COPY
流量副本Traffic copy
按比例复制真实请求，不改变原请求去向。Copy a percentage of live requests without changing their destination.
01 ·
新司机能不能先坐副驾跑完全程，却不碰方向盘01 · Can a new driver ride the full route without touching the
wheel?
离线测试都过了，但真实请求里有你没见过的边界。怎样观察新版反应，同时保证客户仍只收到旧版结果？Offline tests pass, but live requests contain unseen edges. How do you observe the new version while guaranteeing customers still receive only the old result?
02调整影子流量比例，对比新旧输出并把差异案例送入审阅队列Adjust Shadow Traffic, Compare Old and New Outputs, and Queue Differences for Review
发布轨道 · 双车道release rail · parallel lanes
状态 AMetric A
—
—
状态 BMetric B
—
—
状态 CMetric C
—
—
所有概率、分数、成本、token 与输出均为固定种子的教学模拟，不代表真实模型输出。把主变量拖回即可复核结论。All probabilities, scores, costs, tokens, and outputs are fixed-seed teaching simulations, not real model output. Move the main variable back to verify the conclusion.
03把刚才的变化拆成 4 个零件Break the Change into Four Parts
Shadow ModePAIR
并排输出Paired output
同一请求的新旧结果必须能一一对应。Old and new results for the same request must remain paired.
EFFECT
旧版生效Old version effective
影子输出只记录，不能驱动工具、通知或业务状态。Shadow output is recorded only; it cannot drive tools, notices, or business state.
QUEUE
差异审阅Difference review
有意义的差异进入带请求 ID 的人工审阅队列。Meaningful differences enter a review queue with request IDs.
04最小机制：主交互能逐项验证Minimal Mechanism: Verify Each Link in the Interaction
MECHANISM
review_queue = {request | old(request) ≠ new(request)}
主交互按固定种子生成成对请求；只有差异项进入审阅队列，生效列始终指向旧版。The interaction creates fixed-seed request pairs. Only
differences enter review, while the effective column always
points to the old version.
复制流量Copy traffic新旧并行run old and new
计算差异Compute difference案例入队queue cases
隔离副作用Isolate effects仅旧版生效old only is effective
05把影子比例拉满并尝试让新版直接生效Set Shadow Traffic to 100% and Attempt to Make the New Version Effective
主动制造失败Create a failure边界实验Boundary experiment
影子比例可以很高，但“收到流量”不等于“拥有决策权”。越过隔离边界就不再是影子运行。Shadow share can be high, but receiving traffic is not decision authority. Crossing that boundary is no longer shadow mode.
尚未执行。先在主交互中观察正常机制。Not run yet. Observe the normal mechanism in the main interaction
first.
✗ 影子流量低就不需要旧版结果。A small shadow share makes old output
unnecessary.→ ✓ 每个影子样本仍需与旧版成对比较。Every shadow sample still needs a paired old
result.
✗ 影子输出可以悄悄写业务状态。Shadow output may silently write business
state.→ ✓ 副作用必须被技术隔离。Side effects must be technically isolated.
✗ 差异越少就一定能发布。Fewer differences guarantee release
readiness.→ ✓ 差异还要按质量、延迟与安全分类审阅。Differences still need quality, latency, and safety
review.
06 · 一句话带走06 · One line to keep