract",
      "cases": [
        {
          "id": "langchain-sql-query",
          "observedName": "sql_db_query",
          "eventType": "tool.completed",
          "expectedCapabilityId": "data.query",
          "rationale": "The SQL query tool reads structured database results in the documented toolkit contract."
        },
        {
          "id": "langchain-sql-schema",
          "observedName": "sql_db_schema",
          "eventType": "tool.completed",
          "expectedCapabilityId": "data.query",
          "rationale": "Schema inspection is a read-only database operation."
        },
        {
          "id": "langchain-sql-tables",
          "observedName": "sql_db_list_tables",
          "eventType": "tool.completed",
          "expectedCapabilityId": "data.query",
          "rationale": "Listing tables reads database metadata."
        },
        {
          "id": "langchain-sql-checker",
          "observedName": "sql_db_query_checker",
          "eventType": "tool.completed",
          "expectedCapabilityId": "data.query",
          "rationale": "The query checker analyzes a query without establishing a write outcome."
        }
      ],
      "status": "passed",
      "metrics": {
        "total": 4,
        "expectedKnown": 4,
        "expectedUnknown": 0,
        "exact": 4,
        "falseUnknown": 0,
        "misclassified": 0,
        "falseKnown": 0,
        "exactMatchRate": 100,
        "falseUnknownRate": 0
      }
    },
    {
      "id": "cloudflare-agents-email",
      "category": "messaging",
      "project": "Cloudflare Agents Email",
      "repository": "https://github.com/cloudflare/agents",
      "sourceRef": "365ced36bab5f713f1669f3b79f6810ba1535729",
      "sourceUrl": "https://github.com/cloudflare/agents/blob/365ced36bab5f713f1669f3b79f6810ba1535729/docs/email.md",
      "framework": "cloudflare-agents",
      "calibrationMode": "source_pinned_contract",
      "cases": [
        {
          "id": "cloudflare-email-receive",
          "observedName": "onEmail",
          "eventType": "tool.completed",
          "expectedCapabilityId": "messaging.read",
          "rationale": "The email handler receives an inbound message for inspection."
        },
        {
          "id": "cloudflare-email-raw",
          "observedName": "getRaw",
          "eventType": "tool.completed",
          "expectedCapabilityId": "messaging.read",
          "rationale": "Reading the raw email body is a message read."
        },
        {
          "id": "cloudflare-email-send",
          "observedName": "sendEmail",
          "eventType": "tool.completed",
          "expectedCapabilityId": "messaging.send",
          "rationale": "sendEmail creates an external communication side effect."
        },
        {
          "id": "cloudflare-email-reply",
          "observedName": "replyToEmail",
          "eventType": "tool.completed",
          "expectedCapabilityId": "messaging.send",
          "rationale": "A reply sends an external message."
        },
        {
          "id": "cloudflare-email-forward",
          "observedName": "forward",
          "eventType": "tool.completed",
          "expectedCapabilityId": "messaging.send",
          "rationale": "Forwarding sends message content to another recipient."
        }
      ],
      "status": "passed",
      "metrics": {
        "total": 5,
        "expectedKnown": 5,
        "expectedUnknown": 0,
        "exact": 5,
        "falseUnknown": 0,
        "misclassified": 0,
        "falseKnown": 0,
        "exactMatchRate": 100,
        "falseUnknownRate": 0
      }
    },
    {
      "id": "stripe-agent-toolkit",
      "category": "finance",
      "project": "Stripe Agent Toolkit and MCP",
      "repository": "https://github.com/stripe/ai",
      "sourceRef": "c29cd23cfd27830bf10961d58646a9fd127fa6df",
      "sourceUrl": "https://docs.stripe.com/mcp",
      "framework": "stripe-agent-toolkit",
      "calibrationMode": "source_pinned_contract",
      "cases": [
        {
          "id": "stripe-balance",
          "observedName": "retrieve_balance",
          "eventType": "mcp.tool.completed",
          "expectedCapabilityId": "finance.read",
          "rationale": "Balance retrieval reads financial state."
        },
        {
          "id": "stripe-payment-intents",
          "observedName": "list_payment_intents",
          "eventType": "mcp.tool.completed",
          "expectedCapabilityId": "finance.read",
          "rationale": "Listing payment intents reads financial records."
