diff --git a/docs/architecture/rfcs/monorepo-distribution-split-v0.md b/docs/architecture/rfcs/monorepo-distribution-split-v0.md
index 0623c0b194..41f6fc2e1d 100644
--- a/docs/architecture/rfcs/monorepo-distribution-split-v0.md
+++ b/docs/architecture/rfcs/monorepo-distribution-split-v0.md
@@ -2567,6 +2567,65 @@ R0/V10/R1, D2/D5/D6/D3, context/model/split-install, PostgreSQL host/wheel and
 full-cost/adoption requirements remain open. No main merge or host upgrade is
 included.
 
+### 2026-11-08 — Feedback validation without mutation-adapter loading
+
+Initial runtime `2c5212fdb15035c26c3e158420d67197d45715f2` makes the two existing
+retained-state mutation adapters in `loopx.feedback` load on demand. Pure
+feedback validation now works when those adapters are unavailable; the parent
+fails at module import. Reward mutation resolves both adapters before registry
+reads or writes, while the original guard and writer call sites still look up
+the current module binding. Adapter object identity, mid-call patch behavior,
+index-before-state lock ordering, TypeScript guard ownership, credential refusal
+and partial-write handling remain. A local compatibility map, module `__getattr__`
+and `__dir__`, and a computed `__all__` preserve all 50 original star-export
+names, including the later append/render functions. There is no new capability,
+provider, persisted vocabulary, decision owner or cross-request authority cache.
+
+The two relevant groups pass 291 feedback-policy and 48 adapter/write/recovery
+cases. The independent unavailable-adapter tests fail on the parent before the
+change; head preserves pure validation and rejects a missing writer before
+reward mutation. Focused mypy and Ruff pass. The production semantic advisory
+finds no supported new carriers; explicit test-file inclusion is outside that
+scanner's supported `loopx/` scope. Test refusal strings remain local fixtures.
+
+The rebuilt monolithic wheel first passes isolated installed-module ownership,
+then matches all 1,840 packaged tracked sources outside the checkout with
+`PYTHONPATH` cleared. The five source-only omissions remain; relative to the
+frozen runtime candidate, only `effect_runtime.py` and `feedback.py` bytes change.
+The existing dashboard Chat build supplies the required bundle after the first
+wheel preparation failed for its absence. Four new candidate source/wheel ×
+File/SQLite 101-Todo CLI lifecycles use default bytecode and pass the original
+artifact, completion-time, scope, four-receipt, one-debit, restart, refusal and
+replay oracles. Both managed native daemons exit in every case. These are new
+runtime correctness checks, not additional matched cost samples.
+
+A normal-cache profile of an already settled synthetic recovery dry-run is
+used only to inspect loading dependencies. Its 410-module closure and nested
+import/compile timings cannot establish default-provider performance, fresh
+authority behavior or lifecycle cost. The initial stdlib-shadowing profile
+filename failed before CLI execution and remains a preparation failure. No
+frozen sample or preceding attribution pair is repeated, expanded or relabeled.
+No new whole-path benefit, PostgreSQL host/wheel, model/attention or installed
+adoption qualification is claimed. Preserve the peer hard-cut hold and all
+earlier failures, I3 for at least one minor release, M1/#8011, full R0/V10/R1,
+D2/D5/D6/D3 and context/model/split-install/full-cost/adoption acceptance.
+The original task and independent exact-head review/readiness remain open;
+the next loading decision stays with the existing quota/context/refresh callers.
+No main merge or installed-host upgrade is included.
+
+During submission preparation, a maintainer merges #8302 at
+`481436a56b47220c7c43bd6e272c1fa59d5038fe`; that head excludes this feedback
+change. The two unsubmitted feedback/checkpoint commits move to a clean successor
+based on target main `0ed909880`, retaining identical feedback and focused-test
+bytes. The successor wheel matches 1,841 packaged tracked sources, with the same
+five omissions and installed ownership checked first. Its fresh source/wheel ×
+File/SQLite lifecycles pass all four original oracle sets and both native exits
+per case. On this changed integration baseline, 48 write/recovery and 265 shared
+feedback-policy cases, mypy and Ruff pass. These checks qualify the successor's
+changed baseline; the initial runs remain historical evidence. No performance
+sample is repeated or added. The successor still requires independent exact-head
+review/readiness; the original task's complete acceptance remains open.
+
 ## Appendix B: Decision log
 
