@@ -156,7 +156,7 @@ The client manages a singleton connection to LaunchDarkly and the associated tel
 |---|---|
 | `init_client(options?)` | Auto-discovers and initializes `launchdarkly-server-sdk` (optional dep, loaded via `importlib`). Optional — the first AI API call triggers lazy init when `LD_SDK_KEY` is set. Accepts optional overrides for SDK key, base URIs, service name, environment, and OTLP endpoint. Returns `Awaitable[LDClientInterface]`. On every successful path, including the already-initialized path, flushes `$ld:ai:sdk:info` for any LaunchDarkly AI packages that have not yet reported. Does not take behavior defaults. |
 | `init_client(client=...)` | **BYOC overload** — accepts a pre-initialized `LDClientInterface`. Stores it directly without calling the SDK. Flushes pending `$ld:ai:sdk:info` events. |
-| `configure(*, options)` | Sets process-wide behavior defaults on the lifecycle singleton. `options` is a frozen `SDKOptions` (`capture_content: bool = False`). Keyword-only. Handlers read it at invocation. An explicit factory or wrapper `capture_content` overrides it. Replaceable until the first AI invocation; an identical value is allowed after. Not read from request data. |
+| `configure(*, options)` | Sets process-wide behavior defaults on the lifecycle singleton. `options` is a frozen `SDKOptions` (`capture_content: bool = True`). Keyword-only. Handlers read it at invocation. An explicit factory or wrapper `capture_content` overrides it, and an explicit `False` strips content everywhere, including offline evaluation spans. Replaceable until the first AI invocation; an identical value is allowed after. Not read from request data. |
 | `get_client()` | Returns the initialized `LDClientInterface`. Throws if initialization has not completed. |
 | `shutdown()` | Flushes all pending events and telemetry, then closes the client. Must be awaited before the process exits. Clears sdk-info reporting so a later client reports again. Restores the built-in SDK defaults. |
 | `register_ai_sdk_package(name, version)` | Records a LaunchDarkly AI package identity. Handler and convenience packages call this at import time. |
@@ -571,14 +571,15 @@ beside the input count, so its handlers add it in. OpenAI and LangChain already
 input, so theirs pass the figure through. Centralising that rule would double-count for two
 providers out of three. `SpanUsage` is the type that means the folding is already done.
 
-#### Content is off by default
+#### Content is on by default
 
 Every factory takes `capture_content: bool | None = None`. `None` reads `configure()` at
-invocation, which is off until a process sets `SDKOptions(capture_content=True)`. An explicit
-`True` or `False` overrides that default. Guard at the call site as well as inside the helper: the
-helper's guard makes a forgotten call site harmless, and the call site's guard avoids serialising
-JSON that would then be discarded, once per turn, in a loop. Resolve the bool when the handler
-runs, not when it is constructed.
+invocation. The built-in default is on until a process sets `SDKOptions(capture_content=False)`.
+An explicit `True` or `False` overrides that default, and an explicit `False` strips content
+everywhere, including offline evaluation spans. Guard at the call site as well as inside the
+helper: the helper's guard makes a forgotten call site harmless, and the call site's guard avoids
+serialising JSON that would then be discarded, once per turn, in a loop. Resolve the bool when the
+handler runs, not when it is constructed.
 
 #### Finish reasons have three mechanisms, not one
 
 
@@ -542,20 +542,23 @@ This matters most on Anthropic, which reports cache reads and writes beside the
 than inside it. A turn that reads 19,971 tokens from cache and writes 3,580 more reports an input of
 3 from the provider; the span reports 23,554.
 
-### Conversation content is off by default
+### Conversation content is on by default
 
-Prompts, model output, tool arguments and tool results are personal data, so a span carries only
-metadata unless you ask for more: models, token counts, timings, tool names.
+Prompts, model output, tool arguments and tool results are personal data. The built-in default
+sends that text. A span carries only metadata — models, token counts, timings, tool names — when
+capture is turned off.
 