        },
        {
          "id": "stripe-search",
          "observedName": "search_stripe_resources",
          "eventType": "mcp.tool.completed",
          "expectedCapabilityId": "finance.read",
          "rationale": "Resource search is read-only."
        },
        {
          "id": "stripe-payment-link",
          "observedName": "create_payment_link",
          "eventType": "mcp.tool.completed",
          "expectedCapabilityId": "finance.pay",
          "rationale": "Creating a payment link initiates a payment-capable external workflow."
        },
        {
          "id": "stripe-refund",
          "observedName": "create_refund",
          "eventType": "mcp.tool.completed",
          "expectedCapabilityId": "finance.pay",
          "rationale": "A refund moves money and requires payment-level controls."
        },
        {
          "id": "stripe-finalize-invoice",
          "observedName": "finalize_invoice",
          "eventType": "mcp.tool.completed",
          "expectedCapabilityId": "finance.pay",
          "rationale": "Finalizing an invoice makes a payment obligation externally actionable."
        },
        {
          "id": "stripe-customer",
          "observedName": "create_customer",
          "eventType": "mcp.tool.completed",
          "expectedCapabilityId": "finance.manage",
          "rationale": "Creating a customer mutates a financial account resource but is not itself a payment."
        },
        {
          "id": "stripe-subscription",
          "observedName": "update_subscription",
          "eventType": "mcp.tool.completed",
          "expectedCapabilityId": "finance.manage",
          "rationale": "Updating a subscription changes a financial account resource."
        },
        {
          "id": "stripe-cancel",
          "observedName": "cancel_subscription",
          "eventType": "mcp.tool.completed",
          "expectedCapabilityId": "finance.manage",
          "rationale": "Cancelling a subscription is a controlled financial administration action."
        }
      ],
      "status": "passed",
      "metrics": {
        "total": 9,
        "expectedKnown": 9,
        "expectedUnknown": 0,
        "exact": 9,
        "falseUnknown": 0,
        "misclassified": 0,
        "falseKnown": 0,
        "exactMatchRate": 100,
        "falseUnknownRate": 0
      }
    }
  ],
  "controls": {
    "status": "passed",
    "metrics": {
      "total": 5,
      "expectedKnown": 0,
      "expectedUnknown": 5,
      "exact": 5,
      "falseUnknown": 0,
      "misclassified": 0,
      "falseKnown": 0,
      "exactMatchRate": 100,
      "falseUnknownRate": 0
    }
  },
  "results": [
    {
      "adapterId": "browser-use",
      "caseId": "browser-use-navigate",
      "expectedCapabilityId": "browser.navigate",
      "actualCapabilityId": "browser.navigate",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "browser-use",
      "caseId": "browser-use-search",
      "expectedCapabilityId": "browser.navigate",
      "actualCapabilityId": "browser.navigate",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "browser-use",
      "caseId": "browser-use-click",
      "expectedCapabilityId": "browser.interact",
      "actualCapabilityId": "browser.interact",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "browser-use",
      "caseId": "browser-use-input",
      "expectedCapabilityId": "browser.interact",
      "actualCapabilityId": "browser.interact",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "browser-use",
      "caseId": "browser-use-scroll",
      "expectedCapabilityId": "browser.interact",
      "actualCapabilityId": "browser.interact",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "openhands",
      "caseId": "openhands-bash",
      "expectedCapabilityId": "coding.execute",
      "actualCapabilityId": "coding.execute",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "openhands",
      "caseId": "openhands-ipython",
      "expectedCapabilityId": "coding.execute",
      "actualCapabilityId": "coding.execute",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "openhands",
      "caseId": "openhands-editor-view",
      "expectedCapabilityId": "coding.read",
      "actualCapabilityId": "coding.read",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "openhands",
      "caseId": "openhands-editor-create",
      "expectedCapabilityId": "coding.write",
      "actualCapabilityId": "coding.write",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "openhands",
      "caseId": "openhands-editor-replace",
      "expectedCapabilityId": "coding.write",
      "actualCapabilityId": "coding.write",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "langchain-sql-toolkit",
      "caseId": "langchain-sql-query",
      "expectedCapabilityId": "data.query",
      "actualCapabilityId": "data.query",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "langchain-sql-toolkit",
      "caseId": "langchain-sql-schema",
      "expectedCapabilityId": "data.query",
