The AI and Computer-Algebra Solution Audit: Restore Missing Assumptions and Verifiable Steps

Mathematics technology can produce a correct-looking solution faster than a learner can inspect it. That changes the checking problem.

A calculator, computer algebra system, symbolic engine or generative AI may return an expression, decimal answer, transformed equation, proof-like explanation or full worked solution. The output may be correct. It may also omit a domain restriction, assume a branch, simplify under hidden conditions, round too early, use a theorem outside its jurisdiction or present a numerically plausible value without showing why it answers the original question.

The learner therefore needs a technology-era verification routine:

Preserve the input → identify assumptions → reconstruct the mathematical path → test the result independently → return to the original problem → perform a fresh task without relying on the tool.

1. Tool output is a candidate mathematical object

Do not begin by asking whether the solution “looks intelligent”. Ask what claim the output is making.

If the tool says x = 3, the claim is that 3 satisfies the original equation and that no required alternatives have been omitted. If it gives an antiderivative, the claim is that differentiating the result recovers the integrand on the relevant domain. If it gives a factorisation, the claim is algebraic equivalence under the stated conditions.

2. Preserve the original input

Before inspecting the answer, keep the exact problem that was submitted. A verification audit is impossible if the prompt or equation has been rewritten after the fact.

Record:

• the original mathematical statement;
• the domain or context if given;
• the units;
• the requested answer form;
• the tool output;
• any assumptions explicitly stated by the tool.

3. Input errors can produce perfectly coherent wrong solutions

Suppose the intended equation is:

2(x + 3) = 14.

If the user enters 2x + 3 = 14 instead, a solver can correctly return x = 5.5 for the entered equation. The tool is not the source of the mathematical mismatch. The input is.

Verification therefore begins before the first generated line.

4. Parentheses are mathematical structure

Technology often removes the visual spacing cues that paper gives naturally. Compare:

1/(x+2)

with:

1/x + 2.

They are different functions. A typed expression must preserve grouping explicitly.

5. Angle mode can change a trigonometric result

A calculator evaluating sin(30) may interpret 30 as degrees or radians depending on mode. The display can be numerically precise and mathematically wrong for the intended context.

The audit question is not “Did the calculator calculate?” It is “What mathematical object did it calculate?”

6. Exact versus approximate output

Suppose a tool gives:

x = 1.414213562.

This may be an approximation to sqrt(2), but the decimal does not prove the exact form. If the question requests an exact answer, preserve symbolic reasoning or recover the exact relationship.

7. Domain assumptions must be restored

Consider:

(x^2 – 4)/(x – 2).

A symbolic tool may simplify this to x + 2. Algebraically, that equality holds where the original denominator is defined. The original expression excludes x = 2.

The correct audit writes:

(x^2 – 4)/(x – 2) = x + 2 for x != 2.

Removing the restriction changes the function.

8. Square roots and principal values

A tool may state sqrt(49) = 7. That is correct: the radical symbol denotes the non-negative principal square root.

But solving x^2 = 49 requires x = 7 or x = -7.

Technology cannot rescue a learner who has not distinguished evaluation from equation solving.

9. Squaring can introduce candidates

Consider:

sqrt(x + 6) = x.

Squaring produces candidates x = 3 and x = -2. The original equation accepts only x = 3.

A generated solution that stops after factorisation is incomplete even if every algebraic line after squaring is correct.

10. Division can remove branches

For:

x(x – 4) = 0,

dividing by x assumes x != 0 and can erase the valid solution x = 0.

If a tool simplifies by cancellation or division, ask what zero cases were excluded.

11. Logarithms require positive arguments in real-number problems

For log(x – 3), the real domain is x > 3. A candidate outside that domain cannot be repaired by later arithmetic.

12. Even roots require non-negative radicands in real-number problems

If a tool manipulates sqrt(2x – 5), record 2x – 5 >= 0 before transformations. The condition belongs to the original object.

13. Rational expressions require non-zero denominators

Before clearing denominators, record every excluded value. Multiplying through can hide which values made the original expression undefined.

14. A full audit has four layers

1. Input layer: Was the problem represented correctly?
2. Transformation layer: Does each step preserve or appropriately control the solution set?
3. Output layer: Does the final value satisfy the original mathematical conditions?
4. Context layer: Does the result answer the original real-world or examination question?

15. Correct algebra can still fail the context

A quadratic model for a physical length may produce 5 and -8. The algebraic roots can both be valid roots of the equation. A length of -8 is inadmissible in the physical model.

The final answer must re-enter the context.

16. Units are part of the proof of meaning

A result of 12 for an area question is incomplete until the unit is square units. A rate with kilometres divided by minutes is not interchangeable with kilometres per hour without conversion.

Unit mismatch is often detectable before numerical checking.

17. Computer algebra systems can transform forms without explaining usefulness

A CAS may rewrite an expression into a mathematically equivalent form that is less useful for the task. Expanded form may be good for coefficient comparison; factorised form may reveal roots; completed-square form may reveal the vertex.

Equivalence does not imply equal usefulness.

18. Simplest-looking is not always best

For a quadratic, x^2 – 6x + 5, factorised form (x – 1)(x – 5) exposes roots. Completed-square form (x – 3)^2 – 4 exposes the vertex. “Simplify” has to be interpreted relative to the reader job.

19. AI explanations can insert plausible but unnecessary reasons

A generated explanation may mention a theorem that sounds relevant without being the actual justification used by the mathematics.

Audit every named theorem by asking:

• What are its hypotheses?
• Are those hypotheses satisfied?
• Does the theorem prove this exact step?

20. A fluent proof can still contain a converse error

If the task is to prove P implies Q, a generated argument that proves Q implies P has solved a different problem.

Write the logical direction explicitly before reading the proof.

21. A proof can assume the conclusion

Generated prose can conceal circularity because the wording changes while the logical content repeats. Identify what is assumed and what must be established.

22. Numerical evidence is not universal proof

An AI system may check many cases and conclude a pattern always holds. Unless the statement is finite and all cases were exhaustively covered, successful tests are evidence, not proof.

