diff --git a/integrations/google-adk/README.md b/integrations/google-adk/README.md
index 692c0bb..a7c4c66 100644
--- a/integrations/google-adk/README.md
+++ b/integrations/google-adk/README.md
@@ -60,6 +60,7 @@ record = plugin.build_record(
     invocation_id,
     subject="spiffe://example.org/agent/research-bot",
     policy_bundle=open("policy.cedar", "rb").read(),
+    enforcement_mode="declared",
     workload_digest="sha256:...",
     data_class="internal",
     model_provider="google",
@@ -68,9 +69,11 @@ signed = sign_record(record, generate_key())
 plugin.discard(invocation_id)
 ```
 
-`enforcement_mode` defaults to `declared`: the policy is bound into the signed
-record, but Google ADK itself did not evaluate it. Override that value only when
-a separate enforcement layer actually evaluated the policy.
+`enforcement_mode` is required. A bare ADK run is `declared`: the policy is bound
+into the signed record, but Google ADK itself did not evaluate it. TRACE spec
+section 4.3 says `declared` MUST NOT be a default, so the caller states it. Pass
+another value only when a separate enforcement layer actually evaluated the
+policy.
 
 One plugin can observe concurrent invocations. It retains state by ADK
 invocation id until `discard()` is called, so long-running processes should
@@ -89,6 +92,6 @@ python -m pytest test_google_adk_interop.py -q
 The first suite exercises evidence construction without installing ADK. The
 second uses the released runner and checks success, tool failure, cancellation,
 concurrent invocations, payload exclusion, signed TRACE validation, and Level 0
-conformance for both the default `declared` policy mode and the optional
+conformance for both the `declared` policy mode and the optional
 externally enforced `advisory` path. `declared` needs `agentrust-trace-tests`
 0.5.1 or later; 0.5.0 rejects it with `TR-POL-002`.
diff --git a/integrations/google-adk/google_adk_to_trace.py b/integrations/google-adk/google_adk_to_trace.py
index 5f838fa..6897d3e 100644
--- a/integrations/google-adk/google_adk_to_trace.py
+++ b/integrations/google-adk/google_adk_to_trace.py
@@ -7,9 +7,10 @@
 and responses, tool arguments and results, and exception messages are
 deliberately never read into evidence.
 
-Google ADK does not evaluate a TRACE policy. Records therefore default to
-``policy.enforcement_mode: declared`` and ``appraisal.status: none``. They have
-no ``origin`` block because this is first-party, in-process observation.
+Google ADK does not evaluate a TRACE policy, so ``enforcement_mode`` has no
+default: a bare ADK run passes ``declared`` (TRACE spec section 4.3 says
+``declared`` MUST NOT be a default). Records carry ``appraisal.status: none``
+and no ``origin`` block because this is first-party, in-process observation.
 """
 
 from __future__ import annotations
@@ -352,7 +353,7 @@ def build_record(
         *,
         subject: str,
         policy_bundle: bytes,
-        enforcement_mode: str = "declared",
+        enforcement_mode: str,
         workload_digest: str,
         data_class: str,
         model_provider: str | None = None,
@@ -409,7 +410,7 @@ def build_record(
     *,
     subject: str,
     policy_bundle: bytes,
-    enforcement_mode: str = "declared",
+    enforcement_mode: str,
     workload_digest: str,
     data_class: str,
     model_provider: str | None,
diff --git a/integrations/google-adk/test_google_adk_interop.py b/integrations/google-adk/test_google_adk_interop.py
index a29b62f..0a1a6eb 100644
--- a/integrations/google-adk/test_google_adk_interop.py
+++ b/integrations/google-adk/test_google_adk_interop.py
@@ -136,6 +136,7 @@ def build_signed(plugin: GoogleAdkTracePlugin, invocation_id: str) -> dict:
             invocation_id,
             subject="spiffe://example.org/agent/google-adk",
             policy_bundle=b'{"rules":["no-payload-egress"]}',
+            enforcement_mode="declared",
             workload_digest=DIGEST,
             data_class="confidential",
             model_provider="test-provider",
diff --git a/integrations/google-adk/test_google_adk_to_trace.py b/integrations/google-adk/test_google_adk_to_trace.py
index f7d410b..ca22980 100644
--- a/integrations/google-adk/test_google_adk_to_trace.py
+++ b/integrations/google-adk/test_google_adk_to_trace.py
@@ -39,6 +39,7 @@ def kwargs(**overrides):
     values = {
         "subject": SUBJECT,
         "policy_bundle": b'{"rules":["no-payload-egress"]}',
+        "enforcement_mode": "declared",
         "workload_digest": DIGEST,
         "data_class": "confidential",
         "model_provider": "test-provider",
@@ -396,6 +397,23 @@ def test_invalid_workload_digest_is_refused() -> None:
         )
 
 
+def test_enforcement_mode_has_no_default() -> None:
+    """TRACE spec section 4.3: declared MUST NOT be a default. The caller states it."""
+    values = kwargs(model_id="test-model")
+    del values["enforcement_mode"]
+    with pytest.raises(TypeError, match="enforcement_mode"):
+        build_record(**values, transcript=b"{}", tool_count=0)
+
+
+def test_declared_is_accepted_when_the_caller_states_it() -> None:
+    record = build_record(
+        **kwargs(model_id="test-model", enforcement_mode="declared"),
+        transcript=b"{}",
+        tool_count=0,
+    )
+    assert record["policy"]["enforcement_mode"] == "declared"
+
+
 def test_invalid_enforcement_mode_is_refused() -> None:
     with pytest.raises(MissingEvidence, match="enforcement_mode"):
         build_record(
diff --git a/integrations/langchain/README.md b/integrations/langchain/README.md
index b4b0bad..adc1e9f 100644
--- a/integrations/langchain/README.md
+++ b/integrations/langchain/README.md
@@ -50,22 +50,22 @@ agent.invoke({"input": "..."}, config={"callbacks": [handler]})
 record = handler.build_record(
     subject="spiffe://example.org/agent/research-bot",
     policy_bundle=open("policy.cedar", "rb").read(),
-    # enforcement_mode defaults to "declared"; see below
+    enforcement_mode="declared",  # required; see below
     workload_digest="sha256:...",
     data_class="internal",
 )
 signed = sign_record(record, generate_key())
 ```
 
