📚 Why Teacher Assessment and Exam Assessment Produce Systematic Gaps | 教师打分与考试打分为何会有系统性差距
Every results season produces the same argument. Teachers insist that a student who has worked consistently for two years deserves a higher grade than a single three-hour paper suggests; examiners reply that only a common, externally marked instrument can compare one candidate with another. In 2020 and 2021 that argument stopped being theoretical, as centre assessed grades and teacher assessed grades replaced examinations across the UK and A-Level results rose sharply before falling back again. But the gap did not begin in 2020 and it did not end in 2022. It is a structural feature of two different measurement systems, and understanding it is essential for anyone preparing for A-Level examinations.
每到放榜季，同样的争论就会重演。老师坚持认为，一个持续努力两年的学生应该比一场三小时的考试得到更高的评价；考官则回应说，只有统一标准、外部评分的工具才能让考生之间可比。2020 和 2021 年，这场争论不再是纸上谈兵——英国用中心评估成绩和教师评估成绩取代考试，A-Level 成绩大幅上升，随后又回落。但这一差距并非始于 2020 年，也没有在 2022 年终结。它是两套不同测量体系的结构性特征，对每一位备考 A-Level 的学生来说都值得弄清。
1. What a "Systematic Gap" Actually Means | "系统性差距"的确切含义
A systematic gap is not the same as ordinary disagreement. If a teacher marks one script two marks above an examiner and another script two marks below, the errors cancel out and the average gap is roughly zero. That is random noise. A systematic gap is directional: it persists across a whole cohort, in the same direction, year after year. Statistically, if we let T be the teacher assessed mark and E the examination mark, the systematic component is simply the difference of the means:
系统性差距不等于普通的意见分歧。如果老师给一份答卷比考官高两分、给另一份低两分，误差互相抵消，平均差距接近零——这只是随机噪声。系统性差距是方向性的：它在一整届学生中持续存在、方向一致、年年如此。用统计语言说，设 T 为教师评估分数、E 为考试分数，系统性成分就是两个均值的差：
G = mean(T) − mean(E)
noise, and the noise shrinks as the sample grows. The standard error of a mean is SE = σ ÷ √n, so a centre with 20 candidates carries far more random variation than a national cohort of 800,000 entries. This is why individual class-level comparisons are almost meaningless, while national patterns are extremely stable.
观察到的差距永远是偏差与抽样噪声的混合体，而样本越大噪声越小。均值的标准误为 SE = σ ÷ √n，因此一个只有 20 名考生的中心，其随机波动远远大于 80 万人次的国家级样本。这就是为什么班级层面的个别比较几乎没有意义，而全国性的规律却极其稳定。
It also helps to separate three questions that are often confused. Accuracy asks whether the mean is right. Reliability asks whether the measurement is consistent, usually expressed through an intraclass correlation or Cohen's kappa:
此外，还应区分三个常被混为一谈的问题。准确性关注均值是否正确；信度关注测量是否稳定，通常用组内相关系数或 Cohen's kappa 表示：
κ = (p₀ − pₑ) ÷ (1 − pₑ)
Validity asks something deeper: does the instrument measure the construct we actually care about? A gap between teacher grades and exam grades may be evidence of teacher bias, examiner narrowness, or both. The number alone cannot tell you which.
效度则追问更深的问题：这个工具是否测量了我们真正关心的构念？教师分数与考试分数之间的差距，可能说明教师存在偏差，也可能说明考试过于狭窄，或者两者兼有。单看数字无法判断。
2. Two Instruments Measuring Different Constructs | 两种工具，测量两种不同的构念
An A-Level examination is a standardised, anonymised, time-limited sample of performance. Every candidate answers the same questions in the same window under the same conditions, and marks are awarded against a published scheme by examiners who do not know the candidate. Its great strength is comparability: a grade means roughly the same thing in Cornwall, Cardiff and Chengdu. Its weakness is sampling: a paper can only test a slice of a two-year specification, and it rewards performance on the day rather than performance over time.
A-Level 考试是一种标准化、匿名、限时的表现取样。所有考生在同一时间窗、同样条件下作答同一套题，由不认识考生的考官依据公开的评分方案给分。它最大的优势是可比较性：在康沃尔、加的夫和成都，同一个等级大致意味着同样的水平。它的弱点在于抽样：一份卷子只能覆盖两年课程规范的一部分，奖励的是"当天状态"而非"长期表现"。
Teacher assessment is the mirror image. It is longitudinal, evidence-rich and context-sensitive. A teacher sees the first draft, the correction, the oral explanation, the practical work, the improvement between October and March. That gives genuine access to constructs a timed paper struggles to reach: sustained project work, practical competence, the ability to act on feedback. But it is also local, unblinded and socially embedded, and the person judging is the person who taught, encouraged and sometimes predicted the grade.
