How Error Analysis Fails | Why Recording Mistakes Does Not Automatically Repair Them

Error analysis can fail even when the learner records every mistake. A correction book can be full. An error log can be beautifully organised. Every wrong answer can be labelled, highlighted and rewritten. Yet the same errors can return in the next worksheet, the next test and the next examination.

The reason is simple: recording an error is not the same as repairing the mechanism that produced it. An error log stores information about the past. Recovery requires a change that survives a fresh future task. Between those two states sit diagnosis, rule change, practice, retrieval, transfer, checking and fading support.

This guide is the cross-subject failure-mechanism owner inside the Learning Problems Hub. It does not replace subject-specific error-analysis guides such as PSLE English Error Analysis, Mathematics repair articles, or How to Translate a Bad Examination Result Into a Repair Plan. Those pages operate inside specific subjects or recovery tasks. This page owns the earlier diagnostic question: why does analysing mistakes sometimes fail to stop the mistakes?

The wider evidence base on metacognition and self-regulated learning helps explain why. The Education Endowment Foundation’s current Metacognition and Self-Regulated Learning guidance emphasises planning, monitoring and evaluating learning, with explicit teaching and modelling of strategies. An error list can support that process, but only if the learner uses it to change later decisions rather than treating the list itself as evidence of improvement.

---

An error is evidence, not a diagnosis

A wrong answer tells us that the final performance failed. It does not tell us why. The same wrong answer can arise from different mechanisms.

Visible error	Possible mechanisms underneath
Wrong Mathematics answer	Concept gap, algebra error, wrong method selection, transcription error, unit error, time-pressure shortcut, arithmetic slip, misread condition.
Weak comprehension answer	Misread demand, wrong evidence, insufficient inference, answer-scope problem, vocabulary misunderstanding, vague expression.
Weak Science explanation	Incorrect concept model, missing causal link, wrong variable relationship, keyword without mechanism, evidence not connected to claim.
Repeated grammar error	Rule unknown, rule known but not retrieved, cue not recognised, editing attention overloaded, correction not transferred.

If the learner labels all of these as “careless”, the error log has compressed several different mechanisms into one unusable category.

Failure mechanism 1: the category is too broad

“Careless mistake”, “concept error”, “language problem”, “didn’t understand” and “wrong method” may be true descriptions, but they are often too broad to change the next attempt.

A useful error category should imply a testable repair. Instead of “careless”, write “copied 0.45 as 0.54 after moving from rough work to answer line”. Instead of “weak evidence”, write “selected a sentence that describes the situation but does not support the inference asked”.

The narrower statement creates a place to intervene.

Failure mechanism 2: the learner records the error after seeing the answer

Once the correct solution is visible, the cause can look obvious. The learner writes “forgot to divide by 2” or “should have quoted paragraph 3”. But this retrospective explanation may not describe what actually happened during the attempt.

Whenever possible, capture the learner’s reasoning before showing the full solution. Ask: “What did you think the question wanted?” “Why did you choose this method?” “Where did you first feel uncertain?” “What did you check?”

This preserves diagnostic evidence that disappears after the correct answer has reorganised the problem.

Failure mechanism 3: the log records the last wrong step, not the first wrong step

Later errors can be consequences of an earlier break. A Mathematics solution may contain three wrong lines after the student chose the wrong representation. Correcting the final arithmetic misses the first decision that caused the chain.

In English, a weak paragraph may contain awkward expression because the central idea was never clear. Editing sentences does not repair the planning decision. In Science, a missing keyword may be downstream of an incorrect causal model.

Diagnostic rule: trace backwards until you reach the first point at which the learner’s reasoning diverged from a workable route.

Failure mechanism 4: the learner corrects the answer but not the decision rule

A corrected answer can be perfect and still fail to transfer. The learner may know “what the answer was” without knowing what future cue should trigger a different action.

Every useful error analysis should end with a rule of the form:

When I notice ______, I will ______ because ______.

For example: “When the question asks for an inference about attitude, I will select evidence that shows the attitude indirectly, then explain the connection, because repeating the literal statement does not answer the inference.”

Or: “When a ratio problem gives a changed total after adding the same amount to two groups, I will model before and after separately because the original ratio does not remain constant.”

Failure mechanism 5: the learner never attempts a fresh question

The original item is contaminated by memory. Once corrected, it can no longer prove that the learner has repaired the mechanism. A fresh question is required.

The fresh question should preserve the underlying decision while changing the surface. If the learner succeeds, the repair may be working. If the error returns, the log has documented the issue without changing it.

Failure mechanism 6: the fresh question is too similar

Near-copy questions can produce false transfer. A student learns to recognise a pattern of numbers, words or layout rather than the underlying condition.

After one close transfer question, vary the task more strongly. Remove topic headings. Change representation. Mix the repaired skill with nearby alternatives. Ask the learner to explain why the repaired rule applies before solving.