-## `enforcement_mode` defaults to `declared`
+## `enforcement_mode` is required
 
-**LangChain enforces no policy.** It has no policy engine, and this handler is an observer that cannot block anything. So the default is `declared`: the policy is named and bound into the signed record, and nothing evaluated it.
+**LangChain enforces no policy.** It has no policy engine, and this handler is an observer that cannot block anything. So a bare LangChain run is `declared`: the policy is named and bound into the signed record, and nothing evaluated it. `declared` needs `agentrust-trace>=0.9`.
 
-That value did not exist when this adapter was written. `enforce`, `advisory` and `silent` all presuppose that *something evaluated the policy*, so the adapter refused to default the field and made the caller pick a value that overstated their run. TRACE 0.9.0 added `declared` for exactly this case, and it needs `agentrust-trace>=0.9`.
+The adapter does not pick the mode for you. `enforce`, `advisory` and `silent` all presuppose that *something evaluated the policy*, so defaulting to any of them claims an evaluation nobody observed, and TRACE spec section 4.3 says `declared` MUST NOT be a default. Leaving the argument out raises `TypeError`.
 
 | Value | When it is true here |
 |---|---|
-| `declared` | The default. The policy is named and bound; nothing evaluated it. |
+| `declared` | A bare LangChain run. The policy is named and bound; nothing evaluated it. |
 | `enforce` | Only when a real enforcement layer (cMCP, a policy proxy) sat in front of the tools. |
 | `advisory` | Only when something evaluated the policy and chose not to act on it. |
 | `silent` | Only when something enforced it with the operational logs suppressed. |
diff --git a/integrations/langchain/langchain_to_trace.py b/integrations/langchain/langchain_to_trace.py
index 79ac7a2..fec073b 100644
--- a/integrations/langchain/langchain_to_trace.py
+++ b/integrations/langchain/langchain_to_trace.py
@@ -15,15 +15,13 @@
 