-Pass `capture_content=True` to a handler factory to include the conversation, or turn it on for every handler from process startup. See [Process-wide defaults](packages/client/README.md#process-wide-defaults).
+Pass `capture_content=False` to a handler factory to leave the conversation off that handler, or
+turn it off for every handler from process startup. An explicit `False` wins everywhere, including
+on offline evaluation spans. See [Process-wide defaults](packages/client/README.md#process-wide-defaults).
 
 ```python
 from launchdarkly_ai_claude_messages import create_claude_messages_handler
 
-handler = create_claude_messages_handler(capture_content=True)
+handler = create_claude_messages_handler(capture_content=False)
 ```
 
-Turning this on sends the text of every request and response to whatever collector the SDK points
+Leaving this on sends the text of every request and response to whatever collector the SDK points
 at.
 
 ### Streaming
 
@@ -98,6 +98,7 @@ able to tell from the trace which path ran.
 | `launchdarkly.stream.abandoned` | `True`, only when abandoned | `end_span_once` |
 | `gen_ai.evaluation.name` | judge config key, judge roots only | `with_judge_evaluation`, see section 4a |
 | `gen_ai.evaluation.score.value` | numeric score, judge roots only | `with_judge_evaluation`, see section 4a |
+| `gen_ai.evaluation.explanation` | judge reasoning, judge roots only, gated by section 7 | `with_judge_evaluation`, see section 4a |
 
 The root also carries one span event, `feature_flag`, with these event attributes:
 
@@ -226,17 +227,18 @@ SDK writes a `gen_ai.evaluation.result` span event on that `invoke_agent` span:
 |---|---|
 | `gen_ai.evaluation.name` | judge config key |
 | `gen_ai.evaluation.score.value` | numeric score, only when the judge returned a finite number |
+| `gen_ai.evaluation.explanation` | the judge's reasoning, gated. See section 7 |
 
 The same keys are mirrored as span attributes, so section 2 lists them too.
 `gen_ai.evaluation.score.label` is not invented.
 The existing `track(evaluationMetricKey)` call is unchanged and still feeds AI Config Monitoring —
 a judge that returns a non-numeric score emits no evaluation event but still tracks the metric.
 
-`gen_ai.evaluation.explanation` is deliberately **not** emitted. The judge's reasoning is
-model-generated prose about the user's conversation — content, under section 7 — and content
-attributes require `captureContent` / `capture_content`, a handler-factory option this layer does
-not receive. The reasoning is still returned to the caller in `judgeResults` / `judge_results`;
-only the telemetry copy is withheld. Exporting it needs its own opt-in.
+`gen_ai.evaluation.explanation` is content under section 7. The judge's reasoning is
+model-generated prose about the user's conversation, so it follows the judge handler's own
+`capture_content` decision. It is written by default, and an explicit `False` on the judge's
+handler or the process default withholds it. The reasoning is always returned to the caller in
+`judgeResults` / `judge_results`; only the telemetry copy is gated.
 
 ---
 
@@ -397,12 +399,13 @@ error for an abandoned stream, so two dashboards would describe the same run dif
 Every handler factory takes `capture_content: bool | None = None`.
 
 `None` reads the process-wide default at invocation (`configure(options=SDKOptions(...))`), which
-is off. `True` or `False` overrides that default for the handler, and an explicit `False` wins
-over a configured `True`. A convenience wrapper's `capture_content` is the per-call override: when
-the caller passes it, that value is what the handler is built with; when they omit it, the wrapper
-passes `None` and the process default applies. Conversation content is personal data. A run emits
-only metadata, meaning models, token counts, timings and tool names, until capture is turned on.
-Turning it on sends the text of every request and response to whatever collector the SDK points at.
+is on. `True` or `False` overrides that default for the handler, and an explicit `False` wins
+over a configured `True`, including on offline evaluation spans. A convenience wrapper's
+`capture_content` is the per-call override: when the caller passes it, that value is what the
+handler is built with; when they omit it, the wrapper passes `None` and the process default
+applies. Conversation content is personal data. The default sends the text of every request and
+response to whatever collector the SDK points at. An explicit `False` emits only metadata, meaning
+models, token counts, timings and tool names.
 