教师评估恰恰相反。它是纵向的、证据丰富的、对情境敏感的。老师看到过初稿、修改稿、口头解释、实验操作，以及学生从十月到三月的进步。这使其能够触及限时笔试难以抵达的构念：长期项目、实操能力、根据反馈改进的能力。但它同时是局部的、非匿名的、嵌入社会关系之中的——评分者正是那个教导、鼓励并有时曾给出预测成绩的人。
Both instruments therefore suffer from construct under-representation, but in opposite directions. Examinations under-sample the long-run and the practical; teacher assessment under-samples the unseen and the independent. Any comparison between them is a comparison of two imperfect proxies, not of a proxy against a gold standard. That single fact explains a large share of the persistent gap.
因此，两种工具都存在"构念覆盖不足"，只是方向相反。考试低估了长期表现与实操能力，教师评估则低估了"面对陌生题目独立作答"的能力。二者之间的比较，是两个不完美代理指标之间的比较，而不是代理指标与黄金标准之间的比较。仅这一事实就解释了长期差距的很大一部分。
3. Assessment Objectives and the Limits of Examinability | 考核目标与"可考性"的边界
A-Level specifications are built around assessment objectives. Typically AO1 covers knowledge and understanding, AO2 application, AO3 analysis and evaluation, and AO4 skills such as practical technique or independent research. Written papers can reach AO1 to AO3 very efficiently. AO4 is far harder, which is why practical endorsements, non-exam assessment and coursework exist at all.
A-Level 的课程规范围绕考核目标（AO）构建。通常 AO1 考查知识与理解，AO2 考查应用，AO3 考查分析与评价，AO4 考查如实验技能或独立研究等能力。笔试能非常高效地覆盖 AO1 到 AO3，而 AO4 要难得多——这正是实践认证、非考试评估和课程作业存在的原因。
The proportion of non-exam assessment varies enormously between subjects, and so does the room for systematic divergence. Where a large share of marks comes from extended work judged inside the centre, teacher and examiner judgements have more space to drift apart.
非考试评估所占比例在各学科之间差异巨大，系统性偏差的空间也随之不同。当相当比例的分数来自中心内部评定的长期作业时，教师与考官的判断就有更大的空间发生分歧。
| Subject | 学科 | Non-exam assessment | 非考试评估占比 | Typical gap direction | 典型差距方向 |
|---|---|---|
| Mathematics, Further Mathematics | 0% | Smallest gap; right-or-wrong marking |
| Physics, Chemistry, Biology | Practical endorsement (separate) | Small to moderate |
| History, Geography, English | NEA in some specs | Moderate; essay judgement varies |
| Art and Design, Design and Technology | High (portfolio based) | Largest; holistic judgement |
The pattern is consistent: the more a grade depends on holistic judgement of sustained work, the larger the systematic gap between internal and external marking tends to be. This is not a moral failing of teachers or examiners; it is a direct consequence of what each instrument can see.
规律相当一致：一个成绩越依赖于对长期作业的整体判断，内部评分与外部评分之间的系统性差距往往越大。这既不是教师的道德缺陷，也不是考官的失职，而是两种工具各自"能看到什么"的直接结果。
4. Reliability: The Signal-to-Noise Ratio | 信度：信噪比
Classical test theory defines reliability as the proportion of observed variance that is true variance:
经典测量理论把信度定义为观测方差中属于真分数的比例：
r = σ²true ÷ (σ²true + σ²error)
Reliability limits the correlation any two measures can achieve. If measure T has reliability 0.85 and measure E has reliability 0.90, the maximum observable correlation between them is √0.85 × 0.90 that ceiling are not evidence that the two instruments disagree about the student; they are partly just measurement error. This attenuation effect is routinely ignored in public debate.
信度限制任意两个测量之间可达到的相关上限。若测量 T 的信度为 0.85、测量 E 的信度为 0.90，二者可观测相关最高只有 √0.85 × 0.90 ≈ 0.87。低于这一上限的观测相关，并不能证明两种工具对学生水平的判断不一致——其中一部分只是测量误差。这种衰减效应在公共讨论中常被忽略。
Exam marking is not perfectly reliable either. Examiner variability, borderline scripts, and mark schemes with room for interpretation all introduce error, and exam boards spend substantial sums on standardisation, seeding scripts and re-marking precisely because of this. Similarly, teacher assessment has well-documented error sources: the halo effect (one strong essay colouring everything else), order effects (the last portfolio seen seems best), evidence selection, and the single-marker problem in small departments where no second opinion exists.
考试评分同样并非完全可靠。考官之间的差异、临界卷子、以及留有解释空间的评分方案都会引入误差——各考试局之所以在标准化、种子卷和重评上投入巨资，正是因为这个原因。教师评估也有充分记录在案的误差来源：光晕效应（一篇出色作文染整份档案）、顺序效应（最后看到的作品集显得最好）、证据选择性，以及小型学科组中无人复核的"单人评分"问题。
So part of the gap is simply noise that has not averaged out, and part is genuine bias. The practical task for schools and boards is to reduce both, which is exactly what moderation is designed to do.