 **The honest limit, stated where you cannot miss it.** LangChain enforces no
 policy. It has no policy engine, and this handler is an observer with no ability
-to block anything. ``enforcement_mode`` therefore defaults to ``declared``: the
-policy is named and bound into the signed record, and nothing evaluated it.
-
-That value did not exist when this adapter was written. The three original modes
-all assert that something evaluated the policy, so the adapter refused to default
-the field and made the caller choose a value that overstated their run. TRACE
-0.9.0 added ``declared`` for exactly this case. Override the default only when a
-real enforcement layer sat in front of the tools; a record claiming ``enforce``
-from a bare LangChain run describes enforcement that did not happen.
+to block anything. So ``enforcement_mode`` has no default and the caller states
+it. For a bare LangChain run that is ``declared``: the policy is named and bound
+into the signed record, and nothing evaluated it. TRACE spec section 4.3 says
+``declared`` MUST NOT be a default, and any other default would claim an
+evaluation nobody observed. Pass ``enforce`` only when a real enforcement layer
+sat in front of the tools; a record claiming ``enforce`` from a bare LangChain
+run describes enforcement that did not happen.
 
 Callback signatures are taken from ``langchain_core.callbacks.base`` and are the
 public, documented API. Payloads never enter the record: ``on_tool_start``
@@ -215,7 +213,7 @@ def build_record(
         *,
         subject: str,
         policy_bundle: bytes,
-        enforcement_mode: str = "declared",
+        enforcement_mode: str,
         workload_digest: str,
         data_class: str,
         model_provider: str | None = None,
@@ -225,9 +223,9 @@ def build_record(
     ) -> dict[str, Any]:
         """Assemble the unsigned Trust Record for this run.
 
-        ``enforcement_mode`` defaults to ``declared`` because LangChain itself
-        evaluates no policy. Override it only when a separate enforcement layer
-        actually evaluated the declared bundle. See the README.
+        ``enforcement_mode`` is required. LangChain itself evaluates no policy,
+        so a bare run passes ``declared``; pass another mode only when a
+        separate enforcement layer actually evaluated the bundle. See the README.
 
         ``attestation`` is ``{"platform": ..., "measurement": ...}`` when the
         deployment runs in a TEE, which lifts the record to Level 1. Absent, the
@@ -257,7 +255,7 @@ def build_record(
     *,
     subject: str,
     policy_bundle: bytes,
-    enforcement_mode: str = "declared",
+    enforcement_mode: str,
     workload_digest: str,
     data_class: str,
     model_provider: str | None,
@@ -284,9 +282,9 @@ def build_record(
         )
     if enforcement_mode not in ENFORCEMENT_MODES:
         raise MissingEvidence(
-            f"enforcement_mode must be one of {', '.join(ENFORCEMENT_MODES)}. The default "
-            "is 'declared', which is what a LangChain run actually is: the policy is named "
-            "and bound, and nothing evaluated it."
+            f"enforcement_mode must be one of {', '.join(ENFORCEMENT_MODES)}. A bare "
+            "LangChain run is 'declared': the policy is named and bound, and nothing "
+            "evaluated it."
         )
     if not model_provider or not model_id:
         raise MissingEvidence(
diff --git a/integrations/langchain/test_langchain_to_trace.py b/integrations/langchain/test_langchain_to_trace.py
index 8516934..0c8b5dc 100644
--- a/integrations/langchain/test_langchain_to_trace.py
+++ b/integrations/langchain/test_langchain_to_trace.py
@@ -119,16 +119,17 @@ def test_transcript_is_order_sensitive() -> None:
 # --- refusals --------------------------------------------------------------
 
