Secondary Mathematics Tuition | How to Stop Repeating Math Mistakes: Error Analysis and Self-Correction

Secondary Mathematics Tuition · Error analysis and self-correction for secondary learners

Mira has written x = 7 at the bottom of the page. The answer is wrong. Her tutor could circle the seven, write the correct answer and move on. Instead, he asks a different question: “Where did the mathematics first stop being true?”

They read upward. The final division is correct. The previous subtraction is correct. Two lines earlier, Mira expanded −3(2 − x) as −6 − 3x. That is the first invalid line. The error is smaller and more useful than the final wrong answer: one sign has been mishandled while distributing a negative factor.

To stop repeating mathematics mistakes, do not begin with the final wrong answer. Find the first wrong or unsupported decision, identify what kind of control failed, repair that relationship, and then prove the repair on a changed problem after the original solution is no longer visible. Self-correction becomes reliable when the learner can detect, explain and repair the error with progressively less external help.

This worldwide guide is written for secondary learners across school systems. It does not own Singapore SEC-specific error propagation, Secondary 3 Additional Mathematics diagnosis or examination post-mortem procedures. Those existing BTT pages keep their specialist roles. This article owns the transferable system for everyday mathematical error analysis and self-correction across topics.

The recurring learners Mira, Ben and Clara are fictional teaching characters. They illustrate possible mistakes and repair conversations, not real student histories or fixed ability labels. A learner who makes a sign error in one task may be highly secure in another.

Choose a route: if you keep correcting the answer but repeat the same mistake, start with the first wrong line. If you do not know what kind of error occurred, use the error families. If you want a routine for correcting your own work, go to the self-correction protocol. If mistakes appear only in tests, use errors under load. For independent practice, attempt the error-analysis checkpoint.

This guide connects to the wider BTT longform lane. Use How to Study Maths Effectively for the overall learning system, How to Understand Mathematics Instead of Memorising It for meaning and conditions, How to Solve Math Problems for complete route-building, and How to Revise for Maths for integrating repairs into a revision cycle.

1. The first wrong line is more useful than the final wrong answer

Consider Mira’s work on 5 − 3(2 − x). She writes 5 − 6 − 3x and simplifies to −1 − 3x. The final expression is wrong, but the last simplification is internally correct. If feedback says only “wrong answer”, it points to the least informative part of the page.

The first invalid line is the expansion. Multiplying −3 by −x should produce +3x. The corrected line is 5 − 6 + 3x, which simplifies to 3x − 1. That single diagnosis changes the repair task.

A numerical check confirms the distinction. Let x = 4. The original expression is 5 − 3(2 − 4) = 5 − 3(−2) = 11. The incorrect expression −1 − 3x gives −13. The corrected expression 3x − 1 gives 11. The substitution does not explain distribution by itself, but it independently rejects the wrong equivalence.

Now imagine giving Mira ten more expansion questions without identifying the sign issue. She may repeat the same rule ten times. Practice volume has increased while the error mechanism remains untouched. The page becomes thicker, not more reliable.

This is why correction should travel upward through the solution. Start at the answer and ask whether the last line follows from the previous one. If yes, move one line higher. Continue until the first transition that is false, unsupported or inconsistent with the problem’s conditions. That is usually where the repair belongs.

The first wrong line can be a missing line. Suppose Ben solves x² = 16 and writes x = 4. Nothing on the page is algebraically false if he intends to name one solution, but the problem asked for all real solutions. The missing branch x = −4 is the first incomplete decision. Error analysis must include omissions, not only false statements.

It can also be a wrong model before any algebra begins. Clara reads “A is five more than B” and writes B = A + 5. She then solves the resulting equation perfectly. The arithmetic is not the repair target. The relationship was reversed when translated from words.

Or the first wrong line can be an unjustified cancellation. From x(x − 6) = 0, dividing by x gives x = 6 and loses the solution x = 0. The division is valid only in the case x ≠ 0. The problem is not cancellation in general; it is cancellation without preserving the excluded case.

A learner who can find the first wrong line has begun to separate symptoms from causes. The final answer shows that something failed. The first wrong line shows where the mathematical contract was broken.

Use the focused page Study Wrong Solutions to Find the First Invalid Step for the narrow skill. This article takes the next step: once the first wrong line is found, what kind of repair should follow?

2. An error is an event in a solution, not an identity of the learner

“Careless”, “weak” and “bad at algebra” are poor error descriptions because they do not tell anyone what mathematical action should change. A useful diagnosis describes the event: copied 0.06 as 0.6; distributed a negative factor incorrectly; used final amount as the percentage base; assumed a diagram was drawn to scale.

This matters for self-correction. If Ben writes “careless” beside a wrong answer, he has recorded blame without a control. If he writes “I changed only one side of the equation on line three”, he can build a specific check for the next problem.

The same learner can make different error types in adjacent questions. Mira may understand percentages conceptually but copy a calculator value incorrectly. She may execute algebra accurately but choose the wrong equation for a word problem. A single label cannot explain both.

Likewise, the same visible wrong answer can have different causes. Suppose two students answer 24 for 20% of 150. One multiplied by 0.16 after a decimal-entry error. Another divided 150 by 5 incorrectly. A third found 30 correctly and then copied 24 from a previous line. Correction should follow the actual route, not the final number.

Error analysis is therefore local before it is global. Ask what happened in this solution. Only after several examples should you infer a recurring pattern. Four sign errors across different topics may justify a broader sign-control repair. One sign error may not.

Clara uses language that can be tested. Instead of “I panic in graphs”, she writes “I read the vertical scale as one unit per square when it was two”. The next graph can deliberately vary the scale to see whether the repair holds.

This does not deny that confidence, fatigue or pressure can affect performance. It simply keeps the mathematical description separate. “Under time pressure, I skipped the domain restriction” is more actionable than “I always fail under pressure”.

Self-correction grows when the learner can name a decision without turning it into a permanent self-description. “I used the wrong base in this percentage problem” invites a changed example. “I am terrible at percentages” offers no next move.

Tutors and parents can support the same distinction. Ask “Which line changed the meaning?” or “What quantity did this number represent?” rather than “Why were you careless again?” The first questions produce evidence. The second may produce defensiveness without mathematics.

An error event has a time, a location, a mathematical relationship and a possible repair. Treating it that way makes correction technical enough to improve.

3. Five broad error families make diagnosis faster

Many mathematics mistakes can be organised into five working families: representation, selection, execution, condition and completion errors. These are practical categories for this guide, not a universal psychological taxonomy. A single question can contain more than one.

Representation errors happen when the problem is translated incorrectly. “Three less than twice x” becomes 3 − 2x instead of 2x − 3. A diagram labels a sloping side as the perpendicular height. A table places cost where quantity should be.

Selection errors happen when the learner has relevant methods but chooses the wrong one. A percentage increase is treated as direct addition. A graph is assumed to show direct proportion merely because it is linear. An arithmetic-sequence formula is applied to a geometric sequence.

Execution errors occur after an appropriate route has been chosen. Sign mistakes, arithmetic slips, incorrect distribution, copied coefficients and calculator-entry mistakes belong here when the model and method were otherwise correct.

Condition errors lose or invent restrictions. Dividing by a variable without considering zero, accepting a negative length, using a right-triangle ratio without a right triangle, forgetting “without replacement”, or treating a rounded measurement as exact are examples.

Completion errors stop too early or answer the wrong requested quantity. One root is given when all are required. A radius is calculated but the question asks for diameter. A maximum is found but the answer is not converted to a whole-number count. Units are omitted where they carry meaning.

These families help because the repair differs. Representation errors need a better model or translation. Selection errors need contrasts and method cues. Execution errors need accurate practice and checking controls. Condition errors need explicit domains and boundary cases. Completion errors need target reading and final-answer discipline.

Consider the equation √(x + 5) = x − 1. A student squares both sides, solves the resulting quadratic and reports every algebraic root without checking the original. The first route may be appropriate. The error is condition/completion: squaring can create candidates that do not satisfy the original equation, so each must return to the starting relation.

In a word problem, a learner may first represent the situation incorrectly and then make a calculation error. Correcting only the arithmetic still leaves the model wrong. Find the earliest family in the chain. Later errors may disappear automatically once the earlier one is repaired.

Do not force every mistake into one box. “Used 60 minutes as 60 hours” could be representation of units, execution of conversion or both. The category is useful only if it clarifies the next repair.

Mira writes one letter beside an error: R, S, E, C or F for representation, selection, execution, condition or finish. Then she adds the actual sentence: “S — treated straight line as direct proportion despite nonzero intercept.” The sentence carries the real diagnostic value.

Across several weeks, the distribution of error families can reveal where revision should shift. If execution errors decline but selection errors persist, more blocked drill may be less useful than mixed comparison. If condition errors cluster, build domain checks into the revision sheet.

4. Representation errors are often invisible because the later algebra can be perfect

A correct calculation inside a wrong model is dangerous because it looks disciplined. The symbols line up, the arithmetic works and the final answer may even be plausible. Error analysis must therefore inspect what each variable and equation means, not only whether transformations are legal.

Suppose a taxi charges a fixed 8 units plus 3 units per kilometre. A 20-kilometre journey costs 68 units. A learner writes C = 3(8 + d). At d = 20, this gives 84. The algebra is easy; the model is wrong because the fixed charge has been multiplied by three as though it were part of the per-kilometre quantity.

The correct model is C = 8 + 3d. The distinction can be checked at d = 0. A fixed-charge model should still cost eight units. The incorrect equation gives 24. A boundary case reveals the representation error quickly.

Percentage problems produce similar mistakes. “After a 20% discount, the price is 96” means 0.8p = 96. Writing p − 0.2(96) = 96 uses the final amount as the discount base. The equation is not a bad rearrangement; it describes a different transaction.

Geometry representation errors often begin with a diagram. A learner sees a triangle with a visually central line and assumes it bisects the base. Unless symmetry, equal sides or another theorem justifies the bisection, the picture does not grant it.

Tables can misrepresent too. In a mixture problem, placing “4” and “7” in a row without labels can obscure whether they are unit costs, masses or total costs. A labelled table—mass, unit cost, contribution—prevents unlike quantities from being added.