23. One counterexample can overturn a universal claim

If a generated solution asserts “for every real x…”, actively search boundaries: zero, negative values, fractions, excluded denominators, endpoints, equality cases and parameter extremes.

24. Restore quantifiers

A proof may silently move from “for some” to “for all”. Rewrite the claim with explicit quantifiers where necessary.

25. Restore domain

Write the domain at the top of the audit. Integers, positive integers, real numbers, complex numbers and non-zero reals support different statements.

26. Restore parameter conditions

If solving ax = b, a generated solution x = b/a requires a != 0. At a = 0, the equation behaves differently depending on b.

27. Parameter case split example

Equation:

ax + 2 = 5.

If a != 0, x = 3/a.

If a = 0, the equation becomes 2 = 5, so there is no solution.

A tool that gives x = 3/a without a condition has produced an incomplete general solution.

28. Restore branch logic

If an inverse operation has multiple branches, list them. Squared equations, absolute values and trigonometric equations often need case control.

29. Absolute value example

|x – 2| = 5 means:

x – 2 = 5 or x – 2 = -5.

Thus x = 7 or x = -3.

A one-branch output is incomplete.

30. Trigonometric solutions need interval control

A CAS may return a general periodic family. An examination question may ask only for solutions in a specified interval. The audit must connect the general solution to the requested domain.

31. A graphing result needs window control

If a tool reports visible roots, ask whether additional roots can lie outside the window. Graphs are evidence, not automatic global proof.

32. Numerical root finders need starting conditions

Iterative methods can converge to different roots depending on initial values. A solver returning one root does not necessarily establish completeness.

33. Residual checking for numerical roots

If x approximately equals 1.769 is proposed as a root of f(x)=0, evaluate f(1.769). A small residual supports the numerical claim. It does not by itself prove uniqueness or exactness.

34. Differentiate an antiderivative to verify it

If a tool returns F(x) as an antiderivative of f(x), compute F'(x). This is often cheaper and stronger than trusting the integration steps.

35. Differentiate with domain awareness

Equivalent-looking antiderivatives can differ by constants on connected intervals, and expressions with absolute values or logarithms may need domain-aware interpretation.

36. Integrate a derivative only with the constant

If a tool returns one antiderivative without +C in an indefinite integration context, the family is incomplete.

37. Matrix outputs require dimension checks

Before multiplying matrices, verify dimensions. A tool may reject an invalid multiplication, but the learner should understand why the product is undefined.

38. Inverse matrices require invertibility

A formula involving A^{-1} requires A to be invertible. If det(A)=0, the inverse does not exist.

39. Numerical inverse versus symbolic inverse

A floating-point inverse can be unstable for an ill-conditioned matrix. At advanced levels, a small numerical residual may need interpretation rather than blind acceptance.

40. Statistics outputs need model assumptions

A regression line can be computed even when the relationship is strongly nonlinear. A p-value can be produced even when sampling assumptions are violated. Technology makes calculation easier; it does not make assumptions disappear.

41. Regression audit

• Plot the data.
• Check whether a linear form is plausible.
• Inspect residuals.
• Check influential outliers.
• State the domain of prediction.
• Avoid causal language unless the design supports it.

42. Statistical significance is not practical importance

A tool can produce a small p-value. The learner still has to interpret effect size, context and practical relevance.

43. AI-generated data are not empirical evidence

If an AI system invents an illustrative dataset, label it as synthetic. Do not treat generated numbers as observations about real learners, markets, schools or populations.

44. Source and computation must stay separate

A tool may explain an official examination rule or syllabus fact while also doing mathematics. Current policy claims require current authoritative sources; algebraic derivations require mathematical verification. Do not let one form of confidence substitute for the other.

45. The independent-route principle

Checking by repeating the same tool call is weak if the same hidden assumption remains.

Use a different route when possible:

• substitute a root into the original equation;
• differentiate an antiderivative;
• multiply factors to recover the original polynomial;
• estimate before accepting a decimal;
• graph a symbolic result for a reasonableness check;
• prove a pattern after numerical exploration;
• check units and bounds in a model.

46. Audit Case 1: equation solving

Tool output: for sqrt(x+6)=x, solutions are x=3 and x=-2.

Audit:

1. Original square root requires x >= 0.
2. Squaring may create candidates.
3. Substitute both values into the original.
4. x=3 works; x=-2 fails.
5. Final solution set: {3}.

47. Audit Case 2: rational simplification

Tool output: (x^2-9)/(x-3) simplifies to x+3.

Audit:

1. Original expression excludes x=3.
2. Factor numerator: (x-3)(x+3).
3. Cancel for x != 3.
4. Final statement: expression equals x+3 on its original domain x != 3.

48. Audit Case 3: quadratic context

A model produces x=-4 and x=9 for a length. Algebraic audit: both roots satisfy the polynomial. Context audit: negative length is inadmissible. Final contextual answer: 9 units.

49. Audit Case 4: graph root

Tool reports one visible root near x=2. A symbolic factorisation reveals another root at x=50. Failure source: limited graph window.

50. Audit Case 5: derivative

Tool returns derivative of x^3 sin x as 3x^2 sin x + x^3 cos x.

Audit using product rule confirms the structure. A numerical spot check can support but is not needed once the symbolic rule is correctly applied.

51. Audit Case 6: integral

Tool returns integral of 2x as x^2. For an indefinite integral, restore +C:

x^2 + C.

Differentiate to verify 2x.

52. Audit Case 7: probability

Tool says probability of at least one six in two die rolls is 1/6 + 1/6 = 1/3.

Audit: the events “first die six” and “second die six” overlap when both are six. Correct complement method:

1 – (5/6)^2 = 11/36.

53. Audit Case 8: averages

Tool averages class means 80 and 60 to obtain 70. If class sizes are 10 and 30, individual-weighted mean is 65. The missing assumption was equal group weight.

54. Audit Case 9: units

Tool uses 2 hours directly with a rate in litres per minute. Unit audit forces 2 h = 120 min before substitution.

55. Audit Case 10: theorem condition

Tool uses Pythagoras on a triangle with no right-angle condition. The arithmetic may be flawless; the theorem is inapplicable.