 
-def test_enforcement_mode_defaults_to_declared() -> None:
-    """TRACE 0.9.0 added the value that is actually true of a framework run.
-
-    Before it, the three modes all asserted that something evaluated the policy,
-    so this adapter refused to default the field and made the caller pick a value
-    that overstated their run.
-    """
+def test_enforcement_mode_has_no_default() -> None:
+    """TRACE spec section 4.3: declared MUST NOT be a default, and any other
+    default claims an evaluation nobody observed. The caller states the mode."""
     kwargs = _kwargs()
     del kwargs["enforcement_mode"]
-    record = _handler_with_two_tools().build_record(**kwargs)
+    with pytest.raises(TypeError, match="enforcement_mode"):
+        _handler_with_two_tools().build_record(**kwargs)
+
+
+def test_declared_is_accepted_when_the_caller_states_it() -> None:
+    record = _handler_with_two_tools().build_record(**_kwargs(enforcement_mode="declared"))
     assert record["policy"]["enforcement_mode"] == "declared"
 
 
diff --git a/integrations/langchain/test_langgraph_interop.py b/integrations/langchain/test_langgraph_interop.py
index c60e288..ea791bf 100644
--- a/integrations/langchain/test_langgraph_interop.py
+++ b/integrations/langchain/test_langgraph_interop.py
@@ -59,6 +59,7 @@ def test_langgraph_tool_run_emits_a_valid_trace_record() -> None:
         handler.build_record(
             subject="spiffe://example.org/agent/langgraph",
             policy_bundle=b'{"rules":["no-payload-egress"]}',
+            enforcement_mode="declared",
             workload_digest=DIGEST,
             data_class="confidential",
             model_provider="test-provider",
diff --git a/integrations/llamaindex/README.md b/integrations/llamaindex/README.md
index b904075..36f3dbb 100644
--- a/integrations/llamaindex/README.md
+++ b/integrations/llamaindex/README.md
@@ -50,6 +50,7 @@ async def run_with_record(
     unsigned = tracker.build_record(
         subject=subject,
         policy_bundle=policy_bundle,       # bytes of the declared policy
+        enforcement_mode="declared",       # required; nothing evaluated the policy
         workload_digest=workload_digest,   # digest of your artifact
         model_provider=model_provider,
         model_id=model_id,
@@ -133,9 +134,11 @@ retain `appraisal.status: none`. Signing binds the record to its signing key;
 it does not attest the observer, authenticate model-supplied tool identity,
 prove safe behavior, or establish hardware provenance or runtime integrity.
 
-`policy.enforcement_mode` defaults to `declared`: the caller's policy is named
-and hashed, but LlamaIndex has not evaluated or enforced it. Supplying another
-mode requires an actual external policy layer. Supplied attestation fields are
+`policy.enforcement_mode` is required. A bare LlamaIndex run is `declared`: the
+caller's policy is named and hashed, but LlamaIndex has not evaluated or
+enforced it. TRACE spec section 4.3 says `declared` MUST NOT be a default, so the
+caller states it. Supplying another mode requires an actual external policy
+layer. Supplied attestation fields are
 passed through by the existing record builder; this adapter does not verify
 them or independently establish Level 1 assurance.
 
diff --git a/integrations/llamaindex/llamaindex_to_trace.py b/integrations/llamaindex/llamaindex_to_trace.py
index e710a33..e2ce339 100644
--- a/integrations/llamaindex/llamaindex_to_trace.py
+++ b/integrations/llamaindex/llamaindex_to_trace.py
@@ -209,7 +209,7 @@ def build_record(
         *,
         subject: str,
         policy_bundle: bytes,
-        enforcement_mode: str = "declared",
+        enforcement_mode: str,
         workload_digest: str,
         data_class: str,
         model_provider: str | None = None,
@@ -244,7 +244,7 @@ def build_record(
     *,
     subject: str,
     policy_bundle: bytes,
-    enforcement_mode: str = "declared",
+    enforcement_mode: str,
     workload_digest: str,
     data_class: str,
     model_provider: str | None,
@@ -256,10 +256,11 @@ def build_record(
 ) -> dict[str, Any]:
     """Assemble the unsigned record. Raises rather than inventing a field.
 