To diagnose representation, ask the learner to read the equation back into words. If 0.8p = 96 is read as “eighty percent of the original price equals ninety-six”, does that match the story? If not, the model needs repair before any algebra continues.

Use units as another representation check. If kilograms are added to currency units, or square centimetres appear where length is requested, the model may have combined incompatible quantities. Dimensional consistency does not prove a formula correct, but inconsistency can expose a wrong one.

Clara draws only what the problem guarantees. Given parallel lines are marked; inferred equal lengths are not. Variables are defined in sentences before equations. This reduces the number of silent assumptions entering the work.

A representation error should be retested with a changed surface. After repairing a fixed-fee model in a taxi context, use a subscription, rental or delivery example. If the learner can still separate fixed and variable contributions, the relationship has begun to transfer.

The global problem-solving guide How to Solve Math Problems covers representation in complete solutions. Error analysis uses a narrower lens: when the answer is wrong, did the mistake enter before the mathematics was even represented correctly?

5. Selection errors happen when known mathematics is chosen for the wrong structure

A learner can know several methods and still choose the wrong one. Selection errors are common in mixed work because the topic heading no longer announces which tool to use.

Suppose Ben sees y = 4x + 7 and says it is direct proportion because it is a straight line. He remembers the graph shape but selects the wrong classification. Direct proportion in the ordinary school model has y = kx and passes through the origin. The +7 changes the relationship.

The repair is not another lesson on plotting straight lines. It is a contrast: y = 4x beside y = 4x + 7. Ask what stays the same—gradient four—and what changes—the intercept and constant ratio y/x. Selection improves by learning which feature decides the category.

Quadratics provide another example. A learner tries to factor x² − 6x − 2 using integer factors and spends five minutes searching. The problem is not inability to factor familiar quadratics. It is failure to recognise that integer factorisation is not available here. Completing the square or the quadratic formula is a better route.

In probability, adding event probabilities without checking overlap is a selection error. The addition rule for disjoint events is being applied where the events share outcomes. The learner knows addition but has not selected the correct event model.

Selection errors need comparative practice. Present near-neighbour problems that require different methods. Ask the learner to choose and justify before calculating. The answer can then be solved normally.

Use a two-column repair: “feature seen” and “method chosen”. For a right triangle with two sides known, the feature may suggest Pythagoras. For a non-right triangle, that method may be unavailable. The learner begins to associate methods with conditions rather than with superficial topic words.

Mira keeps one question above her mixed work: “What fact makes this method legal or useful here?” If she cannot answer, she delays the calculation long enough to inspect the structure.

Do not overcorrect by demanding lengthy explanations before every routine question. The goal is not to make selection slow. It is to make the deciding feature retrievable. With practice, the justification can become internal and rapid.

When a selection error repeats, study the alternative methods side by side. The BTT guide Compare Different Methods Until You Can Choose provides the focused route.

A self-correcting learner should eventually detect a selection error early: “This is becoming more complicated than expected; perhaps the route does not exploit the structure.” That metacognitive signal is valuable, but it must lead back to mathematics: which feature did the chosen method fail to use?

6. Execution errors need controls that operate while the mathematics is being done

Execution errors occur after the learner has chosen an appropriate model and method. They include arithmetic slips, sign errors, copied digits, bracket mistakes and calculator-entry failures. Because the route is otherwise sound, the temptation is to dismiss them as “careless”. That word does not improve the next solution.

Suppose Clara solves 3(2x − 5) = 21. She expands correctly to 6x − 15 = 21, adds fifteen and writes 6x = 34. The error is a simple addition slip: 21 + 15 should be 36. The repair is not a new lesson on distribution or equations. It is a control at the high-risk arithmetic line.

One control is reverse estimation. Before exact arithmetic, 21 + 15 is clearly above thirty-five and near thirty-six. The written thirty-four should look suspicious. Another is to substitute the final value back into the original equation. If Clara continues from 34 and obtains x = 17/3, substitution will fail.

Some execution errors recur because working is compressed. A learner may perform three sign changes mentally and write only the result. Expanding one extra line at the risky point can be faster overall than repeatedly repairing invisible mental steps.

For example, with 7 − 2(4 − 3x), write 7 − 8 + 6x before simplifying to 6x − 1. The intermediate line makes the two multiplications visible. Once the learner is reliably accurate, the working can become more compact again.

Copying errors need a different control. If 0.06 becomes 0.6, the learner may not need more decimal theory. They need a deliberate visual check when moving numbers between lines or from calculator to paper. Underline the decimal in one place, or compare the magnitude before proceeding.

Calculator-entry errors can be reduced by writing the mathematical expression first. If the intended quantity is (18 + 6)/3, entering 18 + 6 ÷ 3 produces a different result under standard order of operations. The machine has executed the input correctly; the translation into input failed.

Execution repair should stay narrow enough to preserve confidence in what was correct. If the equation model, method selection and algebraic structure were sound, say so. The learner’s next task is to control the arithmetic, not relearn the entire topic.

Ben keeps a personal “high-risk line” list: negative distribution, subtraction of negatives, decimal transcription and fraction denominator changes. He checks those lines more deliberately. The list can shrink when an error family stops recurring.

Do not add a check to every line indiscriminately. That would make routine mathematics unbearably slow. Place controls where recent evidence shows risk. A learner with stable decimal copying does not need a special decimal ritual merely because another learner does.

Execution accuracy also improves through fluency. If basic signed arithmetic is slow and effortful, sign errors can appear inside larger problems because too much attention is spent on the component. Targeted practice can reduce this load once the underlying meaning is sound.

The self-correction principle is therefore: preserve the correct route, add a control at the actual failure point, and retest in a fresh problem. Do not replace a local execution repair with a global judgement about mathematical ability.

7. Condition errors appear when a rule is used outside the circumstances that make it valid

Mathematics rules often carry conditions. Denominators must be nonzero. Right-triangle trigonometric ratios require the relevant right-triangle structure. Square-root expressions over the reals require nonnegative radicands. Probability models depend on assumptions such as replacement and equal likelihood.

A condition error happens when those requirements disappear from the learner’s working. The calculation may look familiar and even produce a neat answer.

Take x(x − 4) = 0. Dividing by x produces x = 4 and loses x = 0. The division is legal only in the branch x ≠ 0. The repair is not “never divide by x”. It is “division by a variable requires the divisor to be nonzero, so preserve any excluded zero case.”

Factorisation gives x(x − 4) = 0, hence x = 0 or x = 4. An alternative is to split cases: if x = 0, one solution is obtained; if x ≠ 0, division gives x = 4. Both methods retain the condition.

Absolute-value and square-root equations provide another example. Solving √(x + 6) = x by squaring gives x + 6 = x², with candidates 3 and −2. The original equation requires the right side to be nonnegative, and direct substitution rejects −2. Squaring produced a necessary consequence, not an automatically equivalent relation in both directions.

Geometry condition errors can be subtle. A learner uses Pythagoras because a triangle looks right-angled. If no right angle is given or proved, the theorem has been selected outside its condition. The repair is to label supplied facts separately from visual impressions.

In probability, “without replacement” changes the second denominator. Starting with five red and three blue counters, a red first draw leaves seven counters, not eight. Repeating 5/8 for the second red probability silently assumes replacement.

Condition errors benefit from attaching a short condition phrase to the method. “Cancel common factor, provided it is nonzero.” “Use tangent in a right triangle relative to the chosen angle.” “Use direct proportion when the ratio is constant and the model has zero intercept.”

These phrases need not appear in every final solution. They should be recoverable when a close counterexample appears. Revision can test that by placing valid and invalid cases side by side.

Mira writes conditions in the margin only when they are at risk. For a rational expression she writes x ≠ 3. For a parameter equation she separates a = 0 from a ≠ 0. For a geometry proof she marks the stated parallel lines before using angle properties.

A self-correcting learner asks not only “Do I know this rule?” but “What made it legal here?” This question prevents rules from becoming free-floating gestures.

When condition errors repeat across topics, the revision system should include boundary cases. Ask what happens when the denominator becomes zero, when the triangle is no longer right-angled, when the replacement condition changes, or when the parameter reaches the excluded value.

The goal is not fear of exceptions. It is controlled scope. A mathematical rule is powerful because it says exactly when a relationship is guaranteed.

8. Completion errors happen after most of the mathematics is already correct

Some mistakes occur at the final metre. The learner finds a radius when the question asks for diameter, gives one root instead of all roots, forgets units, rounds too early or fails to filter an algebraic candidate through the context.

These mistakes can be frustrating because the difficult mathematics may already be complete. They need a finishing protocol rather than a complete reteach.

Suppose a circle has area 49π square units. Solving πr² = 49π gives r = 7. If the question asks for diameter, the finished answer is 14 units. Writing 7 is not a failure of circle-area knowledge; it is a failure to return from the intermediate quantity to the target.

Or consider x² = 25. If the instruction is “solve over the reals”, x = ±5. If the expression is √25, the value is 5. A completion error can come from forgetting what kind of mathematical object the question requested.

Optimization requires a special final check. If the problem asks for a maximum whole number of boxes under a capacity, a real-valued inequality boundary may need conversion to an integer. If the problem asks for the minimum number of containers, round in the direction required by sufficiency rather than by ordinary nearest-integer convention.

Units are part of completion. Area is measured in square units; volume in cubic units; rates carry numerator and denominator units. A number without the requested unit may be mathematically incomplete even when the arithmetic is right.

Clara uses a target echo. Before solving, she writes a five-word target such as “find maximum whole number of vans”. At the end, she rereads that target. This catches situations where the algebra has found a raw quotient but the problem asks for a practical count.

Another control is the “all conditions used?” question. If a problem states that a length is positive and your final list contains a negative candidate, the condition has not yet been applied. If a problem provides two equations and the pair was checked in only one, the work may be incomplete.

Completion errors should be retested with the same mathematics but a changed final request. Find radius, then diameter. Find roots, then state the positive root only. Find exact value, then round to a stated precision. This trains attention to the endpoint rather than just the route.