 | Date | Decision | Owner / approval | Alternatives | Normative sections changed |
diff --git a/docs/architecture/rfcs/monorepo-distribution-split-v0.zh-CN.md b/docs/architecture/rfcs/monorepo-distribution-split-v0.zh-CN.md
index 0c3d5fec27..c50298df31 100644
--- a/docs/architecture/rfcs/monorepo-distribution-split-v0.zh-CN.md
+++ b/docs/architecture/rfcs/monorepo-distribution-split-v0.zh-CN.md
@@ -1993,6 +1993,55 @@ caller 的必需依赖闭包，再选一个完整可回退的加载改动并验
 context／model／split-install、PostgreSQL host／wheel、full-cost／adoption 要求
 继续开放；本轮不合入 main、不升级 host。
 
+### 2026-11-08 — 反馈校验不加载写入适配器
+
+初始 runtime `2c5212fdb15035c26c3e158420d67197d45715f2` 将 `loopx.feedback` 中
+两个既有 retained-state 写入适配器改为按需加载。适配器不可用时，纯反馈校验
+仍可执行；parent 会在模块导入时失败。Reward mutation 在 registry 读取或
+写入前解析两个适配器，原 guard 与 writer 调用位置仍查找当前模块 binding。
+保留适配器对象 identity、调用中途 patch 行为、先 index 后 state 的锁顺序、
+TypeScript guard owner、凭证拒绝及 partial-write handling。本地兼容映射、
+模块 `__getattr__`／`__dir__` 与 computed `__all__` 保留全部 50 个原 star-export
+名称，包括后面的 append／render 函数。不新增 capability、provider、持久词汇、
+规则 owner 或跨请求 authority cache。
+
+两个相关组分别通过 291 个反馈策略及 48 个适配器／写入／恢复用例。独立的
+适配器不可用用例在 parent 上失败；head 保留纯校验，并在 reward mutation 前
+拒绝缺失 writer。聚焦 mypy／Ruff 通过。Production semantic advisory 未发现
+支持范围内的新 carrier；显式纳入测试文件超出该 scanner 支持的 `loopx/` 范围。
+测试中的 refusal 字符串仍是本地 fixture。
+
+重建单体 wheel 先断言隔离安装的模块所有权，再在非 checkout 工作目录、清空
+`PYTHONPATH` 后核对全部 1840 个 packaged tracked source。相同五个 source-only
+omission 保留；相对冻结 runtime candidate，仅 `effect_runtime.py` 和
+`feedback.py` 字节变化。首次 wheel 准备因缺少 Chat bundle 失败后，既有 dashboard
+Chat build 提供所需 bundle。四条新的 candidate source／wheel × File／SQLite
+101-Todo CLI 生命周期使用默认 bytecode，全部通过原 artifact、completion time、
+scope、四份 receipt、一次 debit、restart、refusal 与 replay oracle。每条的两个
+managed native daemon 均退出。这是新 runtime 正确性检查，不是新增 matched
+cost 样本。
+
+正常缓存 profile 仅检查已结算 synthetic recovery dry-run 的加载依赖。其
+410-module 闭包及嵌套 import／compile timing 不能证明默认 provider 性能、fresh
+authority 行为或生命周期成本。初始 profile 文件名遮蔽 stdlib，在 CLI 前失败，
+仍保留为准备失败。冻结样本及上一轮归因 pair 不重跑、不扩充、不重标签。
+不新增 whole-path 收益、PostgreSQL host／wheel、model／attention 或 installed
+adoption 资格。保留 peer hard-cut hold、全部历史失败、至少一个 minor release
+的 I3、M1/#8011、完整 R0/V10/R1、D2/D5/D6/D3 及
+context／model／split-install／full-cost／adoption 验收。原 Task 与独立 exact-head
+评审／readiness 继续开放；下一段加载决策仍属于既有 quota／context／refresh
+caller。本轮不合入 main，不升级 installed host。
+
+提交准备期间，维护者在 `481436a56b47220c7c43bd6e272c1fa59d5038fe` 合入
+#8302，该 head 未包含本次反馈改动。两个尚未提交远端的反馈／checkpoint commit
+移到基于目标 main `0ed909880` 的干净后续分支，反馈源码及聚焦测试字节相同。
+后续 wheel 先验证安装所有权，再匹配 1841 个 packaged tracked source，保留
+相同五个 omission。新的 source／wheel × File／SQLite 生命周期全部通过四组
+原 oracle，每条两个 native 均退出。在此变化的集成基线上，48 个写入／恢复及
+265 个共享反馈策略用例、mypy／Ruff 通过。这些检查验收后续提案的变化基线，
+初始运行保留为历史证据，不重跑或新增性能样本。后续提案仍需独立 exact-head
+评审／readiness，原 Task 的完整验收继续开放。
+
 ## 附录 B：决策日志
 