-    ``enforcement_mode`` defaults to ``declared``: the policy is named and bound
-    into the signed record and nothing evaluated it, which is what a LlamaIndex
-    run is. TRACE 0.9.0 added that value; before it, every available value
-    overstated a bare run and this adapter refused to default the field.
+    ``enforcement_mode`` is required. A bare LlamaIndex run is ``declared``: the
+    policy is named and bound into the signed record and nothing evaluated it.
+    TRACE spec section 4.3 says ``declared`` MUST NOT be a default, and any
+    other default would claim an evaluation nobody observed, so the caller
+    states the mode.
     """
     if not _SUBJECT_RE.match(subject or ""):
         raise MissingEvidence(
@@ -274,9 +275,9 @@ def build_record(
         )
     if enforcement_mode not in ENFORCEMENT_MODES:
         raise MissingEvidence(
-            f"enforcement_mode must be one of {', '.join(ENFORCEMENT_MODES)}. The default "
-            "is 'declared', which is what a LlamaIndex run actually is: the policy is named "
-            "and bound, and nothing evaluated it."
+            f"enforcement_mode must be one of {', '.join(ENFORCEMENT_MODES)}. A bare "
+            "LlamaIndex run is 'declared': the policy is named and bound, and nothing "
+            "evaluated it."
         )
     if not model_provider or not model_id:
         raise MissingEvidence(
diff --git a/integrations/llamaindex/test_llamaindex_interop.py b/integrations/llamaindex/test_llamaindex_interop.py
index 4a01f92..142d56e 100644
--- a/integrations/llamaindex/test_llamaindex_interop.py
+++ b/integrations/llamaindex/test_llamaindex_interop.py
@@ -130,6 +130,7 @@ def signed_record(tracker, *, subject=SUBJECT):
         tracker.build_record(
             subject=subject,
             policy_bundle=b'{"declared-policy":true}',
+            enforcement_mode="declared",
             workload_digest="sha256:" + "a" * 64,
             data_class="internal",
             model_provider="local-test-double",
diff --git a/integrations/llamaindex/test_llamaindex_to_trace.py b/integrations/llamaindex/test_llamaindex_to_trace.py
index d2a22ea..613fea8 100644
--- a/integrations/llamaindex/test_llamaindex_to_trace.py
+++ b/integrations/llamaindex/test_llamaindex_to_trace.py
@@ -167,11 +167,17 @@ def test_transcript_is_order_sensitive() -> None:
 # --- refusals --------------------------------------------------------------
 
 
-def test_enforcement_mode_defaults_to_declared() -> None:
-    """TRACE 0.9.0 added the value that is actually true of a framework run."""
+def test_enforcement_mode_has_no_default() -> None:
+    """TRACE spec section 4.3: declared MUST NOT be a default. The caller states it."""
     kwargs = _kwargs()
     del kwargs["enforcement_mode"]
-    assert _handler().build_record(**kwargs)["policy"]["enforcement_mode"] == "declared"
+    with pytest.raises(TypeError, match="enforcement_mode"):
+        _handler().build_record(**kwargs)
+
+
+def test_declared_is_accepted_when_the_caller_states_it() -> None:
+    record = _handler().build_record(**_kwargs(enforcement_mode="declared"))
+    assert record["policy"]["enforcement_mode"] == "declared"
 
 
 def test_unknown_enforcement_mode_is_still_refused() -> None:
diff --git a/integrations/openai-agents/README.md b/integrations/openai-agents/README.md
index e4d7aa6..1098b54 100644
--- a/integrations/openai-agents/README.md
+++ b/integrations/openai-agents/README.md
@@ -19,14 +19,15 @@ would produce a worse description, not a safer one.
 Where the deployment runs inside a TEE, passing an attestation lifts the same
 record from Level 0 to Level 1 and nothing else about the call changes.
 