 Guard twice, on purpose. The helper returns without writing when capture is off, which makes a
 forgotten call site harmless. The call site also checks before building the argument, which avoids
 
@@ -1,11 +1,13 @@
 """
-Example: enable conversation content for every handler at process startup.
-
-The handlers in the global registry are built with no ``capture_content``
-argument. They read the process default when they run. ``configure`` is
-separate from ``init_client``: ``main.py`` initializes the client, and this
-example calls ``configure`` once before the first AI invocation. An explicit
-``True`` or ``False`` on a factory would override this default.
+Example: pin conversation content on for every handler at process startup.
+
+The built-in default is already on. The handlers in the global registry are
+built with no ``capture_content`` argument and read that default when they
+run. ``configure`` is separate from ``init_client``: ``main.py`` initializes
+the client, and this example calls ``configure`` once before the first AI
+invocation so the choice is visible at startup. Pass
+``SDKOptions(capture_content=False)`` to opt out. An explicit ``False`` on a
+factory removes the transcript everywhere, including offline evaluation spans.
 
 Usage (via main.py):
     python main.py capture-content <flag-key> "<user input>"
 
@@ -9,10 +9,10 @@
      the same id: root, ``chat``, ``execute_tool``, and the judge's own ``invoke_agent``.
   2. Each turn shows a score badge, sourced from the ``gen_ai.evaluation.result`` span event on
      the judge span.
-  3. No judge reasoning anywhere in the telemetry. The score and the judge's config key are
-     exported; the explanation is not, because it is model prose about the user's conversation
-     and content attributes require ``capture_content``. The reasoning IS printed below, straight
-     from ``judge_results`` — that is the caller's copy, and it is unaffected.
+  3. The judge explanation is on the span, because capture is on by default. The score and
+     the judge's config key are exported either way. An explicit ``False`` removes the
+     explanation from telemetry. The reasoning IS printed below, straight from
+     ``judge_results`` — that is the caller's copy, and it is unaffected.
 
 The flag key must point at an AI Config with a ``judge_configuration``, otherwise there are no
 judge turns to look at.
 
@@ -40,7 +40,7 @@ asyncio.run(main())
 
 | Option | Default | Description |
 |---|---|---|
-| `capture_content` | `None` | Put prompts, model output, and tool arguments and results on the emitted spans. `None` follows `configure()`, which is off by default. |
+| `capture_content` | `None` | Put prompts, model output, and tool arguments and results on the emitted spans. `None` follows `configure()`, which is on by default. An explicit `False` leaves them off. |
 | `cwd` | the process's working directory | Working directory for the Claude Code process the Agent SDK starts. A `str` or path-like. |
 | `setting_sources` | `[]` (none) | Which Claude Code settings sources to load: `"user"` (`~/.claude/settings.json`), `"project"` (`.claude/settings.json`, and `CLAUDE.md`), `"local"` (`.claude/settings.local.json`). |
 
 
@@ -2,7 +2,7 @@
 
 > **Span shape.** This package emits `invoke_agent` → `chat {model}` → `execute_tool {name}`, with
 > tool spans as siblings of `chat`. Span construction lives in `spans.py` beside the handler.
-> Conversation content is off unless the caller passes `capture_content=True`, or the process sets it. See [Process-wide defaults](../client/README.md#process-wide-defaults).
+> Conversation content is on unless the caller passes `capture_content=False`, or the process turns it off. See [Process-wide defaults](../client/README.md#process-wide-defaults).
 > `TELEMETRY-CONTRACT.md` at the repo root is the authority; read it before changing span code.
 
 
 
@@ -739,8 +739,8 @@ def create_claude_agents_handler(
 
     Set *capture_content* to put prompts, model output, tool arguments and tool results on the
     emitted spans. ``None`` (the default) reads the process-wide default at invocation,
-    which is off until ``configure`` sets it. ``True`` or ``False`` overrides that default
-    for this handler. See TELEMETRY-CONTRACT.md section 7.
+    which is on until ``configure`` turns it off. ``True`` or ``False`` overrides that default
+    for this handler. An explicit ``False`` emits only metadata. See TELEMETRY-CONTRACT.md section 7.
 