 | 日期 | 决策 | 负责人 / 批准 | 备选 | 变更的规范章节 |
diff --git a/loopx/feedback.py b/loopx/feedback.py
index fc78946de0..1b44e1339b 100644
--- a/loopx/feedback.py
+++ b/loopx/feedback.py
@@ -3,10 +3,9 @@
 from typing import Any, cast
 
 from contextlib import nullcontext
+from importlib import import_module as _import_module
 
 from .file_lock import exclusive_cross_runtime_file_lock
-from .control_plane.coordination.legacy_writer_fence import require_prose_state_write_allowed
-from .control_plane.runtime.document_io import atomic_write_state_text
 
 import re
 from datetime import datetime, timezone
@@ -24,6 +23,30 @@
 from .control_plane.actor_identity import normalize_owner_controller_actor
 
 
+_MUTATION_ADAPTERS = {
+    "require_prose_state_write_allowed": "loopx.control_plane.coordination.legacy_writer_fence",
+    "atomic_write_state_text": "loopx.control_plane.runtime.document_io",
+}
+
+
+def __getattr__(name: str) -> Any:
+    owner = _MUTATION_ADAPTERS.get(name)
+    if owner is None:
+        raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
+    value = getattr(_import_module(owner), name)
+    globals()[name] = value
+    return value
+
+
+def __dir__() -> list[str]:
+    return sorted(set(globals()) | set(_MUTATION_ADAPTERS))
+
+
+def _mutation_adapter(name: str) -> Any:
+    # Preserve the released module-level patch points and original objects.
+    return globals()[name] if name in globals() else __getattr__(name)
+
+
 GOAL_ID_RE = re.compile(r"^[A-Za-z0-9][A-Za-z0-9_.-]*$")
 