-## `enforcement_mode` defaults to `declared`
+## `enforcement_mode` is required
 
 The Agents SDK enforces no policy. Guardrails exist and can stop a run, but they
-are the operator's own code, not a policy engine evaluating a bundle. So the
-default is `declared`: the policy is named and bound into the signed record, and
-nothing evaluated it.
+are the operator's own code, not a policy engine evaluating a bundle. So a bare
+run is `declared`: the policy is named and bound into the signed record, and
+nothing evaluated it. TRACE spec section 4.3 says `declared` MUST NOT be a
+default, so the caller states it; leaving it out raises `TypeError`.
 
-Override it only when a real enforcement layer sat in front of the tools. A
+Pass another value only when a real enforcement layer sat in front of the tools. A
 record claiming `enforce` from a bare Agents SDK run describes enforcement that
 did not happen.
 
@@ -79,6 +80,7 @@ Runner.run_sync(agent, "...")
 record = processor.build_record(
     subject="spiffe://example.org/agent/support-bot",
     policy_bundle=open("policy.cedar", "rb").read(),
+    enforcement_mode="declared",
     workload_digest="sha256:...",
     data_class="internal",
     model_provider="openai",     # the span that reports it is not read; see above
diff --git a/integrations/openai-agents/openai_agents_to_trace.py b/integrations/openai-agents/openai_agents_to_trace.py
index 3bff5cb..cf92ba8 100644
--- a/integrations/openai-agents/openai_agents_to_trace.py
+++ b/integrations/openai-agents/openai_agents_to_trace.py
@@ -17,10 +17,12 @@
 **The honest limit, stated where you cannot miss it.** The Agents SDK enforces
 no policy of its own. Guardrails exist and can stop a run, but they are the
 operator's code, not a policy engine evaluating a bundle. So
-``enforcement_mode`` defaults to ``declared``: the policy is named and bound
-into the signed record, and nothing evaluated it. Override that only when a
-real enforcement layer sat in front of the tools. A record claiming ``enforce``
-from a bare Agents SDK run describes enforcement that did not happen.
+``enforcement_mode`` has no default and the caller states it. A bare Agents SDK
+run is ``declared``: the policy is named and bound into the signed record, and
+nothing evaluated it. TRACE spec section 4.3 says ``declared`` MUST NOT be a
+default. Pass ``enforce`` only when a real enforcement layer sat in front of
+the tools; a record claiming ``enforce`` from a bare Agents SDK run describes
+enforcement that did not happen.
 
