Evidence and records roles: study guide

Worked examples and pitfalls

Word-by-word or letter-by-letter: two correct answers, one rule: File these four names: Van Dyke, Vandenberg, Van Horn, Vance. Under word-by-word filing, each space-separated unit is compared in turn and 'nothing files before something', so the first unit 'Van' sorts ahead of both 'Vance' and 'Vandenberg'. The order is Van Dyke, Van Horn, Vance, Vandenberg. Under letter-by-letter filing, spaces are ignored entirely, so the keys are VANCE, VANDENBERG, VANDYKE, VANHORN, and the order is Vance, Vandenberg, Van Dyke, Van Horn. Both orders are correct filing; only one is correct for the system you are working in. Test items state the convention in the instructions, and the candidates who lose marks are almost always the ones applying whichever convention their previous employer used. The same fork appears with prefixes and punctuation: whether Mc and Mac interfile, whether St is treated as Saint, whether a hyphen counts as a space. Read the rule, restate it to yourself in one sentence, then apply it mechanically and do not let a name that 'obviously' belongs somewhere override it.

Alphanumeric codes: A-102 at the front or the back of the drawer: Sort the references A-7, A-12, A-70, A-102, B-3. Under natural numeric ordering, the way a person reads them, the answer is A-7, A-12, A-70, A-102, B-3. Under plain string ordering, the way most software sorts by default, each character is compared in turn, so '1' comes before '7' and the answer is A-102, A-12, A-7, A-70, B-3, with A-102 first rather than last. Both are defensible; a filing item is testing which one the stated system uses. This is also why serious record systems zero-pad their references: rewrite the set as A-007, A-012, A-070, A-102 and the string sort and the numeric sort agree, permanently. If you are ever asked to design or clean a reference scheme, pad to a fixed width and the whole class of problem disappears. In a test, the giveaway is a set that deliberately mixes one-, two- and three-digit suffixes; that mixture exists only to separate candidates who apply the stated rule from candidates who apply the intuitive one.

Coding to a key: every mark is at a boundary: A claims routing key: under 500 pounds with no injury reported goes to CS1; under 500 with injury goes to CI2; 500 to 4,999 with no injury goes to CS3; 500 to 4,999 with injury goes to CI4; 5,000 and over goes to CE5 regardless of injury. Now code five records. R-118 at 499.99, no injury: CS1, since 499.99 is under 500. R-119 at exactly 500.00, no injury: CS3, because 'under 500' excludes 500 itself and the second band starts there. R-120 at 4,999.00 with injury: CI4, since the band is inclusive at its top. R-121 at exactly 5,000.00 with no injury: CE5, because the top band is 'and over' and its 'regardless of injury' clause overrides the injury split that governs the lower bands. R-122 at 86.40 with injury: CI2. Four of those five decisions turn on a boundary, and that is not an accident, coding items are written so that the interior cases are trivial and every discriminating mark sits on an edge. Before coding anything, underline the boundary words: under, up to, and over, between, inclusive. Then decide, once, what each one does to the endpoint, and apply that decision to every record in the batch.

Duplicates that are not textually identical: Three rows arrive in a merge. Row 1: SMITH, JANE | 07/04/1988 | AB123456C. Row 2: Smith, Jane | 1988-04-07 | AB 123456 C. Row 3: SMITH, JANE | 04/07/1988 | AB123456C. Rows 1 and 2 are the same person: normalise case, strip the spaces from the reference, and convert the ISO date and they match exactly. Row 3 is the genuinely hard one. If the file is in day/month order it is a different date of birth and possibly a different person; if that row came from a system using month/day order it is the same record again. You cannot tell from the row itself, so the correct action is to send it to the exception queue with the ambiguity noted, not to merge it and not to discard it. This is the judgement that separates competent records work from the appearance of it: the goal is not to make every row disappear, it is to make every decision defensible. In test form the item usually asks 'how many distinct individuals are represented', and the answer is often given as a range or accompanied by a 'cannot determine' option for exactly this reason.

Where accuracy actually decays on a long batch: Run a self-measurement rather than trusting a number from anyone: take 200 record pairs, split them into four blocks of 50, and log errors per block. Most people find block 1 slightly worse than block 2, a warm-up cost, and then a rise across blocks 3 and 4 as vigilance falls. The reason to measure it is that the arithmetic of small percentages is brutal at volume. Checking 200 pairs at 98 percent accuracy passes 4 bad records; at 99.5 percent it passes 1. Scale that to a realistic month of 20,000 records and the same two accuracy rates mean 400 defects against 100: a four-fold difference in downstream rework from a 1.5 point difference that would look like noise on a single test. Once you know where your own curve turns, the intervention is cheap: a deliberate twenty-second break at that point, or splitting the batch so the hardest records fall in your strongest block. Practice sessions on this site are recorded on your own device, so building a block-by-block picture across several sessions costs nothing but the logging.