Do not train completion by adding verbose closing sentences to every tiny exercise. Use explicit target checks where the question has several plausible stopping points or where the learner has a history of stopping one step early.

A strong final line names the quantity: “The minimum number of vans is 5,” not merely “5”. This reduces ambiguity and makes the problem-answer connection visible.

When self-correction reaches the end of a solution, ask four things: Is the mathematics valid? Is the answer allowed by the domain? Is it in the requested form and units? Is the requested set complete?

9. One early error can create many later wrong lines without creating many independent weaknesses

A solution can contain six wrong lines after one wrong decision. Counting each line as a separate error exaggerates the problem and can lead to unfocused revision.

Suppose Ben models a discounted price incorrectly as p − 0.2f = f, where f is the final price, instead of 0.8p = f. He then rearranges, substitutes and calculates perfectly. Every later numerical line may differ from the correct solution, but the chain has one primary cause: the wrong percentage base in the model.

This is error propagation. A wrong state feeds correct operations that preserve the wrong state. The repair should attack the earliest cause rather than every downstream consequence.

Algebra shows the same pattern. If Mira expands (x − 3)² as x² − 9, later factorisation, roots and graph features will all be wrong. The first missing term is −6x. Correcting the final roots without repairing expansion will not prevent recurrence.

Use a dependency question: “Would this later line become correct automatically if the earlier line were fixed?” If yes, it is probably downstream. If no, there may be a second independent error.

For example, after a wrong expansion, a learner may also make an unrelated arithmetic error. Repairing the expansion does not fix that second issue. Error analysis should therefore separate primary causes from independent secondary mistakes.

Draw arrows if necessary. Line 2 wrong model → line 3 valid rearrangement of wrong model → line 4 valid arithmetic → wrong answer. This map shows that only the first transition needs conceptual repair.

The SEC-specific page How SEC Mathematics Error Propagation Works owns the Singapore SEC implementation. The transferable principle is broader: do not confuse the size of the visible damage with the number of underlying causes.

Error propagation also explains why correcting from the bottom can waste time. If the denominator was copied wrongly three lines earlier, recomputing the final decimal does nothing. Move upward until the state first diverges.

In long problems, mark checkpoints. After finding an intermediate value, ask whether it is plausible before passing it forward. A wrong early value can contaminate several later stages. A ten-second checkpoint can contain the error before it propagates.

Clara labels intermediate quantities with meaning: “discounted price = 180”, “height = 12”, “remaining probability mass = 7/12”. Labels make it easier to notice when a later use changes the quantity silently.

Error containment is part of self-correction. The goal is not to guarantee zero mistakes. It is to make mistakes easier to locate, less likely to cascade and more likely to be repaired before the final answer.

10. Replay the solution as a sequence of claims when the cause is unclear

Some wrong solutions do not reveal the first error immediately. The working may be compressed, several steps may be mental, or the final answer may fail a check without showing why. A diagnostic replay reconstructs the reasoning one claim at a time.

Take a student’s solution to 2(x + 4) = 18: “2x + 4 = 18, 2x = 14, x = 7.” The final answer fails substitution because 2(7 + 4) = 22. Replay the first transition: does 2(x + 4) equal 2x + 4? No. Distribution should give 2x + 8.

The later subtraction and division are valid relative to the wrong equation. The replay identifies the first false equivalence.

Ask the learner to justify each arrow, not narrate every symbol. “Distributed two across the bracket.” “Subtracted eight from both sides.” “Divided both sides by two.” A weak justification often exposes the fragile step.

If the learner says “moved the four across”, ask what operation that phrase abbreviates. In some contexts it means subtracting four from both sides; in others a term lies inside a bracket or denominator and cannot be moved in the same way.

Diagnostic replay can use a numerical witness. Choose an easy input and compare two allegedly equivalent expressions. For 2(x + 4) and 2x + 4, x = 0 gives 8 and 4. One counterexample proves the expressions are not identical.

For geometry, replay means identifying the reason for each relationship. “These angles are equal because the lines are parallel.” “These sides are equal because the triangle is isosceles.” If the reason relies only on appearance, the claim is unsupported.

For a word problem, replay starts before the equations. “What does x represent?” “Why does this product represent cost?” “Why is the total twelve?” The first error may live in the model rather than the manipulation.

For a calculator-heavy question, replay the entered expression. The written mathematics may be correct while the machine input lacks brackets. Reconstruct the exact keystroke expression and compare it with the intended one.

Do not replay every correct line in every ordinary problem. Use the technique when a wrong result lacks an obvious cause, when working is compressed, or when the learner repeatedly says “I don’t know where I went wrong.”

Over time, the learner should internalise this process. A surprising answer triggers a backward scan: last claim, previous claim, first unsupported transition. Self-correction becomes faster because the learner knows what to inspect.

11. Study wrong solutions for the decision they reveal, not for entertainment

A wrong solution can be an excellent learning resource because it exposes a decision that a correct answer may hide. The point is not to stare at mistakes. It is to identify which line first breaks the mathematical relationship and why.

Consider the proposed simplification (x² − 9)/(x − 3) = x for x ≠ 3. The cancellation looks plausible because both numerator and denominator contain x and 3. Factor the numerator: (x − 3)(x + 3). The valid simplification is x + 3 for x ≠ 3. The wrong solution cancelled pieces of a sum rather than a common factor.

Ask the learner to repair only the first invalid step. Do not rewrite the whole problem immediately. “Factor the numerator before cancelling” is the decisive change. Then reconstruct the rest from that point.

A second wrong solution might say 1/2 + 1/3 = 2/5. The error reveals a false operation model: numerator and denominator are being added independently. Repair by expressing both fractions in sixths: 3/6 + 2/6 = 5/6. The common unit explains the addition.

Use close wrong solutions, not absurd ones. A mistake should be plausible enough that the learner must examine the mathematics. An obviously ridiculous answer trains recognition of teacher intention rather than discrimination of structure.

After diagnosing, ask for a counterexample to the wrong rule. If “cancel matching pieces of a sum” were valid, (2 + 3)/3 would simplify to 2. But 5/3 ≠ 2. A numerical witness destroys the false generalisation quickly.

Then ask for a case where cancellation is valid: 3(x + 2)/3 = x + 2. The contrast identifies the common-factor condition. Learners need both the failure case and the valid case so the repair does not become a new overgeneralised prohibition.

Clara keeps wrong solutions short. One or two lines are enough to test the idea. Long pages filled with invented errors can create unnecessary noise. The focus is the first invalid decision.

The existing BTT guide Study Wrong Solutions to Find the First Invalid Step develops the local skill. This global article uses wrong solutions as one component of a larger correction loop.

A wrong solution becomes useful when the learner can finish three sentences: “The first wrong line is…”, “It is wrong because…”, and “A changed problem that tests the repair is…”.

12. Correct one line first; do not erase the whole solution before learning from it

When a page contains an error, the instinct to cross out everything and restart can destroy diagnostic information. Preserve the original route long enough to see what was correct and where it first changed.

Suppose Ben solves 4(x − 2) = 3x + 5 as 4x − 2 = 3x + 5, then x = 7. The expansion is wrong because the factor four was not applied to the two. Everything after that belongs to the wrong equation.

Repair the line to 4x − 8 = 3x + 5. Continue from there: x = 13. Substitution confirms it: 4(13 − 2) = 44 and 3(13) + 5 = 44.

The visible correction shows which part of Ben’s original knowledge survived. He recognised expansion and equation solving as the right route; he mis-executed one distributive step. That is more precise than marking the whole solution “wrong”.

Preserving the page also helps identify multiple independent errors. If the corrected expansion still leads to a later arithmetic mistake, the second issue remains visible. A total rewrite can accidentally hide the fact that two controls are needed.

Use a repair symbol rather than full erasure. Circle the first invalid line, write the corrected line beside it and continue below. The page becomes a map of learning rather than a performance of cleanliness.

This matters especially in tutoring. If the tutor rewrites the entire solution, the learner may copy a perfect route without knowing which original decision changed. Repair should preserve enough of the student’s route to make the difference visible.

After the local correction, solve one fresh example from scratch. The fresh example answers a different question: can the learner produce the correct line without the old error sitting beside it?

Do not overvalue neatness in the learning draft. A revision page can contain arrows, corrections and notes. The final examination script may need clear working, but learning requires visible comparison between prior and corrected states.

Clara uses two colours only in private study: one for original work and one for repair. The colours are not necessary; the separation is. The important thing is to preserve what the learner originally did long enough to diagnose it accurately.

Correction becomes more efficient when the learner learns to patch the mathematics at the point of failure rather than repeatedly restarting from line one.

13. Compare near-miss problems so the learner can see which feature changes the answer

Some errors repeat because the learner has grouped two different problem structures together. A near-miss contrast holds most features constant and changes one decisive condition.

Compare 2x = 10 with x² = 10. The first has one real solution, x = 5. The second has two real solutions, x = ±√10. The changed operation changes the solution structure.

Compare √(x²) with solving y² = x². The first equals |x| over the reals; the second permits y = ±x. The square-root symbol and an equation involving squares do not ask the same question.

Compare a 20% increase followed by a 20% decrease with no change. Let the original be 100. Increase gives 120; decreasing 20% of 120 gives 96. The percentages act on different bases. The near-miss exposes why “equal percentages cancel” is false.

Compare a straight line y = 3x with y = 3x + 4. Both have gradient three. Only the first represents direct proportion under the usual model because only it has zero intercept and constant y/x for nonzero x.

Compare drawing two counters with replacement and without replacement. The first keeps the bag composition unchanged; the second changes the next probability. The visible story differs by two words, but the second-stage calculation changes.

A good contrast asks the learner to predict before calculating. “Which answer should change and why?” Prediction forces the decisive feature into working memory.

Then calculate both and explain the divergence. If the learner cannot explain it, the contrast has found a useful revision target.

Near-miss work should not become a list of traps. Its purpose is to build discrimination. The learner should leave with a clearer condition, not a feeling that mathematics is full of arbitrary exceptions.

