安全与质量分轴 ·
质量可以渐进,安全硬门不能被平均Safety-Quality Separation · Quality Can Be Graded; Hard Safety Gates
Cannot Be Averaged
same version → quality × safety map → independent thresholds →
hard-gate release decision
安全与质量分轴(Safety-Quality
Separation)将正常任务质量与攻击、越权等安全表现分别测量和判定。平均质量高不能抵消一次严重安全穿透。Safety-quality separation measures normal-task quality independently
from attacks, authorization, and other safety behavior. High average
quality cannot offset one severe safety penetration.
01 ·
一辆车舒适度满分,刹车失灵还能靠平均分上市吗01 · Can perfect comfort scores average away failed brakes?
同一版本质量从 82 分涨到 97
分,但越权安全门仍失败。综合平均分变高,发布结论应该改变吗?The same version improves from 82 to 97 on quality, but its
authorization gate still fails. Should a higher combined average
change the release decision?
×
02把同一版本放上质量×安全二维图,拖高质量并观察安全硬门保持独立Place One Version on a Quality × Safety Map and Raise Quality
Without Moving the Safety Gate
二维证据图 · 独立硬门2D evidence map · independent hard gate
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
Safety-Quality Separation
QUALITY
质量轴Quality axis
衡量正常任务是否有用、正确、稳定。Measures usefulness, correctness, and stability on normal
tasks.
SAFETY
安全轴Safety axis
衡量攻击、越权与泄密场景是否被阻断。Measures whether attacks, authorization violations, and leaks are
blocked.
HARD
硬安全门Hard safety gate
严重安全失败是布尔阻断,不进入平均分。A severe safety failure is a Boolean block, not an averaged
score.
VERSION
同版对照Same-version evidence
两轴必须标注相同构建与评测时间。Both axes must identify the same build and evaluation
time.
04最小机制:主交互能逐项验证Minimal Mechanism: Verify Each Link in the Interaction
拖动质量只改变横轴;安全分和越权/泄密硬门保持原状态,任一失败继续阻断。Dragging quality moves only the horizontal axis. Safety score and
authorization/leakage gates stay unchanged; any failure
continues to block.
同版测两轴Measure both axes画出当前点plot current point
分别过门Apply separate gates禁止平均抵消no averaging
任一硬门失败Any hard gate fails阻断发布block release
05把质量拉到 100 分,同时让越权硬门失败Raise Quality to 100 While Failing the Authorization Gate
主动制造失败Create a failure
边界实验Boundary experiment
二维分轴就是为了阻止漂亮平均分掩盖严重穿透。图上的点再靠右,也不能越过安全硬门。The two-axis design prevents a flattering average from hiding
severe penetration. A point farther right still cannot cross a
failed safety gate.
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.
✗ 质量 99 分能补回安全 59 分。Quality 99 can offset safety 59.→ ✓ 两轴分别判定,不做加权抵消。Judge the axes separately; do not trade one for the
other.
✗ 只要安全平均分过线就行。An average safety score is enough.→ ✓ 越权、泄密等硬门仍可单独阻断。Authorization and leakage hard gates can still block
independently.
✗ 不同版本的两轴可以拼成一个点。Axes from different versions can form one
point.→ ✓ 每个点必须来自同一版本与时间窗。Each point must come from the same version and time
window.
06 · 一句话带走06 · One line to keep
质量与安全必须在同一版本上分轴测量;质量提高不会改变安全门,任一硬门失败都不能发布。Measure quality and safety separately on the same version. Higher
quality does not change safety gates, and any hard-gate failure
blocks release.