Setting an attempt rate from the scoring rule: A 120-item checking test with a 10-minute limit, scored as correct minus incorrect. Suppose practice has told you that you hold 92 percent at ten items a minute and 96 percent at seven and a half. Fast: 10 x 10 = 100 attempted, 92 correct and 8 wrong, score 84. Careful: 10 x 7.5 = 75 attempted, 72 correct and 3 wrong, score 69. Fast wins by 15. Now change the rule to correct minus three times incorrect: fast scores 92 - 24 = 68, careful scores 72 - 9 = 63, and the 15-point gap has shrunk to 5. Now suppose your real accuracy at ten a minute is 80 percent rather than 92. A gap most people do not discover until they measure it. Fast now gets 80 right and 20 wrong: under simple correct-minus-incorrect that is 60, already behind the careful strategy's 69, and under the triple penalty it is 80 - 60 = 20 against 63. Same test, same person, opposite advice, and the two deciding variables are the penalty multiplier, which the instructions hand you, and your own accuracy-at-speed, which only measurement gives you. Do not pick a pace from temperament. Measure two rates in practice, write both accuracy figures down, and do this arithmetic before the test rather than during it.

How to practise this skill

• Before the first record of any batch, write the filing or coding rule at the top of the page in your own words. The single largest source of clerical error is applying a remembered rule from a previous system.
• Underline the boundary words in a coding key (under, up to, and over, inclusive), and resolve each endpoint once. Interior cases carry almost no marks; the edges carry nearly all of them.
• Normalise before you compare: mentally strip case, spaces and punctuation from references, and restate dates in one fixed order. Half of apparent duplicates are formatting, and half of apparent non-duplicates are too.
• When a record is genuinely ambiguous, mark it and move on. Time spent resolving one unresolvable row is taken from thirty rows you could have done correctly, and a guessed merge is worse than a flagged one.
• Measure your accuracy at two different speeds in practice and write both numbers down. You cannot choose an attempt rate rationally without them, and the fast rate is almost never as accurate as it feels.
• Practise on batches long enough to hit your own fatigue point, not on ten-item samples. The construct is specifically about sustained accuracy, and a ten-item drill measures the part of the curve that was never in doubt.

Glossary

Word-by-word filing

An alphabetical convention that compares space-separated units in turn, with a shorter first unit filing before a longer one. Under it, Van Horn files before Vance.

Letter-by-letter filing

An alphabetical convention that ignores spaces and punctuation entirely, comparing the run of letters. Under it, Vance files before Van Dyke.

Natural sort

Ordering that reads embedded digit runs as numbers, so A-7 precedes A-12. Contrasted with string ordering, which compares characters one at a time and puts A-102 first.

Zero padding

Writing references to a fixed width by adding leading zeros (A-007 rather than A-7) so that string ordering and numeric ordering produce the same result. The standard structural fix for alphanumeric filing errors.

Primary and secondary sort key

The field sorted on first and the field used to break ties within it. A batch sorted by site then by surname will look wrong if the two keys are applied in the opposite order.

Canonicalisation

Converting records to a single standard form (one case, no stray whitespace, one date order, punctuation stripped from references) before any comparison, so that formatting differences stop masquerading as data differences.

Exception queue

The destination for records that cannot be resolved from the information available. Routing an ambiguous record here is a correct outcome; guessing it into a merge is not.

Boundary condition

The endpoint of a coded band, where wording such as under, up to or and over decides which side a value falls. Coding tests concentrate their discriminating items here.

Study this next

• Data checking and clerical accuracy skill hub
• Clerical accuracy practice in the Cognitive Skills Lab
• Error detection lesson
• Administrative and clerical entry suite
• Clerical checking suite
• Data checking and clerical accuracy lesson on Novus Learn

Where this material comes from

• All filing sets, reference codes, routing keys and records above were invented for this lesson. The two filing orders, the two sort orders and every score calculation in the attempt-rate example were worked through and checked by recomputation.
• Filing and sorting conventions cross-checked against standard public descriptions such as the Wikipedia articles 'Alphabetical order', 'Natural sort order' and 'Collation'. Terminology only; the examples are original.
• Novus Learn aptitude construct registry (catalog seed) for construct scope and suite mapping.
• Public educational framing only: not affiliated with any official exam board, publisher or employer, and no copyrighted test item is reproduced.