      "actualCapabilityId": "data.query",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "langchain-sql-toolkit",
      "caseId": "langchain-sql-tables",
      "expectedCapabilityId": "data.query",
      "actualCapabilityId": "data.query",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "langchain-sql-toolkit",
      "caseId": "langchain-sql-checker",
      "expectedCapabilityId": "data.query",
      "actualCapabilityId": "data.query",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "cloudflare-agents-email",
      "caseId": "cloudflare-email-receive",
      "expectedCapabilityId": "messaging.read",
      "actualCapabilityId": "messaging.read",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "cloudflare-agents-email",
      "caseId": "cloudflare-email-raw",
      "expectedCapabilityId": "messaging.read",
      "actualCapabilityId": "messaging.read",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "cloudflare-agents-email",
      "caseId": "cloudflare-email-send",
      "expectedCapabilityId": "messaging.send",
      "actualCapabilityId": "messaging.send",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "cloudflare-agents-email",
      "caseId": "cloudflare-email-reply",
      "expectedCapabilityId": "messaging.send",
      "actualCapabilityId": "messaging.send",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "cloudflare-agents-email",
      "caseId": "cloudflare-email-forward",
      "expectedCapabilityId": "messaging.send",
      "actualCapabilityId": "messaging.send",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "stripe-agent-toolkit",
      "caseId": "stripe-balance",
      "expectedCapabilityId": "finance.read",
      "actualCapabilityId": "finance.read",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "stripe-agent-toolkit",
      "caseId": "stripe-payment-intents",
      "expectedCapabilityId": "finance.read",
      "actualCapabilityId": "finance.read",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "stripe-agent-toolkit",
      "caseId": "stripe-search",
      "expectedCapabilityId": "finance.read",
      "actualCapabilityId": "finance.read",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "stripe-agent-toolkit",
      "caseId": "stripe-payment-link",
      "expectedCapabilityId": "finance.pay",
      "actualCapabilityId": "finance.pay",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "stripe-agent-toolkit",
      "caseId": "stripe-refund",
      "expectedCapabilityId": "finance.pay",
      "actualCapabilityId": "finance.pay",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "stripe-agent-toolkit",
      "caseId": "stripe-finalize-invoice",
      "expectedCapabilityId": "finance.pay",
      "actualCapabilityId": "finance.pay",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "stripe-agent-toolkit",
      "caseId": "stripe-customer",
      "expectedCapabilityId": "finance.manage",
      "actualCapabilityId": "finance.manage",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "stripe-agent-toolkit",
      "caseId": "stripe-subscription",
      "expectedCapabilityId": "finance.manage",
      "actualCapabilityId": "finance.manage",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "stripe-agent-toolkit",
      "caseId": "stripe-cancel",
      "expectedCapabilityId": "finance.manage",
      "actualCapabilityId": "finance.manage",
      "resolution": "adapter_rule",
      "outcome": "exact"
    },
    {
      "adapterId": "negative-controls",
      "caseId": "unknown-browser-biometric",
      "resolution": "unknown",
      "outcome": "exact"
    },
    {
      "adapterId": "negative-controls",
      "caseId": "unknown-coding-deploy",
      "resolution": "unknown",
      "outcome": "exact"
    },
    {
      "adapterId": "negative-controls",
      "caseId": "unknown-data-model",
      "resolution": "unknown",
      "outcome": "exact"
    },
    {
      "adapterId": "negative-controls",
      "caseId": "unknown-message-voice",
      "resolution": "unknown",
      "outcome": "exact"
    },
    {
      "adapterId": "negative-controls",
      "caseId": "unknown-finance-trade",
      "resolution": "unknown",
      "outcome": "exact"
    }
  ],
  "claims": [
    "Pinned public tool contracts map deterministically to the expected Witnora capability IDs.",
    "Expected-known cases are measured for false-unknown and semantic misclassification separately.",
    "Expected-unknown controls detect broad aliases that would otherwise inflate apparent coverage."
  ],
  "limitations": [
    "Source-pinned contract calibration is not a full runtime compatibility test of every framework release.",
    "Only browser-use includes an existing checked-in real-agent run; the other rows validate public tool interfaces.",
    "A semantic match proves classification only, not authorization, enforcement, outcome verification, or business correctness.",
    "The optional LLM suggestion provider is excluded from this deterministic score."
  ]
}