56. Audit Case 11: logarithm

Tool solves log(x-2)=0 and returns x=3. Domain x>2 is satisfied, so the candidate survives.

57. Audit Case 12: parameter

Tool solves ax=0 by dividing by a and states x=0. Missing case: a=0, when every real x satisfies 0=0. General solutions require parameter branching.

58. AI prose audit: remove filler, keep dependency

Generated solutions can be longer than necessary. Removing prose is safe only if mathematical dependencies remain visible. Keep conditions, justifications, branch points and final interpretation.

59. AI prose audit: do not reward confident tone

Confidence is not evidence. A terse correct line can be stronger than a polished paragraph containing an invalid implication.

60. AI prose audit: separate known from inferred

Mark which statements are given, which are derived, which are assumptions and which are interpretations. This prevents generated explanations from blending layers.

61. Computer algebra audit: request transformations, not authority

Use CAS to expand, factor, solve, simplify or differentiate, but ask what conditions make the transformation valid. The tool is a mathematical instrument, not the owner of the conclusion.

62. Numerical tool audit: estimate first

Before pressing calculate, estimate sign, order of magnitude and plausible range. A result outside the estimate triggers inspection immediately.

63. Fresh unaided retest is the learning check

After auditing a technology-assisted solution, give a new problem with the same underlying issue but different surface features.

If the learner can now recognise the condition unaided, the interaction produced learning. If they can only explain the old tool output, the capability may still belong to the tool.

64. Fresh retest: domain

Solve:

(x^2-16)/(x-4) = 8.

Record x != 4 before cancellation. Simplification gives x+4=8, producing x=4, which is excluded. Therefore the original equation has no solution.

65. Fresh retest: squaring

Solve:

sqrt(x+12)=x.

State x>=0, square, solve candidates and return them to the original.

66. Fresh retest: branch loss

Solve:

x(x+5)=0.

Do not divide by x without retaining x=0.

67. Fresh retest: graph window

Graph:

(x-1)(x-30).

Explain why a standard -10 to 10 window cannot establish the complete root set.

68. Fresh retest: weighted mean

Group A: 5 observations, mean 90. Group B: 45 observations, mean 60. Compute the combined individual-weighted mean and explain why 75 is not that quantity.

69. Fresh retest: probability overlap

Two dice. Find probability at least one die is six. Use complement or inclusion-exclusion rather than adding 1/6 + 1/6 without correcting overlap.

70. Fresh retest: theorem jurisdiction

A triangle has sides 5 and 7 with included angle 40 degrees. Explain why Pythagoras is not the default route.

71. Tool-use record

For important work, record:

• tool used;
• input;
• output;
• assumptions checked;
• independent verification;
• final accepted result;
• reason for accepting or rejecting tool steps.

72. This is not surveillance

The purpose of a process record is mathematical accountability and learning, not suspicion. A learner should be able to explain the mathematics they submit.

73. Legitimate support versus substituted reasoning

A calculator may reduce arithmetic load. A screen reader may provide access. A CAS may expose structure. An AI system may suggest a method. The learning target determines whether the support is appropriate.

If the target is solving a model, technology may be integral. If the target is unaided algebraic manipulation, automatic symbolic solving may replace the assessed capability.

74. Tool removal should be task-specific

Do not assume every support must disappear. Accessibility supports may remain necessary. Computational tools may be part of the authentic mathematical environment. The question is whether the support performs the target reasoning for the learner.

75. Examination rules are separate from learning value

A tool can be pedagogically useful and prohibited in a particular assessment. Another may be allowed but strategically unhelpful. Current examination permissions should be checked against the governing authority rather than assumed from general technology practice.

76. AI can generate useful wrong solutions for training

Deliberately planted errors can train verification if they are labelled as practice material. The learner’s task is to find the first invalid step, repair it and retest on a fresh problem.

77. Do not present generated errors as real learner data

Synthetic examples should remain synthetic. They demonstrate failure modes without making claims about actual students.

78. Verification ladder

1. cheap plausibility check;
2. domain and unit check;
3. substitution or inverse operation;
4. alternate representation;
5. independent method;
6. proof or theorem-based justification when required.

Use the cheapest check strong enough for the claim.

79. Confidence should track evidence

A graph estimate earns approximate confidence. A residual check supports a numerical root. An exact factorisation can establish polynomial roots. A theorem can establish a universal claim under its hypotheses.

Do not report more certainty than the verification route provides.

80. A complete AI/CAS audit template

81. Worked full audit

Task: solve sqrt(x+1)=x-1 over the real numbers.

Domain: x+1>=0 and x-1>=0 because the square root is non-negative and equals x-1; therefore x>=1.

Square:

x+1=(x-1)^2=x^2-2x+1.

So:

x^2-3x=0.

x(x-3)=0.

Candidates: 0 and 3.

Domain already removes 0. Check 3:

sqrt(4)=2 and 3-1=2.

Final solution: x=3.

An automated solver that reports {0,3} without restoring the original domain is incomplete.

82. Worked full audit: optimisation

If a tool finds a stationary point by solving f'(x)=0, the audit still asks whether the point is a maximum, minimum or neither, whether endpoints matter, and whether the point lies in the allowed domain.

83. Worked full audit: statistics

If AI recommends the mean as “the best average”, inspect skew, outliers, reader purpose and denominator. Statistical summaries require context.

84. Worked full audit: modelling

If a fitted curve predicts negative population, volume above physical capacity or impossible time values, the model fails a contextual constraint even if software fit statistics look strong.

85. Worked full audit: proof

If an AI proof checks many numerical cases, ask where the universal step occurs. If none exists, the output is exploration rather than proof.

86. Worked full audit: geometry

If software measures an angle from a diagram, distinguish a constructed invariant from a display measurement. Dynamic geometry can suggest a theorem; proof establishes it.

87. Technology should expose structure

The best use of technology often makes a relationship visible: parameter variation, graph transformation, residual pattern, numerical instability, geometric invariance or counterexample search.

The learner should emerge able to explain what changed and why.