Mira writes one sentence under each pair: “The changed feature is…” This can be brackets, domain, base quantity, replacement condition, graph intercept or requested target.

Later, remove the pairing. Place one problem inside a mixed set. The learner should still retrieve the deciding feature without the contrast sitting beside it.

The existing guide Compare Almost-Same Problems to Find What Changes the Method supplies focused practice. Error correction uses the same idea to prevent a repaired misconception from returning.

14. Build a counterexample when a wrong rule is too broad

A counterexample is one allowed case that makes a universal claim false. In error analysis, it is useful for dismantling overgeneralised rules quickly.

Suppose a learner says, “When the denominator gets bigger, the fraction gets smaller.” This can be true when the positive numerator is fixed, but it is false as an unrestricted claim. Compare 1/2 and 9/10. The second denominator is larger and the fraction is larger.

The repair is to state the condition: for a fixed positive numerator and positive denominators, increasing the denominator decreases the fraction. The counterexample does not destroy the useful local rule; it corrects its scope.

Another learner says, “Squaring and square roots always cancel.” Test x = −3. √(x²) = 3, not −3. The corrected real-number identity is √(x²) = |x|.

A geometry overgeneralisation might say, “Equal diagonals mean a square.” A nonsquare rectangle provides a counterexample. The learner then asks what additional properties are needed to establish a square.

Counterexamples should respect the claim’s domain. If the statement concerns positive integers, a negative decimal is irrelevant. Error analysis requires fair testing.

Ask the learner to generate the counterexample when possible. Constructing one requires understanding what the claim permits and which feature could break it.

A simple way to search is to try boundary cases, zero, one, negative values where allowed, equal values, extreme values or familiar special shapes. These choices often expose hidden assumptions.

After the claim is disproved, repair it. “What narrower statement would be true?” This step prevents counterexample practice from becoming pure destruction.

Ben revises “division makes smaller” by testing 6 ÷ 1/2 = 12. He then repairs the statement: dividing a positive number by a positive number greater than one makes the result smaller than the original; dividing by a positive number between zero and one makes it larger.

The self-correction value of a counterexample is that it gives the learner an independent test. A remembered slogan must survive contact with an allowed case.

Use counterexamples especially when a mistake looks like a rule rather than an arithmetic slip. One decisive case can redirect an entire pattern of future work.

15. Keep an error log that stores repair instructions, not a museum of wrong answers

An error log can become useful or burdensome. A useful log stores just enough information to change future work. A burdensome one becomes a second textbook made entirely of failures.

For each recurring or important error, record four things: the first wrong decision, the corrected relationship, the control that could catch it, and the date or condition for a fresh retest.

Example: “Reverse percentage — used final price as base. Correct model: final = retained proportion × original. Control: name the base before calculating. Retest: one mixed percentage question on Friday.”

Another: “Quadratic — lost second root after square. Correct relationship: z² = a gives z = ±√a for positive a. Control: ask whether operation was many-to-one. Retest: absolute-value or square equation next week.”

Do not copy the entire question unless context is necessary. A short reference or photo may be enough in personal notes, but the mathematical description should remain readable without hunting through several pages.

Classify only if classification helps. A small R/S/E/C/F code can reveal trends, but the sentence about the actual error matters more than the letter.

Retire entries. If a repair survives several fresh delayed tests under relevant conditions, move it out of the active list. A permanent page of every historical mistake can make progress invisible.

Review the active log before mixed practice or a mock. This primes checking controls: domain restrictions, sign lines, percentage bases or units.

After the mock, update the log only for meaningful patterns. One isolated arithmetic slip may need a brief note but not a full revision programme. Repeated or high-consequence errors deserve priority.

Clara limits the active log to a small number of current risks. When a new error enters, she asks whether an older one can move to maintenance. The constraint keeps the log actionable.

The error log should gradually shift from tutor-written to learner-written. Self-correction requires the learner to describe the error in their own mathematically accurate language.

A good log makes future mathematics easier to control. If it merely records embarrassment, redesign it.

16. Use a six-step self-correction protocol until the sequence becomes internal

A practical self-correction routine can be short enough to remember and specific enough to change the mathematics. Use six steps: detect, locate, classify, repair, retest and return.

Detect: notice evidence that something may be wrong. A failed check, impossible unit, negative length, probability above one, unexpected sign or disagreement between two methods can trigger the process.

Locate: identify the first wrong, unsupported or incomplete decision. Move upward through the work rather than beginning with the last line automatically.

Classify: decide whether the failure is mainly representation, selection, execution, condition or completion. The label is optional; the purpose is to choose an appropriate repair.

Repair: correct the relationship at the first weak point. Use a brief explanation, worked example, contrast, counterexample or targeted practice according to the error.

Retest: solve a changed problem without the original correction visible. If the same support is still needed, the repair is not yet independent.

Return: revisit the repaired idea later in mixed work. Immediate success tells you the explanation made sense. Delayed success provides stronger evidence that the correction remains available.

Suppose Ben writes 4/5 + 1/3 = 5/8. Detect: the result is suspicious because adding a positive fraction to four fifths should produce more than four fifths; five eighths is smaller. Locate: the first line itself uses the wrong addition model. Classify: representation/operation meaning. Repair: common denominator fifteen gives 12/15 + 5/15 = 17/15. Retest: 3/4 + 2/5. Return: a mixed fraction problem several days later.

The protocol is not meant to make every correction slow. At first, write the six words at the top of a practice page. With repetition, the learner may move through them mentally in seconds.

Different errors can skip steps when the diagnosis is obvious. If a copied coefficient is visibly wrong, classification may be unnecessary. The protocol is a scaffold, not a bureaucratic requirement.

Mira’s goal is to need the tutor less often for the first four steps. The tutor may still help with repair design, but Mira should increasingly detect and locate the issue herself.

A learner who can solve only after someone says “there is a sign error on line three” has received diagnosis. A learner who finds that line independently has performed self-correction. Both are useful stages; the direction of progress matters.

Use the protocol in ordinary homework, not only after tests. Small daily errors provide frequent low-cost opportunities to practise correction before high-pressure situations arrive.

17. A correction is not complete until it survives a fresh retest

Immediately after an explanation, the correct method is highly available. The learner has just seen the critical line, heard the reason and perhaps copied the repair. Success on the identical question can therefore overstate independence.

A fresh retest changes the numbers, wording or representation while preserving the repaired structure. It asks whether the learner can regenerate the decision rather than remember the corrected page.

Suppose Mira incorrectly expands −2(5 − 3x). After repair, do not retest only by asking her to rewrite the same expression. Give 7 − 4(2 − y). The key feature remains a negative factor multiplying a negative variable term.

Or after repairing reverse percentage with a 25% discount, use a 15% increase in a new context. The base relationship remains, but the retained/growth multiplier changes.

Fresh retests should be close enough to isolate the repaired decision. If the new problem also adds three untaught ideas, failure becomes ambiguous. First test the repair cleanly; later integrate it into mixed work.

Record support. If the learner asks “Is this another percentage base question?” and receives confirmation, selection was supported. A successful calculation still matters, but the retest has not yet shown independent identification.

Use a second retest after delay. The first can happen minutes later. The next might be days later inside a mixed set. There is no universal interval; the principle is decreasing immediacy.

Ben uses a three-level sequence: same structure today, changed context later this week, unlabeled mixed problem next week. Secure performance across these levels moves the error out of the active log.

Retests also reveal overcorrection. After learning that division by a variable can lose zero, a student may become afraid to divide by any variable. Give a problem where the variable is explicitly nonzero and division is efficient. The repair should become precise, not restrictive.

A fresh retest protects against answer memory, teacher cues and visual imitation. It is the bridge between “I understand the correction” and “I can avoid the mistake when the original page is gone.”

Do not require endless retesting after every minor slip. Use it for recurring, conceptual or high-consequence errors. The purpose is efficient evidence, not ritual repetition.

18. Look for recurring patterns only after collecting enough examples

One mistake is an event. Several similar events may form a pattern. Error analysis should avoid both extremes: ignoring obvious recurrence and generalising too quickly from one bad line.

Suppose Clara loses a negative sign once in a long paper. She immediately identifies it and solves the next five sign-sensitive questions accurately. A brief note may be enough. Calling this a persistent sign problem would be premature.

Now suppose she makes sign errors in expansion, coordinate gradient and trigonometric rearrangement across three weeks. The contexts differ, but all involve handling negative quantities during multi-step work. A broader control may be justified.

Patterns can cross topics because the underlying component is shared. Fraction weakness can appear in algebra, probability and ratio. Equation rearrangement can fail inside geometry and functions. Units can fail across speed, area and scale.

Use the error log to count meaningful recurrence, not every red mark. Record first wrong decisions. If several entries reduce to the same relationship, group them.

For example: “Percentage base” may include discounts, tax increases and reverse percentages. The surface contexts differ; the repair can be organised around identifying the reference quantity.

Another pattern might be “question target not completed”: radius versus diameter, decimal versus exact form, maximum real value versus maximum whole count. The common control is target echo at the end.

Do not combine errors merely because they share a topic name. A graph scale-reading error and a wrong gradient formula are different controls even though both occur in graphs.

Patterns should change revision allocation. If representation errors dominate, spend less time on raw calculation and more on modelling. If execution errors dominate after correct method selection, targeted fluency and checking may deserve more attention.

Mira reviews her active log weekly and asks, “Is this one problem or three versions of the same problem?” The answer determines whether she needs one general repair or several local ones.

Pattern recognition also helps measure progress. If sign errors disappear from four different contexts, the repair is more convincing than success on a single repeated worksheet.

Error analysis becomes strategic when it can distinguish noise, local weakness and cross-topic pattern. That distinction prevents both underreaction and overreaction.

19. Errors that appear only under load need practice under gradually increasing load

A learner can be accurate in short exercises and unreliable in long mixed questions. The mathematics may be known, but performance changes when several demands compete: reading, selection, algebra, units, time and checking.

Suppose Ben solves negative-number algebra accurately in isolation. In a four-stage word problem, he loses a sign after ten lines of working. Another introductory worksheet may not reproduce the condition that triggers the error.