     *cwd* and *setting_sources* configure the Claude Code process the Agent SDK starts, not the
     model, so they come from the application and never from a config. Each one given is passed to
 
@@ -862,7 +862,25 @@ async def test_cache_tokens_folded_into_input(
 
 
 class TestContentCapture:
-    async def test_emits_no_content_at_all_by_default(
+    async def test_emits_content_by_default(
+        self, monkeypatch: pytest.MonkeyPatch
+    ) -> None:
+        monkeypatch.setattr(
+            handler_mod,
+            "query",
+            _fake_query(
+                [
+                    assistant_message(content=[TextBlock(text="hi")]),
+                    result_message("hi"),
+                ]
+            ),
+        )
+        await create_claude_agents_handler()(BASE_CONFIG, "q")
+        [chat] = named("chat ")
+        assert "gen_ai.input.messages" in chat.attributes
+        assert "gen_ai.output.messages" in chat.attributes
+
+    async def test_emits_no_content_when_capture_content_is_false(
         self, monkeypatch: pytest.MonkeyPatch
     ) -> None:
         monkeypatch.setattr(
@@ -872,22 +890,22 @@ async def test_emits_no_content_at_all_by_default(
                 [init_message(tools=["Read"]), assistant_message(), result_message()]
             ),
         )
-        await create_claude_agents_handler()(TOOL_CONFIG, "q")
+        await create_claude_agents_handler(capture_content=False)(TOOL_CONFIG, "q")
         for span in (root(), *named("chat ")):
             assert "gen_ai.input.messages" not in span.attributes
             assert "gen_ai.output.messages" not in span.attributes
             assert "gen_ai.system_instructions" not in span.attributes
             assert "gen_ai.tool.definitions" not in span.attributes
 
-    async def test_root_carries_no_content_by_default_either(
+    async def test_root_carries_no_output_when_capture_is_false(
         self, monkeypatch: pytest.MonkeyPatch
     ) -> None:
         monkeypatch.setattr(
             handler_mod,
             "query",
             _fake_query([assistant_message(), result_message("hi")]),
         )
-        await create_claude_agents_handler()(BASE_CONFIG, "q")
+        await create_claude_agents_handler(capture_content=False)(BASE_CONFIG, "q")
         assert "gen_ai.output.messages" not in root().attributes
 
     async def test_tool_call_arguments_and_result_gated_by_capture(
@@ -925,7 +943,7 @@ async def _query(**kwargs: Any) -> AsyncIterator[Any]:
             yield result_message("done")
 
         monkeypatch.setattr(handler_mod, "query", _query)
-        await create_claude_agents_handler()(
+        await create_claude_agents_handler(capture_content=False)(
             TOOL_CONFIG, "q", {"search": lambda _: "r"}
         )
         [tool_span] = named("execute_tool ")
@@ -1610,7 +1628,11 @@ async def test_ld_span_attributes_land_on_root_only(
         assert attrs["launchdarkly.config.key"] == "cfg"
         assert attrs["launchdarkly.variation.key"] == "var"
         assert attrs["launchdarkly.run.id"] == "run-1"
-        assert [e.name for e in root().events] == ["feature_flag"]
+        assert [
+            event.name
+            for event in root().events
+            if not event.name.startswith("gen_ai.content")
+        ] == ["feature_flag"]
 
 
 class TestModelParametersForwarding:
 
@@ -2,7 +2,7 @@
 
 > **Span shape.** This package emits `invoke_agent` → `chat {model}` → `execute_tool {name}`, with
 > tool spans as siblings of `chat`. Span construction lives in `spans.py` beside the handler.
-> Conversation content is off unless the caller passes `capture_content=True`, or the process sets it. See [Process-wide defaults](../client/README.md#process-wide-defaults).
+> Conversation content is on unless the caller passes `capture_content=False`, or the process turns it off. See [Process-wide defaults](../client/README.md#process-wide-defaults).
 > `TELEMETRY-CONTRACT.md` at the repo root is the authority; read it before changing span code.