88. Technology should not hide the problem definition

If the learner cannot state what the symbols mean, what the domain is or what the answer represents, a polished tool output has outrun understanding.

89. Technology should create a verification opportunity

A second representation can be powerful precisely because it can disagree with the first. Use that disagreement diagnostically.

90. The fresh-task rule

Never use the same tool-produced example as the only evidence that a learner has mastered the issue. Change coefficients, context, representation or method cue.

91. Transfer example set

1. Rational simplification with a removable hole.
2. Square-root equation with an extraneous candidate.
3. Quadratic context with a negative algebraic root.
4. Graph with a root outside the initial window.
5. Indefinite integral missing the constant.
6. Probability union with overlapping events.
7. Weighted mean with unequal subgroup sizes.
8. Parameter equation with a zero-coefficient case.
9. Triangle problem where theorem conditions must be checked.
10. Regression result where extrapolation is unsupported.

92. Tutor feedback language

• The tool solved the entered problem, but the entered problem differs from the original.
• This simplification is valid only when…
• This step generates candidates rather than guaranteed solutions.
• This division removes a zero case unless it is handled separately.
• The numerical output is approximate; the question asks for exact form.
• The theorem requires a condition that has not been established.
• The graph supports a local observation, not the global conclusion.
• The model violates a physical bound.
• The average uses the wrong weighting.
• The final value does not answer the requested quantity.

93. Student self-check

• I preserved the original task.
• I know what the tool actually computed.
• I wrote the domain and units.
• I checked branches and excluded values.
• I distinguished exact from approximate output.
• I checked theorem conditions.
• I used a different verification route where practical.
• I returned the result to the original context.
• I can explain why I accepted or rejected the output.
• I can solve a fresh related problem without copying the tool solution.

94. The Technology Atlas rule

Technology is valuable when it changes a mathematical learning function—representation, exploration, calculation, feedback, verification or access—without silently taking ownership of the reasoning being assessed.

95. Continue through BTT Mathematics

Use How to Check Maths Answers for the full verification system, How to Use a Calculator in Maths for input and independent checking, and Accessibility and Inclusive Technology in Mathematics Education when the technology is an access route rather than a temporary computational scaffold.

96. The durable target

An AI or CAS answer should make mathematical judgement more important, not less.

The learner owns the conclusion when they can reconstruct the assumptions, verify the steps, reject invalid branches and carry the reasoning into a fresh problem.

Final Technology Audit Clinic: When the Tool Is Right but the Mathematics Is Still Incomplete

The final layer of technology literacy is knowing that “correct output” and “complete solution” are not synonyms.

1. Correct value, missing justification

A CAS returns x=4. If the task assesses solving and reasoning, the learner still has to show why 4 follows and whether other solutions exist.

2. Correct simplification, smaller domain

Software simplifies (x²−1)/(x−1) to x+1. The original expression excludes x=1. Preserve that exclusion.

3. Correct decimal, wrong answer form

A calculator gives 0.333333333. If the task requires exact form, 1/3 is the mathematical answer and the decimal is an approximation.

4. Correct graph, incomplete global claim

A graph may show all features in the current window and still hide behaviour outside it. Change the window or use theory before making a global claim.

5. Correct regression, unjustified extrapolation

A line can fit observed data strongly and still fail beyond the observed range. Prediction jurisdiction is part of model validity.

6. Correct symbolic integral, missing context

An antiderivative can be correct while a definite integral or physical accumulation question still needs limits, units and interpretation.

7. Correct proof fragment, missing case

An AI proof may handle positive x and forget x=0 or negative x. Compare proof scope with claim scope.

8. Correct probability arithmetic, wrong independence assumption

Multiplication of branch probabilities is justified only by the conditional structure of the model. Without-replacement sampling often changes later probabilities.

9. Correct mean, wrong weighting

A spreadsheet can average subgroup means exactly. If subgroup sizes differ and individuals are the intended unit, the weighting may be wrong.

10. Correct theorem, unmet hypothesis

The theorem name can be correct while its use is not. Check right angle, continuity, differentiability, independence, invertibility or other required conditions.

Decision record

For high-value work, keep a compact audit:

• What did the tool receive?
• What did it return?
• Which assumptions are explicit?
• Which assumptions are hidden?
• What independent check was performed?
• What part of the output was accepted?
• What part was rejected or qualified?
• Can the learner reproduce the reasoning on a fresh task?

Unaided transfer set

1. Simplify a rational expression while preserving excluded values.
2. Solve a radical equation and reject an extraneous candidate.
3. Use a graph to estimate roots, then verify completeness algebraically.
4. Check a CAS antiderivative by differentiation.
5. Audit a weighted average from unequal groups.
6. Audit a probability answer where events overlap.
7. Audit a model prediction against a physical capacity.
8. Audit a proof that appears to prove the converse.
9. Audit a parameter solution that divides by the parameter.
10. Audit a calculator result whose units do not match the requested quantity.

Technology maturity

A learner is not technology-independent because they never use tools. Mathematical maturity means choosing tools deliberately, knowing what they contribute, knowing what they cannot establish, and retaining ownership of the final claim.

Final rule

The tool may calculate, transform, search, render or suggest. The learner still owns domain, assumptions, verification, interpretation and transfer.

Release Check: A Tool-Assisted Answer Must Still Be Auditable

Before accepting a technology-assisted Mathematics solution, require a short audit trail. This does not need to be bureaucratic. It needs to make mathematical ownership visible.

1. Input: write the exact expression, data or problem sent to the tool.
2. Conditions: record domain, units, interval, parameter restrictions and relevant assumptions.
3. Output: copy the proposed result without silently editing it.
4. Risk point: identify the step most likely to change domain, branches, precision or theorem jurisdiction.
5. Independent check: use substitution, inverse operation, alternative representation, bound, derivative, factor expansion or another suitable route.
6. Interpretation: return the result to the original question.
7. Transfer: solve a fresh related task with reduced assistance.

Three final examples

CAS simplification: if (x²−25)/(x−5) becomes x+5, retain x≠5.

AI proof: if the explanation checks five examples of a universal claim, classify the output as evidence or conjecture, not proof.