Build load in layers. First solve the component accurately. Then place it inside a two-step problem. Then mix topics. Then add modest timing. Finally use longer exam-like work where appropriate.

This progression helps locate the threshold. If the sign error appears only after timing is added, the repair may involve pace and high-risk-line checking. If it appears as soon as another topic is mixed in, selection or working-memory load may matter.

Do not jump from comfortable practice directly to a full mock and call the resulting errors “careless”. Full papers combine too many variables to diagnose cleanly.

Mira uses one high-risk control under load: she boxes negative coefficients before distributing. She does not box every number. The control targets the pattern that recent evidence identified.

Long problems also benefit from intermediate checks. After finding a length, cost or probability, ask whether its sign, magnitude and units make sense before passing it into the next stage.

Under time pressure, selective checking matters. Recomputing every line may be impossible. Identify the operations most likely to produce expensive errors and check those.

The dedicated examination route Mathematics Examination Craft owns full performance strategy. This section keeps the error-analysis question: does the mistake appear because the underlying mathematics is weak, or because load changes execution?

If an error disappears when the problem is shortened, do not conclude the learner “knows it perfectly”. Instead, describe the condition honestly: secure in short isolated work, unstable under multi-step load. That description guides the next practice level.

Self-correction under load ultimately means detecting risk without pausing the entire paper. The learner learns which lines deserve extra attention and which can be executed fluently.

20. Calculator and digital-tool errors still belong to mathematical self-correction

A calculator can produce a wrong answer because the input represents the wrong mathematics. A graphing tool can hide roots outside the window. A spreadsheet can copy the wrong cell reference. A general AI tool can supply an invalid explanation. Tool use does not remove the need for mathematical checking.

Start with expression fidelity. If the intended calculation is 5/(2 + 3), entering 5/2 + 3 performs a different operation. Write the expression first when grouping matters.

Estimate. If 18% of 250 is required, expect about 45. A display of 450 should trigger inspection before it is copied into later work.

For trigonometry, check angle mode. Thirty degrees and thirty radians are not interchangeable inputs. The calculator cannot infer the course convention from a bare number.

For graphing, inspect domain and window. A graph that appears to have one root may have another outside the current view. Algebraic factorisation or a wider window can challenge the display.

For spreadsheets, test a copied formula on a row whose result you can calculate manually. Relative and absolute references can change silently across cells.

For automated solvers, verify candidates in the original problem. A system may return one branch, omit a condition or use a method outside the course. Tool output is a proposed result, not a substitute for mathematical validation.

Ben keeps a “tool or math?” question in mind. If the written expression is correct but the display is wrong, inspect input or mode. If the input matches the expression and the result is still surprising, inspect the mathematical expectation too.

Do not blame the calculator for executing what was entered. Likewise, do not blame the learner for every machine-induced slip without examining interface and workflow. Error analysis separates translation, operation and interpretation.

Privacy and assessment rules also matter when digital tools are used. Do not send unnecessary personal information, and do not use assistance where an assessment forbids it. These are not mathematical error categories, but they are part of responsible tool use.

A self-correcting learner treats technology as an accelerator with audit points. The more work is delegated to a tool, the more important it becomes to retain an independent sense of scale, conditions and answer meaning.

21. Feedback should reveal the missing decision without doing every later decision for the learner

Feedback can correct a page and still leave the learner dependent. If a tutor points to every wrong sign, names every method and finishes every calculation, the final work may be correct without showing whether self-correction improved.

Use the smallest feedback that makes productive work possible. If the learner can locate the wrong line after being told “one equality is false”, do not immediately identify the line. If the learner knows the line but not why, ask for a numerical check or the relevant definition.

Suppose Clara writes 5/6 ÷ 2/3 = 10/18. A full correction would show multiplication by the reciprocal and simplify the answer. A smaller cue is “What number multiplied by 2/3 gives 5/6?” This returns the inverse relationship to the learner.

If she still cannot proceed, show the equation q(2/3) = 5/6. If needed, complete the example. The support level should match the actual gap.

Record support when evaluating progress. “Correct after first-line cue” is different from “independent”. Both can be improvements, but they describe different current states.

Feedback should also distinguish what was correct. “Your model and method were right; the arithmetic changed 36 to 34” preserves the successful components. This helps the learner focus repair narrowly.

Avoid feedback that replaces mathematics with approval alone. “Be more careful” or “check your work” names no control. Better feedback is “substitute your candidate into the original equation” or “state which quantity is the percentage base”.

Use delayed feedback strategically when the learner can safely explore. Immediate correction is important when a misconception would otherwise be rehearsed repeatedly. But a short opportunity to detect the error independently can strengthen self-monitoring.

Mira’s tutor often asks, “What evidence would tell you this line is wrong?” The question turns feedback into a search for a check rather than a request for the correct answer.

The long-term goal is feedback compression. Full explanation becomes cue, cue becomes question, question becomes learner-generated check. The mathematics has not changed; control has moved.

Useful feedback leaves the learner with a repair tool that can be used when the tutor is absent.

22. Tutors and parents should separate diagnosis from rescue

When a learner is stuck, adults often want to help quickly. Help is valuable, but immediate rescue can erase the evidence needed to diagnose where independence stops.

Allow an initial attempt. Ask the learner to mark the last line they believe is definitely correct. This creates a starting point for diagnosis.

If the learner says “I don’t know anything”, ask narrower questions: What is the target? Which quantities are given? Is there a diagram or equation you can form? These questions reveal whether the difficulty is representation, selection or execution.

Parents do not need to know the correct mathematical method to support error analysis. They can ask the child to preserve the page, mark the uncertain line and bring the exact question to the tutor or teacher.

Do not invent a shortcut simply to avoid leaving a question unresolved. A clear unresolved question is better than a confident wrong rule that must later be unlearned.

Tutors should resist rewriting the student’s working into the tutor’s preferred style before locating the actual error. Different valid methods can look different. Diagnose validity before standardising presentation.

When the error is local, repair locally. If Ben made one arithmetic slip inside a correct simultaneous-equation solution, a forty-minute reteach of simultaneous equations may be unnecessary.

When the error reveals a prerequisite gap, route backward. A probability fraction problem may expose fraction multiplication weakness. A geometry equation may expose algebra rearrangement weakness. The repair belongs at the earliest unstable component that is actually causing failure.

Ask the learner to explain the repair after support. Then give a fresh problem. Adult help should produce a handover, not a permanent dependency on hints.

Parents can track evidence rather than grades alone. “The sign error appeared twice this week and was self-corrected once” is a useful conversation with a tutor. It describes a changing capability.

Adult support is most successful when the learner increasingly arrives with a diagnosis: “I think the first wrong line is here, but I cannot explain why.” That is much closer to independent correction than arriving only with a final red cross.

23. A mathematics post-mortem should separate knowledge, route and execution failures

After a test or mock, it is tempting to sort questions only by marks lost. Error analysis needs another layer: why were the marks lost?

Begin with untouched original work. For each significant loss, find the first wrong or missing decision. Do not overwrite the evidence with a perfect solution before diagnosis.

Use categories such as knowledge unavailable, model wrong, method wrong, execution wrong, condition lost, answer incomplete and time-limited. These categories can overlap; the first cause matters most.

Suppose a six-mark question loses four marks. The learner chose the correct formula, substituted correctly, but rounded an intermediate value too early and then omitted the requested unit. That is not a six-mark knowledge gap. It is precision and completion control.

Another learner leaves the same question blank because the formula is not recognised. The score loss is identical; the repair is not.

Record time information when relevant. A question left blank with ten minutes remaining differs from one left after time expired. The first may involve selection; the second may involve paper management.

For each important error, write the smallest next action. “Relearn all geometry” is rarely useful. “Repair perpendicular height versus sloping side, then retest on two changed triangles” is.

Use fresh questions for post-mortem retests. Repeating the identical paper can be valuable for understanding, but it should not be the only evidence because answer memory is now present.

The existing How to Do a Mathematics Examination Post-Mortem page owns the dedicated exam workflow. This section connects it to the global self-correction architecture: every post-mortem should feed a repair, a retest and a later return.

Do not turn the post-mortem into a punishment session. Its job is information extraction. A disappointing paper can still be valuable if it reveals a small number of correctable decision patterns.

Likewise, a strong paper deserves analysis. Which previously weak errors disappeared? Which checks worked? Success provides evidence for retiring controls and reallocating revision time.

24. Reliable self-correction builds confidence because the learner knows what to do after a mistake

Confidence in mathematics is often discussed as though the ideal learner never makes mistakes. A more useful form of confidence includes recovery: when an answer looks wrong, the learner has a method for investigating it.

Mira no longer treats a failed check as proof that the entire problem is lost. She asks where the first divergence occurred. That changes the meaning of a mistake from endpoint to information.

Evidence-based confidence uses bounded receipts. “I found my own sign error and repaired it on a new problem” is stronger than “I feel better at algebra”.

Correction receipts can be tracked just like successful answers. The learner may note: detected impossible negative length; found denominator restriction; caught calculator bracket error; rejected extraneous root.

This does not mean mistakes are desirable for their own sake. The goal remains accurate mathematics. The point is that an error need not be terminal if the learner has detection and repair tools.

Self-correction also reduces dependence on external reassurance. A student who can substitute a solution, estimate magnitude and inspect conditions can gather evidence rather than waiting for someone to say “yes, correct”.

Confidence should remain calibrated. One successful repair does not prove a topic is mastered. A learner can say, “I can now correct this error with a check, but I still need to see whether it survives mixed work.”

Parents can reinforce this calibration. Praise the specific process: locating the first wrong line, using a check, explaining the corrected condition. Avoid guaranteeing future marks from one good session.

Ben’s error log gradually changes. Early entries are written by the tutor. Later entries contain Ben’s own diagnosis and retest. The shift itself is evidence of growing control.

Correction competence matters beyond examinations. In later mathematics, science, programming and data work, debugging often means locating the first state that diverges from the intended relationship. The habit is portable.

A learner who knows how to recover can approach harder work with a more realistic kind of confidence: not “I will never be wrong,” but “If something breaks, I know how to investigate it.”