Failure mechanism 7: the learner logs every error equally

Not every wrong answer deserves the same repair investment. Some errors expose a high-dependency weakness that damages many tasks. Others are rare slips with little recurrence.

A useful error system prioritises by recurrence, dependency and mark impact.

Error type	Priority	Why
Recurring prerequisite failure	Very high	Likely to damage many later tasks.
Recurring method-selection error	High	Blocks transfer even when procedures are known.
Recurring paper-execution error	High near examinations	Converts existing capability into lost marks.
One-off unusual misconception	Investigate	May matter if it reveals a deeper model problem.
Random transcription slip with no recurrence	Lower	May need a light check rather than full reteaching.

Failure mechanism 8: the log becomes a museum of old mistakes

An error log can grow until it becomes impossible to use. Pages accumulate, categories multiply and the learner feels increasingly defined by past failure. The document stores history but no longer guides present action.

Retire repaired errors. Keep only active mechanisms, recent evidence and maintenance checks. A healthy error system should shrink as capability stabilises.

Failure mechanism 9: there is no distinction between knowledge, selection and execution

A learner can know a method but choose it at the wrong time. Another learner can choose correctly but execute inaccurately. A third can do both untimed and fail under examination pressure.

These are different errors.

• Knowledge error: required concept, fact, rule or procedure is missing or inaccurate.
• Retrieval error: knowledge exists but cannot be produced when needed.
• Selection error: learner knows several tools but chooses the wrong one.
• Execution error: correct route is chosen but carried out inaccurately.
• Representation error: information is encoded or translated incorrectly.
• Interpretation error: question demand or condition is misunderstood.
• Checking error: a detectable mistake survives because the checking system does not target it.
• Time-control error: knowledge is usable but the paper structure causes incomplete or rushed output.

An error log that collapses these into “wrong answer” cannot prescribe the next task accurately.

Failure mechanism 10: the repair never reaches delayed retrieval

A learner can correct an error immediately and lose the repair a week later. Same-day correction proves that the learner can follow the explanation while it is active. It does not prove durable availability.

Schedule a delayed check. It can be one or two questions. The important feature is that the learner is not warned immediately before the test by rereading the correction.

Failure mechanism 11: the repair is never mixed with nearby alternatives

Blocked practice makes the correct method obvious because every question uses the same method. Real examinations and real work are mixed. The learner must decide what the problem is before deciding what to do.

After the repair stabilises, mix the target with plausible alternatives. If the learner can still choose correctly, the repaired rule is becoming discriminative rather than merely rehearsed.

Failure mechanism 12: the error log is teacher-owned

If the adult always identifies the error, names the cause, chooses the category and assigns the repair, the learner can become very good at receiving diagnosis without learning to diagnose.

Gradually transfer ownership. Ask the learner to classify the error first. Compare the learner’s diagnosis with the teacher’s. Ask which fresh task would test the proposed repair. Ask what evidence would show the problem is gone.

This is where error analysis becomes metacognition rather than record keeping.

Failure mechanism 13: the learner treats all repeated errors as lack of effort

When a mistake repeats, adults and students often conclude that the learner “didn’t check” or “didn’t concentrate”. Sometimes that is true. Often the checking instruction is too vague.

A checking system must specify what to look for, when to look, and what signal should trigger the check. “Check your work” asks the learner to inspect the entire task with no model of likely failure.

Targeted checks are cheaper and more reliable: units after a measurement calculation, sign changes after moving terms, evidence relevance before finalising a comprehension answer, variable relationships before writing a Science explanation.

Failure mechanism 14: the learner repairs symptoms instead of dependencies

A student can accumulate many errors that all arise from one earlier weakness. Treating each as separate creates an unnecessarily large repair programme.

Cluster errors by dependency. If five Mathematics mistakes all require confident fraction operations, repair fractions. If several English tasks fail because the learner cannot distinguish literal evidence from inference, repair that distinction. If Science answers repeatedly miss cause-and-effect links, repair the conceptual model rather than each sentence individually.

This is why the Academic Recovery system begins with the first broken dependency.

The six-question error diagnosis

1. What exactly was wrong? Describe the observable mismatch.
2. Where did the reasoning first diverge? Find the earliest break, not the final wrong line.
3. What kind of error is it? Knowledge, retrieval, interpretation, representation, selection, execution, checking or timing?
4. What rule or capability must change? State the repair in operational language.
5. What fresh task will test the repair? Do not reuse only the corrected item.
6. When will we retest after delay? Durable learning requires time to pass.

A better error log template

Field	What to record
Task	Enough context to understand the original demand.
First break	The earliest point where reasoning stopped working.
Error mechanism	Knowledge, retrieval, interpretation, representation, selection, execution, checking, timing.
Changed rule	What the learner will notice and do differently.
Repair task	The smallest exercise that rebuilds the missing capability.
Fresh transfer	A new question requiring the same decision.
Delayed check	Date/result after time has passed.
Status	Active, improving, stable, or retired.

The last field prevents the log becoming a permanent archive of failure. Once the error is stable across fresh and delayed tasks, retire it.