Graphing calculator: if one root is visible, vary the window or use symbolic reasoning before claiming uniqueness.

What acceptance sounds like

“I accept this result because it satisfies the original equation, respects the domain and agrees with an independent check.”

“I accept the decimal as an approximation but retain the exact form for the final answer.”

“I reject this branch because it violates the original denominator restriction.”

“I reject the global conclusion because the graph only supports behaviour in the displayed window.”

What learning sounds like

The learner can explain not only what the tool returned but why the answer is valid, where it is valid and which part of the reasoning they could reproduce without the tool.

That is the release threshold for technology-assisted Mathematics: the tool can contribute computation or representation, but the mathematical claim remains inspectable and owned.

Closure: Mathematical Ownership After Automation

Automation changes where mistakes hide. When arithmetic and algebra become easier to outsource, domain, assumptions, interpretation and verification become more important.

A mature learner should be able to receive a tool-generated solution and perform four independent actions:

1. state the exact claim the output is making;
2. identify the conditions under which that claim is valid;
3. test the claim through a different route;
4. carry the underlying reasoning into a fresh problem.

This standard applies whether the technology is a basic calculator, graphing calculator, spreadsheet, CAS, numerical solver or generative AI system.

The point is not to remove technology. It is to prevent fluent output from becoming unexamined authority.

A tool-assisted solution becomes learner-owned when the learner can say: “I know why this result works, where it works, what could make it fail, and how I would check it if the tool were wrong.”

That is the Technology Atlas end state: more computational power, with no reduction in mathematical accountability.

Final Note

A technology audit should end with one question that no tool can answer on the learner’s behalf: What evidence makes this conclusion trustworthy?

The evidence may be substitution into the original equation, exact symbolic equivalence, a theorem whose hypotheses are satisfied, a residual small enough for the stated approximation, agreement with an independent representation, a physical bound, a unit check, or a fresh problem solved without copying the original output.

Different claims require different strengths of evidence. A graph can support an estimate. A calculator can support a numerical value. A CAS can support a transformation. An AI explanation can suggest a route. None of those surfaces automatically supplies the missing domain, proof of completeness, contextual interpretation or assessment legitimacy.

The final accepted solution should therefore make its verification route visible. That is how BTT keeps technology inside Mathematics rather than allowing Mathematics to disappear inside the tool.

Verification Remains the Last Mathematical Act

Whatever technology appears earlier in the workflow, the final mathematical act is still verification. The learner must connect the returned expression, number, graph, proof or model to the original conditions and decide what the evidence really establishes.

That final act is what separates automated production from mathematical capability. Tools can shorten computation and widen exploration; they do not remove the need to justify the conclusion.

AI and Computer-Algebra Audit Casebook: Restore the Missing Mathematics

The examples below are synthetic audit exercises. They do not report a comparison of named AI models, computer-algebra systems or calculator products. Their purpose is to train the learner to treat any machine-produced solution as a candidate mathematical object that must still be checked.

A tool can be useful for algebra, graphs, numerical approximations, symbolic manipulation and explanation. The learner still owns the assumptions, domains, logical direction, verification and connection to the original problem.

Audit case 1: cancellation with a hidden zero case

Candidate solution:

x(x-3)=0

“Divide by x to get x-3=0, so x=3.”

The division by x assumes x≠0 and silently removes the valid solution x=0.

Repair:

x(x-3)=0 gives x=0 or x=3.

Audit lesson: before dividing by an expression containing the unknown, record the zero case.

Audit case 2: squaring introduces an extra candidate

Solve:

sqrt(x+2)=x.

A candidate solution squares both sides:

x+2=x²

x²-x-2=0

(x-2)(x+1)=0

giving x=2 or x=-1.

But the original square root is non-negative and the right side must therefore be non-negative. Substitution gives:

• x=2: sqrt(4)=2, valid.
• x=-1: sqrt(1)=1≠-1, invalid.

Audit lesson: non-reversible transformations require a return to the original equation.

Audit case 3: logarithm domain omitted

Solve:

log(x-1)+log(x-3)=log 8.

Before combining logarithms, the real-domain conditions are:

x-1>0 and x-3>0, so x>3.

Combining gives:

(x-1)(x-3)=8

x²-4x-5=0

(x-5)(x+1)=0.

Candidates: 5 and −1. Only 5 satisfies the original logarithm domain.

Audit lesson: domain restrictions belong at the start, not as an afterthought.

Audit case 4: denominator exclusions vanish after clearing fractions

Suppose:

1/(x-2)=3/(x²-4).

The original expression excludes x=2 and x=-2.

Since x²-4=(x-2)(x+2), multiplying through can simplify the equation, but excluded values remain excluded even if they disappear algebraically later.

Audit lesson: transformations can hide where the original expression was undefined.

Audit case 5: square root branch lost

From x²=25, a candidate solution reports x=5.

The complete real solution is x=±5.

Audit lesson: inverse operations may have multiple branches.

Audit case 6: absolute value needs two cases

For |2x-1|=7, solve:

2x-1=7 or 2x-1=-7.

This gives x=4 or x=-3.

A one-branch answer is incomplete even if the chosen branch is correct.

Audit case 7: parameter solution without the singular case

Solve ax+2=5 for x.

A candidate output gives x=3/a.

This is valid only when a≠0. If a=0, the equation becomes 2=5, so there is no solution.

Audit lesson: a symbolic formula may conceal parameter cases.

Audit case 8: candidate root without completeness

A numerical solver returns one approximate root of a nonlinear equation.

Verification questions:

• Does substitution give a small residual?
• Could other roots exist?
• Was a search interval specified?
• Does the function change sign elsewhere?
• Does monotonicity or another theorem establish uniqueness?

A verified candidate root is not automatically a complete solution set.

Audit case 9: graph window hides another root

A graphing tool displays one crossing in the default window.

Audit response:

1. record the displayed window;
2. inspect algebraic or asymptotic structure;
3. expand the window if appropriate;
4. use symbolic reasoning to establish completeness where required.

The graph is evidence about the visible region, not proof about all real numbers.