25. Test whether the learner can correct without being told that an error exists

There are levels of self-correction. At the earliest level, the teacher says the answer is wrong and shows the correction. Later, the teacher says which line is wrong. Later still, the learner is told only that something may need checking. The strongest level is spontaneous detection.

A learner who corrects after a red cross has demonstrated repair but not detection. A learner who notices that a probability exceeds one and investigates without prompting has detected and repaired.

Build independence gradually. First, mark a wrong line. Then mark only the question. Then provide a mixed page with some correct and some incorrect solutions and ask which need repair. Finally, require ordinary work to include self-selected checks.

Use hidden-error tasks carefully. The purpose is not to trick the learner. Include mathematically plausible errors and enough correct work that “everything is wrong” is not a winning strategy.

For example, present three solutions: one correct linear equation, one with a distribution error, and one with a valid alternative method. Ask the learner to classify them. This tests error detection and respect for multiple valid routes.

Spontaneous detection often begins with reasonableness. A negative area, probability 1.4 or average outside the data range should trigger inspection. Later, more subtle checks such as domain restrictions and equivalence become automatic.

Mira’s tutor stops writing “check” beside every problem. Instead, Mira chooses two questions in a set where checking is high value and states which check she will use. The control is becoming learner-owned.

Independence also includes knowing when correction requires help. If a learner detects that a theorem has been misapplied but does not know the correct theorem, asking for instruction is appropriate. Self-correction is not solitary guessing.

A useful independence test asks: Can the learner detect, locate, explain, repair and retest the error without the original correct solution visible? If one stage still requires help, record that stage rather than declaring the entire process independent or dependent.

The goal of teaching error analysis is ultimately to make explicit routines less visible. The learner should not need to recite the protocol forever. The sequence becomes part of mathematical working.

26. Build a weekly repair cycle so corrected mistakes do not disappear from memory

Error analysis becomes durable when it is connected to future practice. A correction made on Tuesday should not vanish into an old workbook. The weekly system needs a path from diagnosis to delayed independent use.

Use four jobs: collect, repair, retest and mix. Collect only recurring or consequential errors. Repair the first wrong relationship. Retest on a changed problem. Mix the repaired idea into ordinary work later.

Imagine Mira’s active errors this week are negative distribution, reverse percentage and forgetting units on area. Monday’s work repairs negative distribution and immediately retests it. Wednesday’s mixed algebra includes one sign-sensitive expression. Friday’s percentage set contains one reverse problem without a label. Sunday’s short mixed set includes an area problem whose numerical calculation is easy but whose unit matters.

The same error should not dominate every session after one successful repair. Space the return enough that memory must be reactivated. If the error returns, move it closer again. If it remains secure, reduce its frequency.

A weekly review asks which entries can retire. An error that survives a fresh direct problem, a delayed mixed problem and relevant exam-like work may move to maintenance. The exact number of tests is not universal; the principle is evidence across changing conditions.

Keep the active repair queue small. If twenty errors are treated as equally urgent, none receives a clear cycle. Prioritise by recurrence, consequence and dependency.

Foundational errors deserve attention because they spread. A fraction weakness can contaminate algebra, probability and ratio. A missing unit may cost less conceptual damage but recur across many measurement questions. The plan should reflect both reach and frequency.

Ben’s week contains one “error-free” goal only in the sense of targeted controls: he chooses two high-risk operations to monitor. He does not demand perfect performance across every topic. The system seeks reducing recurrence, not impossible certainty.

Link error repair to revision rather than maintaining two separate programmes. The How to Revise for Maths guide provides the broader scheduling architecture. This page supplies the error-specific loop that feeds it.

At the end of the week, the learner should be able to say not only “I made fewer mistakes,” but “These two error patterns no longer appeared, this one still needs a cue, and this one has moved into mixed maintenance.”

That is a more useful picture than a raw count of red marks because it describes changing control.

27. Prevention works best when each check is tied to a known risk

Error analysis is retrospective, but its purpose is partly preventive. Once an error pattern is known, build a lightweight control that can operate before the same mistake reaches the answer.

For negative distribution, the control might be “mark the outside sign before expanding”. For reverse percentage, “name the base and retained multiplier”. For rational expressions, “write denominator restrictions before cancelling”. For geometry, “mark given facts before inferring from the diagram”.

A good preventive control is short. If checking takes longer than solving every time, the learner may abandon it under load. Focus on the known risk rather than checking every possible failure.

Suppose Clara repeatedly loses the second root after equations of the form u² = a. Her control is one question: “Did a many-to-one operation create branches?” This catches square and absolute-value structures without requiring a long checklist.

For calculator work, the control may be an estimate before entering the expression. For 19% of 310, an estimate near sixty makes 589 obviously suspect. Estimation is cheaper than repeating every keystroke after the fact.

For unit conversions, write the unit at every multiplication or division until the relationship is stable. Cancelling units can expose inversions such as multiplying by sixty when division is required.

Preventive controls should be retired or compressed when they are no longer needed. A mature learner does not need to circle every negative sign forever. The control has succeeded when accurate execution becomes stable enough that attention can move elsewhere.

Do not add controls based on hypothetical mistakes the learner has never made. Personalisation here means responding to evidence. One learner may need a percentage-base control; another may need a graph-scale control.

Mira’s active controls fit on one small card. Before mixed work she reads them once. During the work she uses them selectively. After several weeks, one control disappears because the error no longer recurs.

Preventive checking is therefore the forward-facing side of error analysis. The past error becomes a future decision rule.

The strongest control eventually becomes mathematical understanding itself. A learner who deeply understands equivalent fractions does not need a reminder not to add denominators. The relationship constrains the operation automatically.

28. Error-analysis checkpoint: twenty wrong or incomplete solutions to diagnose and repair

The following twenty original tasks test diagnosis rather than only answer production. For each, identify the first wrong or incomplete decision, explain the reason, repair it, and state a useful retest or check. These tasks are not a standardised assessment and have no validated cut score.

Task 1 — Distribution

A learner writes 3(x − 4) = 3x − 4.

Diagnosis and repair: The first wrong line is the distribution. The factor three must multiply both terms, giving 3x − 12. A quick counterexample is x = 0: the original equals −12 while 3x − 4 equals −4. Retest with −2(y + 5), where both sign and full distribution must be controlled.

Task 2 — Solving a square equation

A learner solves x² = 36 and writes x = 6.

Diagnosis and repair: The answer is incomplete. Over the reals, both 6 and −6 square to 36, so x = ±6. The missing decision is completeness when reversing a square. Retest with (x − 2)² = 25 and require both branches.

Task 3 — Fraction addition

A learner writes 2/3 + 1/4 = 3/7.

Diagnosis and repair: The operation model is wrong. The fractions must be expressed in a common unit: 8/12 + 3/12 = 11/12. The result 3/7 is also suspicious because adding a positive number to 2/3 should produce more than 2/3, while 3/7 is less. Retest with 3/5 + 1/2.

Task 4 — Reverse percentage

An item costs 80 after a 20% discount. A learner calculates original price as 80 + 20% of 80 = 96.

Diagnosis and repair: The percentage base is wrong. Eighty is 80% of the original, so 0.8p = 80 and p = 100. Adding 20% of the final amount reverses a different transformation. Retest with a 25% discount producing a final price of 90.

Task 5 — Direct proportion

A learner says y = 5x + 2 is direct proportion because its graph is straight.

Diagnosis and repair: The classification is wrong. Direct proportion has y = kx under the standard model and zero intercept. Here y/x is not constant and the line does not pass through the origin. Retest by comparing y = 7x and y = 7x − 3.

Task 6 — Cancellation

A learner simplifies (x + 6)/x to 6 for nonzero x.

Diagnosis and repair: The x in the numerator is part of a sum, not a common factor of the entire numerator. The expression is 1 + 6/x. A numerical check at x = 2 gives 4, not 6. Retest with 3x/x, where cancellation is valid for x ≠ 0.

Task 7 — Perimeter versus area

A 5-by-8 rectangle is said to have perimeter 40 units because 5 × 8 = 40.

Diagnosis and repair: The learner selected the area operation for a boundary question. Perimeter is 2(5 + 8) = 26 units. The product 40 is the area in square units. Retest with a 7-by-9 rectangle asking for both quantities.

Task 8 — Probability without replacement

A bag contains four red and three blue counters. A learner gives probability of two reds without replacement as (4/7)(4/7).

Diagnosis and repair: After one red is removed, three red remain among six counters. The correct probability is (4/7)(3/6) = 2/7. The error is a lost condition: without replacement changes the second state. Retest with the same bag but with replacement.

Task 9 — Inequality direction

A learner solves −3x > 12 as x > −4.

Diagnosis and repair: Dividing by a negative reverses order, so x < −4. Check x = −5: the original gives 15 > 12, true. Check x = 0: false. Retest with −2y ≤ 10.

Task 10 — Combined mean

One group of 10 has mean 12; another group of 20 has mean 18. A learner says combined mean = 15.

Diagnosis and repair: The two means were averaged with equal weights despite unequal group sizes. Total is 120 + 360 = 480 across thirty observations, so combined mean is 16. Retest with equal-sized groups, where averaging the two means would be valid.

Task 11 — Graph scale

A graph’s vertical axis increases by 5 units every square. A learner reads a point three squares above zero as y = 3.

Diagnosis and repair: The representation was read using square count rather than scale value. Three squares represent 15. Retest with an axis where each two squares represent ten units and ask for several coordinates.

Task 12 — Radical equation

A learner solves √(x + 2) = x by squaring to x + 2 = x², gets x = 2 or −1, and reports both.

Diagnosis and repair: Candidates must return to the original equation. x = 2 works because √4 = 2. x = −1 fails because √1 = 1, not −1. Squaring introduced an extraneous candidate. Retest with √(x + 6) = x.

Task 13 — Fixed fee model

A service charges 10 plus 4 per hour. A learner writes C = 4(10 + h).

Diagnosis and repair: The fixed fee has been multiplied by the hourly rate. Correct model is C = 10 + 4h. Check h = 0: the service should still cost ten, not forty. Retest with a subscription having a fixed monthly fee plus per-use cost.