 **Payloads never enter the record.** ``FunctionSpanData`` carries ``input`` and
 ``output``, and ``GenerationSpanData`` carries the full message list. None of it
@@ -46,6 +48,7 @@
     record = processor.build_record(
         subject="spiffe://example.org/agent/support-bot",
         policy_bundle=open("policy.cedar", "rb").read(),
+        enforcement_mode="declared",
         workload_digest="sha256:...",
         data_class="internal",
         model_provider="openai",
@@ -271,7 +274,7 @@ def build_record(
     tools: tuple[ToolCall, ...] = (),
     handoffs: tuple[tuple[str, str], ...] = (),
     agents: tuple[str, ...] = (),
-    enforcement_mode: str = "declared",
+    enforcement_mode: str,
     attestation: dict[str, str] | None = None,
     iat: int | None = None,
 ) -> dict[str, Any]:
@@ -289,9 +292,9 @@ def build_record(
         )
     if enforcement_mode not in ENFORCEMENT_MODES:
         raise MissingEvidence(
-            f"enforcement_mode must be one of {', '.join(ENFORCEMENT_MODES)}. The "
-            "default is 'declared', which is what a bare Agents SDK run is: the policy "
-            "is named and bound, and nothing evaluated it."
+            f"enforcement_mode must be one of {', '.join(ENFORCEMENT_MODES)}. A bare "
+            "Agents SDK run is 'declared': the policy is named and bound, and nothing "
+            "evaluated it."
         )
     if not model_provider or not model_id:
         raise MissingEvidence(
diff --git a/integrations/openai-agents/test_openai_agents_interop.py b/integrations/openai-agents/test_openai_agents_interop.py
index 1885c8b..64b9c4e 100644
--- a/integrations/openai-agents/test_openai_agents_interop.py
+++ b/integrations/openai-agents/test_openai_agents_interop.py
@@ -105,6 +105,7 @@ def test_a_real_run_produces_a_valid_record(processor) -> None:
     record = processor.build_record(
         subject=SUBJECT,
         policy_bundle=POLICY,
+        enforcement_mode="declared",
         workload_digest=WORKLOAD,
         data_class="internal",
         model_provider="openai",
@@ -132,6 +133,7 @@ def test_a_real_tool_call_reaches_the_transcript_without_its_arguments(processor
     record = processor.build_record(
         subject=SUBJECT,
         policy_bundle=POLICY,
+        enforcement_mode="declared",
         workload_digest=WORKLOAD,
         data_class="internal",
         model_provider="openai",
diff --git a/integrations/openai-agents/test_openai_agents_to_trace.py b/integrations/openai-agents/test_openai_agents_to_trace.py
index e14814e..fe29db0 100644
--- a/integrations/openai-agents/test_openai_agents_to_trace.py
+++ b/integrations/openai-agents/test_openai_agents_to_trace.py
@@ -83,6 +83,7 @@ def record(**kwargs) -> dict:
     base = dict(
         subject=SUBJECT,
         policy_bundle=POLICY,
+        enforcement_mode="declared",
         workload_digest=WORKLOAD,
         data_class="internal",
         model_provider="openai",
@@ -95,10 +96,23 @@ def record(**kwargs) -> dict:
 # --- honesty rules --------------------------------------------------------
 
 
-def test_enforcement_mode_defaults_to_declared() -> None:
-    """The SDK enforces no policy. Anything else would describe enforcement that
-    did not happen."""
-    assert record()["policy"]["enforcement_mode"] == "declared"
+def test_enforcement_mode_has_no_default() -> None:
+    """TRACE spec section 4.3: declared MUST NOT be a default, and any other
+    default describes enforcement that did not happen. The caller states it."""
+    base = dict(
+        subject=SUBJECT,
+        policy_bundle=POLICY,
+        workload_digest=WORKLOAD,
+        data_class="internal",
+        model_provider="openai",
+        model_id="gpt-5",
+    )
+    with pytest.raises(TypeError, match="enforcement_mode"):
+        build_record(**base)
+
+
+def test_declared_is_accepted_when_the_caller_states_it() -> None:
+    assert record(enforcement_mode="declared")["policy"]["enforcement_mode"] == "declared"
 
 
 def test_a_bare_run_carries_no_origin_block() -> None:
@@ -205,6 +219,7 @@ def test_no_payload_reaches_the_record() -> None:
         processor.build_record(
             subject=SUBJECT,
             policy_bundle=POLICY,
+            enforcement_mode="declared",
             workload_digest=WORKLOAD,
             data_class="internal",
             model_provider="openai",
@@ -264,6 +279,7 @@ def test_concurrent_runs_do_not_share_a_transcript() -> None:
     args = dict(
         subject=SUBJECT,
         policy_bundle=POLICY,
+        enforcement_mode="declared",
         workload_digest=WORKLOAD,
         data_class="internal",
         model_provider="openai",
@@ -282,6 +298,7 @@ def test_building_before_any_run_is_refused() -> None:
         TraceRecordProcessor().build_record(
             subject=SUBJECT,
             policy_bundle=POLICY,
+            enforcement_mode="declared",
             workload_digest=WORKLOAD,
             data_class="internal",
             model_provider="openai",