Audit case 10: graph suggests equality but scale is deceptive

Two curves appear to overlap. If the y-axis spans millions of units, a difference of 100 may be visually invisible.

Audit lesson: visual agreement depends on scale and resolution.

Audit case 11: exact versus decimal form

A tool returns 0.3333333333 where the exact answer is 1/3.

The decimal may be adequate for a numerical application. In algebraic manipulation or proof, exact form may preserve information and avoid accumulated rounding error.

Audit question: what precision does the task require?

Audit case 12: decimal precision exceeds input precision

A length measured to the nearest centimetre is entered into a calculator, which returns an area to ten decimal places.

The extra digits are computational output, not measurement evidence.

Audit lesson: reported precision should respect input quality and task conventions.

Audit case 13: units omitted from a correct number

Candidate answer: 12.

If the problem asks for area, the result needs square units. If it asks for speed, it needs distance per time.

Units are part of mathematical meaning, not formatting.

Audit case 14: mixed units accepted by the formula

Speed = distance/time. A candidate uses 120 km and 90 minutes to report 120/90=1.333 “km/h”.

The arithmetic quotient is kilometres per minute. Convert 90 minutes to 1.5 hours:

120/1.5=80 km/h.

Audit lesson: formulas do not police unit compatibility automatically.

Audit case 15: derivative correct, interpretation wrong

A tool differentiates s(t) correctly to obtain velocity v(t). The explanation then calls v(t) “distance travelled”.

The symbolic calculation is correct; the interpretation is not.

Audit lesson: label the mathematical object before reading the prose explanation.

Audit case 16: antiderivative without verification

If a tool proposes F(x) as an antiderivative of f(x), differentiate F. This is often the cheapest independent check.

For indefinite integration, also check whether the family needs an arbitrary constant.

Audit case 17: definite integral sign ignored

A negative definite integral over an interval may represent signed area or accumulated change, not necessarily geometric area.

Audit question: what quantity is the integral representing in context?

Audit case 18: matrix multiplication dimension error

Before multiplying matrices, write dimensions. An m×n matrix can multiply an n×p matrix.

A tool may reject an invalid product, but the learner should know why the operation is undefined.

Audit case 19: inverse matrix assumed to exist

A symbolic solution writes x=A^{-1}b.

This assumes A is invertible. If det(A)=0, the inverse does not exist and the system may have no solution or infinitely many.

Audit lesson: symbolic shorthand often carries hidden existence conditions.

Audit case 20: regression line accepted without shape check

A software package returns a linear regression equation and correlation coefficient.

Before interpreting them:

• plot the data;
• look for curvature;
• inspect influential points;
• check whether a linear model matches the reader job.

Calculation availability does not establish model suitability.

Audit case 21: high correlation turned into causal language

A generated explanation says “X causes Y because correlation is high.”

Repair: correlation describes association, not causal direction by itself. Causal claims require a design and evidence capable of supporting them.

Audit case 22: p-value paraphrased incorrectly

A candidate explanation says a p-value is “the probability the null hypothesis is true”. That is not the standard interpretation.

Audit lesson: statistical terminology should be checked against definitions, not trusted because the prose sounds fluent.

Audit case 23: universal claim supported only by examples

A candidate proof checks n=1,2,3,4,5 and concludes the identity holds for every positive integer.

The examples support a conjecture. A universal proof still needs an argument covering all allowed n.

Audit case 24: converse error

The theorem says: if a number is divisible by 6, then it is divisible by 3.

The candidate proof establishes: if a number is divisible by 3, then it is divisible by 6.

The reasoning may be internally coherent and still address the wrong logical direction.

Audit case 25: circular proof hidden by paraphrase

A generated explanation restates the conclusion in different words as if it were a premise.

Audit response: list the assumptions separately from the target. If the target appears among the assumptions, the argument is circular.

Audit case 26: theorem named but hypotheses not checked

A candidate invokes the Mean Value Theorem, Pythagoras, Bayes’ theorem or another familiar result.

Audit questions:

• What are the theorem’s hypotheses?
• Are they satisfied?
• Does the theorem support this exact conclusion?

Audit case 27: proof uses an undefined variable

A generated solution introduces k or c without stating what it represents or what set it belongs to.

Repair by declaring the variable and its domain.

Notation should reduce ambiguity, not create it.

Audit case 28: hidden dependence in probability

A candidate multiplies two probabilities as if events were independent.

Audit question: does the first event change the probability of the second? If sampling is without replacement, independence may fail.

Audit case 29: event denominator wrong under conditioning

A solution computes P(A|B) using the original sample-space denominator.

Repair: conditioning restricts the relevant universe to B.

Audit case 30: combination used where order matters

A candidate uses C(n,r) for a role-assignment problem with distinct positions.

If president and vice-president are different roles, order matters. A permutation or sequential count is needed.

Audit case 31: permutation used where order does not matter

The reverse error overcounts committees or unordered selections.

Audit question: would swapping the selected objects create a genuinely different outcome?

Audit case 32: simplifying form changes usefulness

For a quadratic:

• expanded form exposes coefficients;
• factorised form exposes roots;
• completed-square form exposes the vertex.

A computer-algebra system may return any equivalent form. The learner should choose the representation that serves the question.

Audit case 33: numerical equality versus symbolic identity

Two expressions agree at several tested values. That is numerical evidence, not proof that they are identical functions over the full domain.

Use algebraic transformation or another general argument when identity is required.

Audit case 34: hidden discontinuity after simplification

(x²-1)/(x-1) simplifies to x+1 for x≠1.

The simplified expression alone can hide that the original function is undefined at x=1.

Audit lesson: algebraic equivalence may hold only on a restricted domain.

Audit case 35: piecewise behaviour flattened into one formula

A generated solution may compress an absolute-value, threshold or regime-change problem into one expression without preserving cases.

Audit whether a piecewise representation is required to keep the original logic.

Audit case 36: optimisation answer ignores feasibility

A calculus solution finds a stationary point that mathematically minimises a function but lies outside the allowed interval or violates a physical constraint.

Audit endpoints and feasibility conditions before accepting the optimum.