Task 14 — Similarity and area scale

Two similar figures have length ratio 2:3. A learner says their area ratio is 2:3.

Diagnosis and repair: Area scales with the square of the linear factor, so area ratio is 4:9. The error is applying a one-dimensional scale to a two-dimensional quantity. Retest with length ratio 4:5 and ask for area and volume scale analogies where appropriate.

Task 15 — Parameter division

A learner solves ax = 7 as x = 7/a for every real a.

Diagnosis and repair: Division by a requires a ≠ 0. If a = 0, the equation becomes 0 = 7 and has no solution. The general answer must split the parameter case. Retest with ax = 0, where the a = 0 case instead gives every real x.

Task 16 — Calculator grouping

A learner intends (12 + 8)/5 but enters 12 + 8 ÷ 5 and copies 13.6.

Diagnosis and repair: The machine executed a different expression. The intended value is 20/5 = 4. Write grouping before entry and estimate. Retest with (18 − 6)/3 versus 18 − 6/3.

Task 17 — Units

A learner computes a rectangular area as 42 and writes “42 cm”.

Diagnosis and repair: Area requires square units, so 42 cm². The numerical multiplication may be correct; the completion is not. Retest with volume and rate units to distinguish one-, two- and three-dimensional measures.

Task 18 — Incomplete simultaneous-equation check

A learner solves a system and checks the pair in only one of the two original equations.

Diagnosis and repair: A simultaneous solution must satisfy both conditions. Checking one equation can confirm only that one. Substitute the pair into both originals. Retest with a system where a pair satisfies one line but not the other.

Task 19 — Mean from insufficient data

A set has mean 10. A learner concludes its median must be 10.

Diagnosis and repair: The mean does not uniquely determine the median. Counterexamples: 10,10,10 has both equal to ten; 0,10,20 also has median ten; but 0,1,29 has mean ten and median one. The error is inferring more information than the summary contains. Retest by asking which quantities are determined by total and count alone.

Task 20 — Final target

A problem asks for the minimum number of vans holding eight people each for 41 people. A learner computes 41/8 = 5.125 and answers 5.

Diagnosis and repair: Five vans hold only forty people. The minimum sufficient whole count is six. The algebraic quotient is an intermediate boundary, not the completed practical answer. Retest with a maximum-budget count to contrast rounding direction.

After the checkpoint, group errors by the first wrong decision. If several tasks fail for the same reason, choose one repair and several changed retests rather than memorising twenty corrections. The checkpoint becomes useful when it changes future work.

29. Frequently asked questions about mathematics mistakes and self-correction

Why do I repeat a mistake even after I understand the correction?

Understanding immediately after feedback does not guarantee later retrieval or selection. Use a fresh retest, then return after delay in mixed work. The correction needs to become available when the original explanation is gone.

Should I write down every mistake?

No. Record recurring, important or revealing errors. One isolated arithmetic slip that you immediately detect may need only a brief note. Keep the active log small enough to guide action.

What is the difference between a careless mistake and a knowledge gap?

“Careless” is too broad. Look at the work. If the method and relationship are understood and a digit is copied wrongly once, it is likely an execution event. If the same wrong rule appears repeatedly, there is a conceptual or procedural gap to repair.

Should I redo the same question?

Yes for understanding, but not as the only evidence. Repair the original, then use a changed problem. The identical question now carries answer memory and teacher cues.

How can I stop sign mistakes?

Find which sign operation fails: distribution, subtraction, coordinate differences, inequality division or something else. Use a local control and targeted practice. “Sign errors” may contain several different mechanisms.

Why do my mistakes increase in examinations?

Time, topic switching, long working and sustained attention add load. Confirm the mathematics untimed, then build load gradually. Use selective checks at known high-risk lines. Full exam strategy belongs to the relevant examination owner.

Is checking every answer realistic?

Not with every possible method. Choose high-value checks. Substitute equation solutions, inspect units, verify domain-sensitive candidates and estimate magnitudes. Under time pressure, prioritise recent or costly error patterns.

What if I cannot find my own mistake?

Replay the solution claim by claim. Use numerical witnesses, definitions and original conditions. If the method or theorem is unknown, seek instruction. Self-correction includes knowing when outside explanation is necessary.

Can a correct answer still contain an error?

Yes. Two mistakes can cancel, or an invalid method can happen to produce the right number in one case. Judge the route as well as the endpoint, especially during learning.

Can a wrong answer come from mostly correct mathematics?

Yes. One arithmetic slip, omitted unit or rejected-condition failure can spoil an otherwise valid route. Preserve what was correct and repair locally.

When can I remove an error from my active log?

When the correction survives fresh and delayed work under relevant conditions with little or no support. It can return later if evidence changes.

Should a tutor correct mistakes immediately?

Immediate correction is useful when a misconception would be rehearsed. When safe, a short chance for independent detection can strengthen self-correction. The smallest effective feedback is often preferable to a full solution.

30. The complete error-analysis system: turn every important mistake into a better future decision

Return to Mira’s first page. The final answer was wrong, but the useful discovery was earlier: −3 times −x had been treated as negative instead of positive. Once that first wrong line was visible, the repair became small enough to teach and precise enough to retest.

The complete system is:

DETECT → LOCATE THE FIRST WRONG OR MISSING DECISION → CLASSIFY THE FAILURE → REPAIR THE RELATIONSHIP → ADD A LIGHTWEIGHT CONTROL → RETEST ON A CHANGED PROBLEM → RETURN AFTER DELAY → MIX INTO ORDINARY WORK → RETIRE WHEN SECURE.

Representation errors require better models. Selection errors require better discrimination. Execution errors require local accuracy controls. Condition errors require scope and boundary checks. Completion errors require target discipline. One generic instruction—“be careful”—cannot replace these different repairs.

Self-correction also changes the learner’s relationship with feedback. The tutor’s answer is no longer the only source of truth. Substitution, estimation, counterexamples, units, definitions, domains and independent methods provide evidence the learner can use.

As control grows, the tutor does less diagnosis for the learner. The learner begins to arrive with a specific statement: “My first invalid line is here because I cancelled across a sum,” or “The equation is right; I copied 0.06 as 0.6.” Those sentences are signs of a functioning mathematical debugging system.

Use the Secondary Mathematics Learning Hub for topic routes and repair libraries. Use How to Revise for Maths to schedule repaired ideas over time, and How to Solve Math Problems for route-building in unfamiliar tasks.

For Singapore SEC-specific cascading-error behaviour, use How SEC Mathematics Error Propagation Works. For test-specific analysis, use How to Do a Mathematics Examination Post-Mortem. These remain separate owners.

Scope note: the teaching scenes, error families, protocols and checkpoint tasks in this article are original explanatory material. They are not a validated psychological diagnostic instrument and do not guarantee examination outcomes. They are a practical mathematics-learning framework to be adapted to the learner’s taught content and course conditions.

The most useful question after a wrong answer is not “How could I get this wrong?” It is “What was the first decision that stopped being mathematically justified, and what will I do differently when that decision appears again?”

Appendix A — Ten composite mistakes where the first error is not the loudest error

The following cases extend the checkpoint. Each contains more than one tempting interpretation or several downstream wrong lines. The goal is to isolate the earliest cause and decide which later errors would disappear if that cause were repaired.

Case 1 — The right algebra solves the wrong pricing model

A service charges a fixed 12 units and then 5 units per hour. A learner models a six-hour bill as 5(12 + 6), obtains 90, and then checks the arithmetic twice. The arithmetic is internally correct. The first error is representation: the fixed fee has been placed inside the hourly multiplication. The correct model is 12 + 5(6) = 42.

The useful repair is to separate contribution types. Fixed contribution: 12. Variable contribution: 5 per hour × 6 hours = 30. A boundary check at zero hours is decisive: the service should cost 12, while the wrong model gives 60. A future retest should use a subscription or rental with a different fixed fee and unit rate.

Case 2 — A correct factorisation loses the original domain

A learner simplifies (x² − 25)/(x − 5) to x + 5 and later evaluates it at x = 5 to obtain 10. Factorisation and cancellation were executed correctly on the allowed domain. The first error occurs when the original restriction x ≠ 5 disappears.

The repair is to write x + 5, x ≠ 5. The simplified expression agrees with the original everywhere the original is defined, but it is not the same function on an enlarged domain unless an extension is deliberately defined. A retest can use (x² − 9)/(x − 3).

Case 3 — A graphing window hides a second root

A graphing tool shows one visible zero of a polynomial and the learner reports one solution. Algebraic factorisation later reveals two. The first error is not necessarily the graph or the factorisation. It is treating the current viewing window as proof of completeness.

The repair is to use the graph as local visual evidence and an algebraic or wider-domain check for completeness. If the polynomial factors as (x − 2)(x + 11), a window from −5 to 5 hides the second zero. A future control is: when the question asks for all roots, use a method that justifies completeness.

Case 4 — Early rounding corrupts a later threshold

A learner calculates a length as √52 ≈ 7.2 and then uses 7.2 in a later inequality deciding whether a design fits under a strict 7.21-unit limit. The rounded intermediate suggests the design fits. The exact value √52 ≈ 7.211102… does not.

The first error is premature approximation relative to a later threshold. The repair is to retain exact or sufficient precision through intermediate work and round only at the requested stage. A retest should involve a different boundary where early rounding could change a yes/no decision.

Case 5 — A correct percentage calculation uses the wrong base

A quantity rises from 80 to 100. A learner calculates 20/100 = 20% and says the increase is 20%. The division is correct; the base is not. Percentage increase is change divided by the original amount, so 20/80 = 25%.

The first error is representation of the reference quantity. The repair is to state “change = 20; original = 80” before dividing. A reverse comparison—from 100 down to 80—would indeed be a 20% decrease because the base then is 100. The asymmetry is a useful near contrast.

Case 6 — Two wrong steps cancel and produce the right answer

A learner solves 2(x + 3) = 18 by incorrectly expanding to 2x + 3 = 18, then incorrectly subtracting 6 instead of 3, obtaining 2x = 12 and x = 6. The final answer happens to be correct because the two mistakes compensate.