因此，差距的一部分只是尚未被平均掉的噪声，另一部分才是真正的偏差。学校与考试局的实际任务是把两者都压低——这正是调节（moderation）制度的设计初衷。
5. The Psychology of Teacher Judgement | 教师判断的心理学
Human judgement under uncertainty follows predictable heuristics. Teachers are not exempt. Anchoring occurs when a judgement is pulled towards a prior number, such as a target grade set in Year 12 or a school's historical average. Trajectory bias occurs when recent improvement is extrapolated into the future: a student who rose from a C to a B is often predicted to reach an A, even though improvement curves typically flatten. The halo effect ties a student's grade to their effort, behaviour or likeability, all of which are real qualities but not the construct being graded.
人在不确定条件下的判断遵循可预测的启发式规则，教师也不例外。锚定效应会让判断被先前的数字牵引，比如高二设定的目标等级或学校的历史均值。轨迹偏差则把近期进步外推到未来：从 C 升到 B 的学生常被预测能拿 A，尽管进步曲线通常会走平。光晕效应把成绩与努力程度、课堂表现或讨喜程度绑定——这些是真实品质，却不是被评分的构念。
A particularly important mechanism is best-performance versus typical-performance sampling. A teacher's mental model of a student is built from their better days: the moment the concept clicked, the excellent presentation, the insightful question. An examination samples whatever the student produces under pressure on one morning, which is drawn from a wider and lower distribution. Even with a perfectly unbiased teacher, this sampling asymmetry produces a positive mean gap.
一个尤其重要的机制是"最佳表现"与"典型表现"的取样差异。教师心中的学生形象由他们状态较好的时刻构成：某个概念突然打通的那天、一次出色的展示、一个有洞见的问题。而考试取样的，是学生在某个早晨压力下实际产出的东西，来自一个更宽也更低的分布。即便教师完全无偏，这种取样上的不对称也会产生正的均值差距。
Finally, regression to the mean works against the optimistic view. Extreme performances, whether very good or very bad, tend to be followed by more average ones. A teacher who has just seen a student's best ever mock paper is unusually likely to over-predict, purely because the next performance is statistically expected to be closer to that student's average.
最后，均值回归与乐观预期背道而驰。极端表现——无论极好或极差——之后往往跟随更接近平均的表现。刚看过学生史上最好一次模考的老师，特别容易高估，因为从统计上看，下一次表现更可能靠近该生的平均水平。
6. Institutional Incentives and Centre-Level Effects | 机构激励与中心层面的效应
Assessment does not happen in a vacuum. Schools are judged on results, and results affect league table positions, sixth-form recruitment, parental confidence and head-teacher careers. Departments are judged on the proportion of students reaching top grades. When a marginal decision arises, the surrounding incentives all push in one direction. This does not require conscious dishonesty; it is enough that ambiguous evidence is resolved in the candidate's favour slightly more often than not.
评估不是在真空中发生的。学校的声誉由成绩衡量，而成绩影响排名位置、预科招生、家长信心与校长职业前景；学科组则由高分比例被评价。当出现临界判断时，周围的激励全部指向同一方向。这并不需要刻意作弊——只要模糊证据被判为"偏向学生"的次数略多于"偏向严格"，效果就已经产生。
Centre-level effects are well documented. In the years when centre assessed grades were used, the increase in top grades was disproportionately large in independent schools and in centres with historically strong intakes, while predictions for disadvantaged students were less generous relative to their eventual performance under external assessment. Centre type, department size and prior attainment of the intake all predict the size of the optimism gap.
中心层面的效应有充分记录。在使用中心评估成绩的那些年份，高分比例的上升在私校和历史生源较强的中心尤为突出；而相对弱势学生后来在外部评估中的实际表现，其预测往往并不慷慨。中心类型、学科组规模与生源的既有水平，都能预测"乐观偏差"的大小。
This is the uncomfortable part of the story. A gap that looks like kindness at the individual level can function as unfairness at the system level, because two students of identical ability in different centres can receive different grades for the same work. Comparability, not generosity, is what a qualification is ultimately for.
这是整件事最令人不安的部分。在个体层面看起来像"善意"的偏差，在系统层面可能构成不公：两名能力完全相同但在不同中心就读的学生，可能因同样的表现得到不同的成绩。资格证书的最终价值在于可比性，而非慷慨。
7. What the Empirical Evidence Shows | 实证证据说明了什么
The strongest natural experiment in recent British assessment history is the 2020 to 2023 sequence, when the system moved from examinations to centre assessed grades, then to teacher assessed grades, then back to examinations. The proportion of A-Level entries awarded A* or A moved as follows (approximate reported figures).
英国评估史上最有力的"自然实验"是 2020 至 2023 年这段过程：体系从考试转向中心评估成绩，再转向教师评估
Published by TutorHao | Exam Prep Revision Series | aleveler.com
更多咨询请联系16621398022（同微信）
或截图后用「扫一扫 → 相册」选择本图