Audit case 37: probability answer outside [0,1]

This is an immediate structural failure. A probability below 0 or above 1 cannot be correct under ordinary probability axioms.

Bounds are cheap verification.

Audit case 38: geometric length negative

A quadratic may produce a negative root for a length variable. The root can be algebraically valid for the equation and inadmissible for the model.

Mathematics must return to context.

Audit case 39: rate sign interpreted incorrectly

A negative rate may represent decrease, direction or outflow rather than “wrong answer”. The interpretation depends on the sign convention declared by the model.

Audit case 40: missing endpoint in interval solution

A tool reports x>2 when the original inequality permits equality at 2, or includes an endpoint where a denominator becomes zero.

Audit open and closed boundaries against the original condition.

Audit case 41: integer problem solved over reals

A symbolic solver returns real solutions, but the problem asks for integers. The final set must be intersected with the required domain.

Audit case 42: real problem silently extended to complex numbers

A tool may return complex roots when the school task requests real solutions only.

Audit the requested number system before interpreting completeness.

Audit case 43: answer generated from an incorrect transcription

A mathematically perfect solution to the wrong equation is still wrong for the user’s problem.

The first audit layer is always input fidelity: compare the tool’s representation with the original prompt, diagram, units and constraints.

Audit case 44: image or OCR transcription risk

If a mathematical expression was read from an image or scan, check superscripts, fraction bars, minus signs, decimal points, radical scope and subscripts before solving.

A tiny transcription error can change the entire problem while leaving the downstream algebra internally consistent.

Audit case 45: generated explanation adds a false reason to a correct step

Sometimes the equation transformation is valid but the prose justification names the wrong theorem or principle.

Audit the reason separately from the result. Correct output does not retroactively validate an incorrect explanation.

Audit case 46: generated answer uses a method not permitted by the task

In an assessment context, a method can be mathematically valid but outside the required representation or tool conditions.

Keep examination-permission rules separate from mathematical validity. The current official assessment source controls what is permitted.

Audit case 47: machine arithmetic can be checked by estimation

If a calculator returns 39,842 for a product of numbers near 200 and 20, the scale is plausible. If it returns 398.42, a decimal or entry error is likely.

Estimation is a fast independent layer because it does not repeat the exact calculation.

Audit case 48: substitution as a universal equation check

For many algebraic solution problems, substitute candidate values back into the original equation—not only the transformed final line.

This catches extraneous roots, lost domain conditions and transcription errors.

Audit case 49: differentiate to audit an integral

Given a proposed antiderivative, differentiation provides a strong inverse check.

Given a proposed derivative, numerical finite differences can provide supporting evidence, though symbolic differentiation remains the exact route where appropriate.

Audit case 50: verify a matrix solution

If a tool proposes vector x as a solution to Ax=b, multiply A x and compare with b.

Verification often costs less than re-deriving the full solution.

Audit case 51: verify a regression prediction within domain

Even if the regression arithmetic is correct, prediction far outside the observed x-range may be weakly supported.

Audit whether the query is interpolation or extrapolation.

Audit case 52: model output exceeds a known physical bound

If a tank model predicts 920 L for a 900 L tank, the calculation may be correct for the equation but the model has crossed its physical validity boundary.

Known bounds are powerful machine-output checks.

Audit case 53: probability simulation versus exact probability

A simulation frequency of 0.497 does not contradict an exact probability of 0.5. Simulation contains sampling variation.

Audit whether the generated explanation treats approximation as exact equality.

Audit case 54: calculator display rounds an exact relationship

Values such as sqrt(2), π and repeating decimals may appear rounded on a display. A proof should not use the rounded display as if it were the exact object.

Audit case 55: hidden branch in inverse trigonometry

A calculator gives a principal angle. A trigonometric equation may have multiple solutions in the required interval.

Audit the periodic structure and stated domain after obtaining the principal value.

Audit case 56: tool output is too advanced to be auditable by the learner

If a learner cannot explain what the returned expression means, verify its domain or check even a simple special case, the tool has outrun the learner’s control.

The solution may still be mathematically correct, but it is not yet useful as learning evidence.

Audit case 57: explanation quality versus answer quality

Separate four scores mentally:

• input representation correct?
• answer correct?
• reasoning valid?
• explanation appropriate for the learner?

A response can succeed in one layer and fail in another.

Audit case 58: missing assumptions in modelling

A generated linear model assumes constant rate without saying so.

Restore the assumption explicitly. Then ask whether the evidence supports it over the intended range.

Audit case 59: sensitivity not checked

A model depends strongly on one estimated parameter. A tiny parameter change produces a large output change.

Audit sensitivity before treating the point estimate as stable.

Audit case 60: final answer does not answer the original question

A tool calculates a probability when the question asks for odds, finds a derivative when the task asks for a stationary point, or computes an area when the question asks for volume.

The final audit returns to the user’s target.

The six-pass solution audit

1. Transcription: Is the mathematical problem represented correctly?
2. Conditions: What domain, units, constraints and assumptions apply?
3. Transformations: Did any step lose or introduce cases?
4. Verification: Can the result be checked independently?
5. Completeness: Are all branches, cases and requested values included?
6. Interpretation: Does the answer fit the original mathematical and contextual question?

Student audit template

Write:

• Original problem: …
• Tool candidate: …
• Domain and conditions: …
• First vulnerable step: …
• Independent check: …
• Corrected final answer: …
• What I would do without the tool: …

Fresh audit task 1

Candidate: solve sqrt(3x-2)=x by squaring and accepting both quadratic roots.

Audit the domain, solve, and substitute every candidate into the original equation.

Fresh audit task 2

Candidate: solve (x²-4)/(x-2)=0 by simplifying to x+2=0 and reporting x=-2.

Audit whether x=2 is a solution, excluded value or removable-hole location. Keep original-domain conditions visible.

Fresh audit task 3

Candidate: a graph shows exactly two roots. Ask what viewing window was used and whether algebra can establish completeness.

Fresh audit task 4

Candidate: a regression line is used to predict far beyond the observed data range. Identify why the calculation can be correct while the inference is weak.