This case shows why a correct endpoint does not validate the route. Substitution confirms x = 6 solves the original, but error analysis must still reject the false equality 2(x + 3) = 2x + 3. The retest should use numbers where compensation is unlikely, such as 3(x + 4) = 21.

Case 7 — A mean is correct but the conclusion about every observation is false

A data set has mean 70. A learner concludes that every observation is close to 70. The arithmetic may be perfect. The first error is an inference beyond what the mean determines. Values 40 and 100 have mean 70 and are far apart.

The repair is to separate summary from spread. The mean describes an equal-share centre, not the variation by itself. A retest should ask whether two data sets with the same mean can have different ranges or medians.

Case 8 — A valid theorem is used before its condition has been established

A learner applies Pythagoras to a triangle because the drawing looks right-angled and computes an unknown side accurately. The arithmetic and theorem statement are correct; the theorem selection is unsupported.

The first error is condition control. A right angle must be given or proved. The repair is to mark givens separately from visual appearance. A retest should rotate a genuine right triangle and include a non-right near miss drawn deceptively.

Case 9 — A systematic trial becomes inefficient because the step size is ignored

A learner guesses random pairs for a two-type total problem even though replacing one type with the other changes the total by a fixed amount. The guesses are not mathematically invalid, but the route fails to use structure and produces repeated work.

The repair is selection rather than truth correction: choose a reference case, compute the excess or deficit, and divide by the per-replacement change. This case reminds us that error analysis can include inefficient route choice when the inefficiency causes avoidable failure under time constraints.

Case 10 — A final answer satisfies the transformed equation but not the original problem

A rectangle problem produces a quadratic with roots 4 and −11. The learner reports both because both satisfy the algebraic equation. The first mathematical transformation may be valid, but the original variable represents a positive length.

The repair is to preserve the physical domain from the beginning and filter candidates at the end. “Negative roots are always wrong” would be a bad overcorrection; negative values can be valid in coordinate or temperature models. The correct rule is to return every candidate to the original domain and meaning.

Appendix B — How to measure whether self-correction is actually improving

A self-correction programme should change observable behaviour. Do not measure success only by the number of entries in an error log or the amount of time spent correcting pages.

Track four bounded indicators. First, detection latency: does the learner notice the problem before feedback, after a generic “check”, after a line-specific cue, or only after the full solution is shown? Second, localisation accuracy: can the learner identify the first wrong decision rather than merely the final wrong answer?

Third, repair quality: can the learner explain the corrected relationship and produce a valid line? Fourth, transfer durability: does the repair survive a changed problem and later mixed work?

These indicators are not standardised psychological scores. They are practical observations for instruction. A learner might improve from “needs full correction” to “finds the line after a generic cue” even before error frequency changes dramatically. That is meaningful progress in control.

Use parallel tasks when comparing over time. If Mira originally lost signs in −3(2 − x), retest later with a different negative-distribution expression rather than the identical numbers. If Ben misused reverse percentage, use a changed percentage and context.

Measure recurrence by opportunity. Two sign errors in two sign-sensitive questions are more concerning than two errors across fifty such opportunities. Raw counts without denominator information can mislead.

Also record false alarms. A learner who distrusts every correct line and repeatedly changes correct answers has not reached efficient self-correction. Good monitoring should become more selective, not more anxious.

Self-correction is improving when the learner detects important inconsistencies earlier, locates causes more accurately, needs less external diagnosis, and carries repairs into later independent work. The final aim is fewer repeated errors with less supervisory effort.

Appendix C — A compact error card for daily use

For learners who want one portable prompt, use five questions:

1. WHAT WAS I TRYING TO FIND? Restate the target and required form.

2. WHERE DID MY WORK FIRST DIVERGE? Find the earliest false, unsupported or incomplete decision.

3. WHAT KIND OF FAILURE WAS IT? Representation, selection, execution, condition or completion?

4. WHAT CONTROL WOULD HAVE CAUGHT IT? Substitute, estimate, check units, mark domain, identify base, compare methods or use another relevant test.

5. WHAT FRESH QUESTION WILL PROVE THE REPAIR? Choose a changed problem and return later.

The card is deliberately small. If the learner needs a full chapter beside every mistake, the correction system will not survive ordinary homework. The card should start the investigation; the relevant topic guide supplies depth when needed.

As self-correction becomes habitual, the card can disappear. Its success is measured by the learner no longer needing to read it.

Appendix D — Cross-topic repair matrix: symptom → first check → repair → retest

This matrix is designed for the moment when a learner knows an answer is wrong but does not yet know what kind of follow-up to choose. The same surface topic can contain very different failures, so the first check should be as close as possible to the line that diverged.

Algebra: signs, brackets and equivalence

Symptom: an expression changes sign unpredictably after expansion. First check: isolate the outside factor and multiply it by each term separately. Repair: write one visible distribution line before simplifying. Retest: use a fresh expression in which a negative outside factor meets both a positive and a negative term.

Symptom: an equation has a neat answer but substitution fails. First check: compare each equality from top to bottom and ask which operation was applied to both sides. Repair: restore the first false equality, preserving any nonzero restrictions. Retest: solve a changed equation and substitute into the original, not only the last transformed line.

Symptom: cancellation seems to remove terms inside addition. First check: ask whether the cancelled object is a common factor of the entire numerator or denominator. Repair: factor before cancelling and state domain restrictions. Retest: pair one valid factor cancellation with one invalid attempted cancellation across a sum.

Symptom: only one solution is reported from a square or absolute-value equation. First check: identify whether a many-to-one operation has been reversed. Repair: restore all legitimate branches and check them in the original relation. Retest: use a changed square equation and an absolute-value equation so the branching cue must be recognised rather than copied.

Ratio, proportion and percentage: reference quantities

Symptom: the arithmetic is accurate but a percentage change is wrong. First check: name the reference amount in the denominator before calculating. Repair: write change/original for percentage change, or final = multiplier × original for reverse percentage. Retest: compare an increase from 80 to 100 with a decrease from 100 to 80 so the changing base becomes visible.

Symptom: a ratio problem starts by dividing by the sum of ratio parts even when the given quantity is a difference or one component. First check: state what the supplied number represents. Repair: map the number to total parts, difference parts or a named component before finding one part. Retest: use three questions with the same 3:5 ratio but different meanings for the given number.

Symptom: a linear graph is treated as direct proportion. First check: inspect the intercept and whether y/x is constant for nonzero x. Repair: distinguish y = kx from y = kx + c with c ≠ 0. Retest: compare two lines with the same gradient but different intercepts.

Geometry and measurement: object, theorem and dimension

Symptom: area and perimeter formulas are interchanged. First check: point to the object being measured—surface or boundary. Repair: trace the boundary for perimeter and partition the region for area. Retest: use the same rectangle dimensions but ask separately for area, perimeter and the effect of cutting a notch.

Symptom: a theorem is used because the diagram looks suitable. First check: mark which facts are given or already proved. Repair: name the condition that authorises the theorem, such as a right angle or parallel lines. Retest: rotate a valid figure and place it beside a visually similar invalid one.

Symptom: a scale factor is applied directly to area or volume. First check: count the number of independent length dimensions in the measured quantity. Repair: square the linear factor for area and cube it for volume in similar figures. Retest: begin with a 2:3 length scale and ask for length, area and volume ratios side by side.

Symptom: units are missing or dimensionally wrong. First check: read the final quantity aloud with its unit. Repair: attach units through intermediate work until the dimension is stable. Retest: mix a length, an area, a volume and a rate question so the unit form must be selected independently.

Statistics and probability: weighting, events and changing states

Symptom: two group means are averaged directly despite unequal group sizes. First check: recover each group’s total as mean × count. Repair: combine totals and counts before dividing. Retest: compare equal-sized groups, where the shortcut works, with unequal groups, where it does not.

Symptom: probabilities are added and the result exceeds one. First check: inspect event overlap. Repair: subtract shared outcomes or restructure the event using a complement. Retest: use one pair of disjoint events and one overlapping pair.

Symptom: the second draw uses the original denominator in a without-replacement problem. First check: redraw the state after the first event. Repair: update both favourable and total counts. Retest: solve the same colour event once with replacement and once without replacement.

Symptom: a summary statistic is used to infer information it does not contain. First check: ask whether two different data sets can share the summary but differ in the claimed property. Repair: construct a counterexample. Retest: ask whether mean determines median, range or every observation.

Functions, graphs and modelling: domain and meaning

Symptom: a simplified formula is evaluated at an input excluded from the original expression. First check: recover the original domain before simplification. Repair: carry restrictions beside the simplified form. Retest: use another removable-factor rational expression.

Symptom: a graphing window is treated as proof that no other roots or intersections exist. First check: ask what domain the window actually displays. Repair: combine graph evidence with algebraic structure or a justified wider search. Retest: choose a polynomial with one root inside and one root outside a standard window.

Symptom: a physical model produces impossible negative amounts or values beyond capacity. First check: state the intended domain and boundary of the model. Repair: stop or change the formula when the real rule changes. Retest: use a filling tank, decaying balance or fixed-capacity model with a clear boundary time.

Symptom: a whole-number context produces a fractional algebraic answer that is rounded automatically. First check: ask whether the original equality or inequality can be satisfied by an allowed integer. Repair: distinguish exact impossibility from minimum/maximum rounding constraints. Retest: contrast an exact-count equation with a capacity inequality.

How to use the matrix

Choose the row that resembles the first wrong decision, not merely the topic name. If a geometry problem fails because of algebraic rearrangement, use the algebra repair. If a percentage problem fails because of calculator grouping, use the tool-control repair. The mathematical mechanism outranks the chapter label.

After the repair, create one clean retest and one later mixed retest. If both succeed independently, reduce the control. If the same error returns, move the repair closer and inspect whether the diagnosis was too broad or whether a prerequisite remains unstable.

The matrix is not a substitute for full teaching. It is a routing device: symptom → first check → repair → evidence. Its purpose is to make the next action smaller, clearer and more likely to change future mathematics.