BASE
冻结基线Frozen baseline
记录上线时的数据分布、版本与黄金集结果。Record launch-time distribution, version, and gold-set results.
01 ·
去年合身的尺码表，今年换了人群还能直接用吗01 · Does last year's size chart still fit after the customer mix
changes?
同一版本上线时通过，但六个月后输入语言、订单结构和模型版本都变了。旧黄金集分数还能代表今天吗？The version passed at launch, but six months later the language mix, order patterns, and model build have changed. Does the old gold-set score still represent today?
02推进月份并改变分布或模型版本，观察基线偏移、失败簇与重评触发线Advance Months and Change Distribution or Model Version to Reveal Baseline Shift, Failure Clusters, and Re-evaluation Triggers
时间轴 · 重评触发线timeline · re-evaluation trigger
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
Model and Data DriftSHIFT
分布偏移Distribution shift
输入类别比例变化会让旧样本代表性下降。Changing input-category proportions weakens old sample representativeness.
CLUSTER
失败簇Failure cluster
新增错误集中在哪类请求，比总平均更早报警。Where new failures cluster can warn earlier than an overall average.
TRIGGER
重评触发Re-evaluation trigger
偏移、失败簇或版本变化越线就重跑黄金集与金丝雀门。If shift, clusters, or version changes cross a line, rerun gold-set and canary gates.
04最小机制：主交互能逐项验证Minimal Mechanism: Verify Each Link in the Interaction
MECHANISM
re_evaluate = shift ≥ τd ∨ cluster ≥ τf ∨ version_changed
主交互把月份、分布偏移与模型版本并列；任一触发条件成立就亮起重评线，而不是等待总质量崩溃。The interaction tracks month, distribution shift, and model
version side by side. Any trigger lights the re-evaluation line
instead of waiting for overall quality to collapse.
推进时间Advance time分布 / 版本变化distribution / version changes
对比基线Compare baseline偏移与失败簇出现shift and clusters appear
越过触发线Cross trigger重跑评测与金丝雀rerun evals and canary
05推进到第 12 月，放大分布偏移并切换模型版本Advance to Month 12, Amplify Distribution Shift, and Change Model Version
主动制造失败Create a failure边界实验Boundary experiment
若监控只保留上线时的总分，它不会告诉你证据已经过期。触发线必须把变化转成重评动作。A launch-time total score cannot tell you that evidence has expired. Trigger lines must turn change into a re-evaluation action.
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.
✗ 模型文件没变就不会漂移。No model change means no drift.→ ✓ 输入分布与业务规则也会让结论失效。Input distribution and business rules can invalidate
conclusions too.
✗ 看总准确率就能发现所有漂移。Overall accuracy reveals every drift.→ ✓ 失败簇可能在平均分明显下降前出现。Failure clusters may appear before the average moves
visibly.
✗ 触发告警后只需确认已读。A drift alert only needs acknowledgement.→ ✓ 触发线应连接黄金集重评与金丝雀门。The trigger must connect to gold-set re-evaluation and canary
gates.
06 · 一句话带走06 · One line to keep