Fresh audit task 5

Candidate: an AI explanation claims “mutually exclusive events are independent”. Use the definitions to test the statement.

Fresh audit task 6

Candidate: a formula returns a negative time in a physical problem. Decide whether the negative value represents an earlier time origin, an inadmissible result or a modelling issue.

Release standard for tool-assisted Mathematics

A tool-assisted solution is not ready merely because it looks polished. The learner should be able to:

• state the problem correctly;
• identify domains and assumptions;
• locate any non-reversible step;
• check a candidate independently;
• explain whether all solutions were found;
• interpret units and context;
• distinguish numerical evidence from proof;
• say what the tool did and what the learner still had to decide.

Closing principle

Automation can reduce calculation and search. It does not remove mathematical responsibility.

The stronger the tool, the more important it becomes to preserve the learner’s ability to audit the input, restore hidden conditions, verify the output and know what remains unproved.

Use How to Check Maths Answers for the wider verification system and How to Use a Calculator in Maths for calculator-specific control.

Final Technology Audit: Mathematical Ownership After a Tool Gives an Answer

Release question 1: did the tool solve the problem I actually asked?

Compare the original prompt with the tool’s internal representation. Check signs, exponents, fraction bars, units, graph labels, intervals and conditions. A perfect solution to a mistranscribed problem is still wrong.

Release question 2: what assumptions were added silently?

Look for hidden assumptions such as:

• a denominator is non-zero;
• a square root uses the real domain;
• a rate is constant;
• events are independent;
• a function is differentiable;
• a matrix is invertible;
• a model remains valid outside the observed range.

Write them down before accepting the output.

Release question 3: did any transformation change the solution set?

Squaring, multiplying by expressions that may be zero, taking logarithms, applying inverse functions and clearing denominators can introduce or remove candidates.

Audit non-reversible steps separately.

Release question 4: is the answer complete?

A tool can return one numerical root, one principal angle or one branch. The question may require all solutions in an interval or a general family.

Correctness and completeness are separate tests.

Release question 5: can I verify the answer cheaply?

Prefer a check that is independent of the vulnerable step:

• substitute roots into the original equation;
• differentiate an antiderivative;
• multiply a matrix solution back into the system;
• estimate an arithmetic result;
• check units;
• compare with a graph;
• test a special or boundary case.

Release question 6: does the answer respect known bounds?

Probabilities belong in [0,1]. Lengths and physical quantities may have non-negativity or capacity constraints. Angles may be restricted to a stated interval. Model outputs may have practical limits.

Bounds are fast and powerful checks.

Release question 7: is the explanation logically valid?

A fluent explanation can contain a converse error, circular argument, missing condition or unsupported theorem reference.

Read the proof line by line. Ask what each conclusion depends on.

Release question 8: does the representation help the task?

Equivalent expressions can have different usefulness. Factorised form may reveal roots; completed-square form may reveal a vertex; expanded form may reveal coefficients.

Do not equate “different from my expected form” with “wrong”, and do not equate “simplified” with “best”.

Release question 9: is the output exact or approximate?

A decimal approximation can be useful while still losing exact information. The learner should know whether the task expects an exact expression, a numerical approximation, or a stated degree of precision.

Release question 10: what remains my responsibility?

Even when a tool performs the arithmetic, the learner still owns:

• problem interpretation;
• method suitability;
• domain and assumptions;
• verification;
• completeness;
• context and units;
• final communication.

Audit drill A: wrong input, right algebra

Original: 3x-7=11.

Tool input: 3x+7=11.

The tool may correctly return x=4/3 for its input. The original equation gives x=6.

Input fidelity is the first verification layer.

Audit drill B: right root, incomplete solution set

Original: sin x=1/2 for 0°≤x≤360°.

A calculator returns principal value 30°. The interval also contains 150°.

The tool output is a valid candidate, not a complete interval solution.

Audit drill C: right line, poor model

A regression tool returns a line of best fit. The scatter plot is strongly curved.

The calculation can be correct for linear regression while the linear model remains poorly suited to the data.

Model choice is a mathematical decision above the calculation layer.

Audit drill D: right derivative, wrong requested quantity

The problem asks for the maximum value of a function. The tool returns the derivative.

Differentiation is only an intermediate step. The learner still has to find stationary points, inspect the allowed domain and compare candidates.

Audit drill E: right probability formula, wrong independence assumption

A tool multiplies stage probabilities using a fixed success probability. The original sampling is without replacement.

The formula is valid under a model that the problem does not satisfy.

Audit drill F: right simplification, hidden hole

(x²-9)/(x-3) simplifies to x+3 for x≠3.

The simplified expression is not the same function at x=3 if the original was undefined there.

Domain survives simplification.

Audit drill G: right numerical answer, unjustified precision

A measured input is given to two significant figures. The tool returns ten decimal places.

The extra digits are computational, not evidential.

Audit drill H: right theorem name, missing hypotheses

A solution says “by the Mean Value Theorem” but never establishes the continuity and differentiability conditions required on the relevant interval.

The name of the theorem is not enough. Its jurisdiction must be checked.

Audit drill I: right examples, no proof

A candidate checks 100 integer cases and concludes a universal identity.

The checks support a conjecture. A general argument is still required unless the domain itself contains only those cases.

Audit drill J: right proof direction?

If the task is P⇒Q, confirm that the solution does not accidentally prove only Q⇒P.

Logical direction is easy to lose inside fluent prose.

Student release template

Before submitting tool-assisted work, complete:

• I entered: …
• The tool returned: …
• The required domain/conditions are: …
• The most vulnerable step is: …
• My independent check is: …
• The final answer to the original question is: …

Tutor release standard

Tool use is becoming mathematically controlled when the learner can explain:

• why the input matches the problem;
• what assumptions the method needs;
• which transformations need candidate checking;
• how the output was independently verified;
• whether all cases were included;
• what the result means in context.

Closing principle

The purpose of mathematical technology is not to remove the learner from the reasoning loop. It is to move human attention toward the layers that automation does not settle safely by itself: representation, assumptions, method choice, verification and meaning.

A trustworthy tool-assisted solution is one the learner can audit.