 
@@ -413,6 +436,10 @@ def append_human_reward(
     state_file_override: Path | None = None,
     write_active_state_summary: bool = False,
 ) -> dict[str, Any]:
+    # Resolve before registry reads or any append. An unavailable adapter must
+    # fail before the reward index can commit; TS still owns the write check.
+    _mutation_adapter("require_prose_state_write_allowed")
+    _mutation_adapter("atomic_write_state_text")
     actor = normalize_owner_controller_actor(actor_kind, required=not dry_run)
     validate_goal_id(goal_id)
     registry = load_registry(registry_path)
@@ -486,7 +513,7 @@ def append_human_reward(
                     original, str(active_state_update["entry"]),
                     updated_at=str(reward.get("recorded_at") or now_local()),
                 )
-                require_prose_state_write_allowed(
+                _mutation_adapter("require_prose_state_write_allowed")(
                     registry_path=registry_path, runtime_root=runtime_root, goal_id=goal_id,
                     state_path=state_file_to_write, original_text=original, planned_text=planned,
                 )
@@ -495,7 +522,7 @@ def append_human_reward(
             append_run_index_record(index_path, index_record)
             if state_file_to_write and state_text_to_write is not None:
                 try:
-                    atomic_write_state_text(state_file_to_write, state_text_to_write)
+                    _mutation_adapter("atomic_write_state_text")(state_file_to_write, state_text_to_write)
                 except OSError as error:
                     raise HumanRewardSummaryWriteError(str(error)) from error
                 active_state_update["written"] = True
@@ -622,3 +649,7 @@ def render_reward_markdown(payload: dict[str, Any]) -> str:
             ]
         )
     return "\n".join(lines)
+
+
+# Keep the historical star-import surface, including the deferred adapters.
+__all__ = [name for name in globals() if not name.startswith("_")] + list(_MUTATION_ADAPTERS)
diff --git a/tests/control_plane/test_feedback_adapter_loading.py b/tests/control_plane/test_feedback_adapter_loading.py
new file mode 100644
index 0000000000..29f3652148
--- /dev/null
+++ b/tests/control_plane/test_feedback_adapter_loading.py
@@ -0,0 +1,90 @@
+"""Pure feedback callers do not require retained-state mutation adapters."""
+from __future__ import annotations
+
+import subprocess
+import sys
+from pathlib import Path
+
+
+ROOT = Path(__file__).resolve().parents[2]
+
+
+def _isolated(code: str) -> subprocess.CompletedProcess[str]:
+    return subprocess.run(
+        [sys.executable, "-c", code], cwd=ROOT,
+        capture_output=True, text=True, check=False,
+    )
+
+
+def test_text_validation_remains_available_without_mutation_adapters() -> None:
+    result = _isolated('''
+import importlib.abc
+import sys
+
+class UnavailableWriters(importlib.abc.MetaPathFinder):
+    def find_spec(self, fullname, path, target=None):
+        if fullname in {
+            'loopx.control_plane.coordination.legacy_writer_fence',
+            'loopx.control_plane.runtime.document_io',
+        }:
+            raise ImportError('retained-state mutation adapters unavailable')
+
+sys.meta_path.insert(0, UnavailableWriters())
+from loopx.feedback import validate_local_control_text, validate_public_safe_text
+validate_local_control_text('next_action', 'Read the current task and continue.')
+validate_public_safe_text('summary', 'The artifact passed independent validation.')
+try:
+    validate_local_control_text('next_action', 'Bearer ' + 'x' * 32)
+except ValueError:
+    pass
+else:
+    raise AssertionError('the existing credential refusal must remain')
+''')
+    assert result.returncode == 0, result.stderr
+
+
+def test_feedback_adapter_exports_retain_owner_identity_and_star_import() -> None:
+    result = _isolated('''
+from loopx import feedback
+from loopx.control_plane.coordination import legacy_writer_fence
+from loopx.control_plane.runtime import document_io
+
+namespace = {}
+exec('from loopx.feedback import *', namespace)
+assert namespace['append_human_reward'] is feedback.append_human_reward
+assert namespace['render_reward_markdown'] is feedback.render_reward_markdown
+assert namespace['require_prose_state_write_allowed'] is legacy_writer_fence.require_prose_state_write_allowed
+assert namespace['atomic_write_state_text'] is document_io.atomic_write_state_text
+assert feedback.require_prose_state_write_allowed is namespace['require_prose_state_write_allowed']
+assert feedback.atomic_write_state_text is namespace['atomic_write_state_text']
+assert 'require_prose_state_write_allowed' in dir(feedback)
+assert 'atomic_write_state_text' in dir(feedback)
+''')
+    assert result.returncode == 0, result.stderr
+
+
+def test_missing_writer_dependency_fails_before_reward_mutation() -> None:
+    result = _isolated('''
+import importlib.abc
+import sys
+from pathlib import Path
+
+class UnavailableWriters(importlib.abc.MetaPathFinder):
+    def find_spec(self, fullname, path, target=None):
+        if fullname == 'loopx.control_plane.runtime.document_io':
+            raise ImportError('mutation writer unavailable')
+
+sys.meta_path.insert(0, UnavailableWriters())
+from loopx.feedback import append_human_reward
+try:
+    append_human_reward(
+        registry_path=Path('registry-does-not-exist.json'),
+        runtime_root_override=None, goal_id='synthetic-feedback',
+        run_generated_at=None, reward={}, actor_kind='owner',
+    )
+except ImportError as error:
+    assert str(error) == 'mutation writer unavailable'
+else:
+    raise AssertionError('an unavailable writer must refuse before opening the registry')
+''')
+    assert result.returncode == 0, result.stderr