English: when error analysis becomes answer collecting

English corrections often become model-answer collections. The learner copies a stronger comprehension answer or essay paragraph but never identifies the decision that made it stronger.

A better analysis distinguishes:

• Did I misunderstand the question?
• Did I select weak or irrelevant evidence?
• Did I fail to infer?
• Did I know the idea but express it vaguely?
• Did I write beyond the required scope?
• Did I use language inaccurately?
• Did I run out of time before checking?

Each category leads to different practice. The aim is not to own more model answers. It is to need fewer external corrections.

Mathematics: when error analysis becomes arithmetic policing

Students often circle the line where the number became wrong. But the deeper issue may be method selection, representation or a missing prerequisite.

Trace the solution from the question forward. Did the learner model correctly? Choose the right relationship? Translate words into equations? Manipulate accurately? Maintain signs and units? Check plausibility?

If the first break is early, fixing the arithmetic at the end will not prevent recurrence.

Science: when error analysis becomes keyword replacement

A Science learner may correct “heat” to “thermal energy”, “light” to “photosynthesis”, or add a missing term without repairing the explanatory relationship.

Ask what changed, what mechanism links the change to the result, and which evidence or condition matters. Error analysis should repair the model, not only the vocabulary.

Repeated errors are sometimes retrieval problems

A learner can understand a correction and still fail to retrieve it under pressure. This is common when the corrected rule is seen often but rarely produced from memory.

Turn the changed rule into a retrieval prompt. Ask the learner to state it before practice, after delay and inside mixed work. If the rule only appears when the correction book is open, it has not yet become available knowledge.

Repeated errors are sometimes cue-recognition problems

Other learners know the rule but do not notice when it is needed. The repair must therefore train cue recognition.

Compare examples and non-examples. Ask, “What feature of this question tells you that the rule matters?” Mix nearby alternatives. Have the learner identify the cue before solving.

Repeated errors are sometimes environment problems

A learner may stop making an error during calm tutoring and repeat it during timed school tests. The repaired capability has not transferred to the performance environment.

Recreate the relevant constraint gradually: less prompting, mixed questions, timed sections, longer sequences, realistic paper order. Recovery is complete only when the repaired rule survives the conditions in which it is needed.

Do not confuse a growing error log with worsening learning

When practice becomes harder or more varied, the learner may initially record more errors. That can be useful. A mixed set exposes weaknesses that blocked practice hid. Fresh transfer tasks reveal whether knowledge is portable.

Judge the system by the quality of the errors and whether active mechanisms are being repaired, not by the raw count of wrong answers.

Do not confuse a shrinking error log with mastery

The log can also shrink because the learner avoids difficult tasks. If all practice is familiar, fewer errors may mean less challenge rather than stronger learning.

Use representative fresh work. The right amount of difficulty generates useful evidence without overwhelming the learner.

The weekly repair review

1. Which error mechanism appeared most often?
2. Which error caused the greatest downstream damage?
3. Which repair succeeded on fresh questions?
4. Which repair survived delay?
5. Which error can be retired?
6. Which error needs a deeper prerequisite diagnosis?
7. What one checking rule should be installed next week?

The review should reduce the number of active priorities. A repair system becomes stronger when it knows what not to carry forward.

When to stop error logging

Error logs are temporary tools. They are not the curriculum and they are not a personality profile. Stop logging an error when the learner can use the repaired rule independently across fresh, varied and delayed tasks and can detect the risk without prompting.

At that point, move the skill into ordinary maintenance and let the learner’s attention return to current work.

When error analysis needs feedback repair too

Sometimes the learner’s error log fails because the feedback entering it is already too vague, too broad or not operational. In that case, start with How Feedback Fails.

When the learner understands the feedback but does not convert it into behaviour, use How to Translate Feedback Into Change.

When repeated errors are part of a broader plateau, use Why Marks Stop Improving.

When the examination is close and not every error can be repaired, use How to Decide What Not to Study When Time Is Short.

How to know error analysis is working

• Error categories become more precise.
• The learner identifies the first break rather than only the final wrong line.
• Corrections produce explicit changed rules.
• Fresh questions are used after correction.
• Repairs survive changed contexts.
• Delayed retests remain successful.
• Prompts and external diagnosis decrease.
• High-dependency errors disappear from multiple topics.
• The active error list shrinks because repaired mechanisms are retired.
• The learner begins to predict and prevent likely errors independently.

The purpose of error analysis is not to become expert at describing failure. It is to make the same failure less likely to happen again.

---

Error-analysis control line: wrong output → preserve original reasoning → find first break → classify mechanism → state changed rule → repair prerequisite if needed → fresh attempt → varied attempt → delayed retest → retire when stable. Return to the Recovery Atlas when the error pattern reveals a larger recovery problem.

The proof of good error analysis is not a longer log. It is a learner who repeats fewer mechanisms, diagnoses new failures more accurately and increasingly intercepts mistakes before submission.