diff --git a/API-Reference.md b/API-Reference.md
index 1b6a249..5c29ffd 100644
--- a/API-Reference.md
+++ b/API-Reference.md
@@ -11,7 +11,8 @@ SecureChatCompletion(
     base_url: str = "https://api.nomyo.ai",
     allow_http: bool = False,
     api_key: Optional[str] = None,
-    secure_memory: bool = True
+    secure_memory: bool = True,
+    max_retries: int = 2
 )
 ```
 
@@ -21,6 +22,7 @@ SecureChatCompletion(
 - `allow_http` (bool): Allow HTTP connections (ONLY for local development, never in production)
 - `api_key` (Optional[str]): Optional API key for bearer authentication
 - `secure_memory` (bool): Enable secure memory protection (default: True)
+- `max_retries` (int): Number of retries on retryable errors (429, 500, 502, 503, 504, network errors). Uses exponential backoff. Default: 2
 
 ### Methods
 
@@ -73,10 +75,30 @@ A dictionary containing the chat completion response with the following structur
         "prompt_tokens": int,
         "completion_tokens": int,
         "total_tokens": int
+    },
+    "_metadata": {
+        "payload_id": str,
+        "processed_at": int,          # Unix timestamp
+        "is_encrypted": bool,
+        "response_status": str,
+        "security_tier": str,         # "standard", "high", or "maximum"
+        "memory_protection": dict,    # server-side memory protection info
+        "cuda_device": dict,          # privacy-safe GPU info (hashed identifiers)
+        "tpm_attestation": {          # TPM 2.0 hardware attestation (see Security Guide)
+            "is_available": bool,
+            # Present only when is_available is True:
+            "pcr_banks": str,         # e.g. "sha256:0,7,10"
+            "pcr_values": dict,       # {bank: {pcr_index: hex_digest}}
+            "quote_b64": str,         # base64-encoded TPMS_ATTEST (signed by AIK)
+            "signature_b64": str,     # base64-encoded TPMT_SIGNATURE
+            "aik_pubkey_b64": str,    # base64-encoded TPM2B_PUBLIC (ephemeral AIK)
+        }
     }
 }
 ```
 
+The `_metadata` field is added by the client library and is not part of the OpenAI API response format. See the [Security Guide](security-guide.md) for how to interpret and verify `tpm_attestation`.
+
 #### acreate(model, messages, **kwargs)
 
 Async alias for create() method.
@@ -92,13 +114,18 @@ The `SecureCompletionClient` class handles the underlying encryption, key manage
 ### Constructor
 
 ```python
-SecureCompletionClient(router_url: str = "https://api.nomyo.ai", allow_http: bool = False)
+SecureCompletionClient(
+    router_url: str = "https://api.nomyo.ai",
+    allow_http: bool = False,
+    max_retries: int = 2
+)
 ```
 
 **Parameters:**
 
 - `router_url` (str): Base URL of the NOMYO Router (must use HTTPS for production)
 - `allow_http` (bool): Allow HTTP connections (ONLY for local development, never in production)
+- `max_retries` (int): Number of retries on retryable errors (429, 500, 502, 503, 504, network errors). Uses exponential backoff. Default: 2
 
 ### Methods