Worked examples and pitfalls

Read the header: (£000) changes every answer by a factor of a thousand: Table titled 'Regional revenue, year to March (£000)': North 1,240; South 986; East 1,455; West 719. Total = 1,240 + 986 + 1,455 + 719 = 4,400, so the business turned over 4,400 thousand pounds, that is 4.4 million. East's share is 1,455 / 4,400 = 33.1 percent. The whole set: North 28.2 percent, South 22.4, East 33.1, West 16.3, summing to 100. Two things go wrong here. The first is reading East's revenue as 1,455 pounds and then reporting a business with a total turnover of 4,400 pounds, which nobody notices because every option is scaled the same way, until the question asks for revenue in millions and only one option is right. The second is the comparison wording. East's share is (1,455 - 719) / 4,400 = 16.7 percentage points above West's, and East's revenue is (1,455 - 719) / 719 = 102 percent more than West's, that is slightly more than double. 'Sixteen point seven' and 'a hundred and two' both describe the same two cells honestly, and the question decides which one is correct. Note that the percentage-point figure must be computed from the unrounded shares rather than by subtracting the rounded ones, or the last digit will not survive.

One pair of rows, three correct increases: A complaints table: 2023, 120,000 orders, complaint rate 4.0 percent; 2024, 150,000 orders, complaint rate 5.0 percent. Three defensible answers to 'how much did complaints increase?'. The rate rose by 1.0 percentage point. The rate rose by (5.0 - 4.0) / 4.0 = 25 percent in relative terms. And the count of complaints rose from 0.04 x 120,000 = 4,800 to 0.05 x 150,000 = 7,500, which is (7,500 - 4,800) / 4,800 = 56.25 percent. All three are arithmetically right; only one answers the question in front of you. The pattern to internalise is that a rate and a count move together only when the denominator is fixed, and here it is not. Order volume grew 25 percent as well. If the question is about customer experience, the rate is the honest figure; if it is about how many complaint handlers to hire, the count is. Test items usually ask for the one you would not have chosen.

Index numbers: five points is not five percent: A cost index with 2020 = 100 reads 104 in 2021, 111 in 2022 and 109 in 2023. From 2021 to 2023 the index rose 5 points, but the percentage change is 5 / 104 = 4.8 percent, because the base for the comparison is 104, not 100. From 2022 to 2023 it fell 2 points, which is -2 / 111 = -1.8 percent, and note that costs fell even though the index remains 9 percent above the 2020 base. A level and a change are different claims. The only comparison where points and percent coincide is against the base year itself: 2020 to 2023 is 100 to 109, exactly plus 9 percent. Watch also for a rebased series, where a table switches to 2022 = 100 partway down; the two segments cannot be compared directly without converting one of them, and an item that quietly rebases is testing whether you read the column heading.

Joining two tables: totals and per-head figures disagree on purpose: Table 1, headcount by site: Leeds 84, Derby 47, Bristol 129. Table 2, absence days recorded in the same period: Leeds 630, Derby 300, Bristol 903. 'Which site has the worst absence problem?' On raw totals Bristol is worst at 903 days. Normalise per head and the ranking changes: Leeds 630 / 84 = 7.50 days per employee, Bristol 903 / 129 = 7.00, Derby 300 / 47 = 6.38. Leeds is worst, Bristol is merely biggest. The organisation-wide figure is 1,833 / 260 = 7.05 days per head, which is a useful reference line: Leeds is above it, the other two below. The general rule is that any comparison between units of different size demands a denominator, and the denominator has to come from the other table. Items are built so the raw-total answer and the per-head answer are both on the option list, and so the site with the biggest total is never the site with the highest rate.

The truncated axis: measure the numbers, not the bars: A quarterly satisfaction chart with a y-axis running from 78 to 82 shows bars at 79.2, 79.8, 80.4 and 81.1. Visually the last bar looks several times taller than the first, because only the top 4 points of a 100-point scale are drawn. The actual movement is 81.1 - 79.2 = 1.9 points, which on the score's own scale is a relative rise of 1.9 / 79.2 = 2.4 percent. If the question asks 'by approximately what percentage did satisfaction improve', the answer is about 2 percent, and the distractor built from the bar heights will be something like 40 or 400 percent. Related presentation effects to check before answering: a dual-axis chart where two series use different scales and appear to cross meaningfully when they do not; a cumulative series, where a flattening line still means the total is growing, just more slowly; and a logarithmic axis, where equal vertical distances are equal ratios rather than equal amounts. In every case the defence is the same. Find the printed numbers, or read the gridline values, and compute.

Cannot say: revenue is not profit: A product table shows units sold and total revenue. Product P: 4,200 units, 71,400 pounds. Product Q: 1,800 units, 41,400 pounds. Average selling price is 71,400 / 4,200 = 17.00 for P and 41,400 / 1,800 = 23.00 for Q, so Q earns more per unit while P earns more in total. Now the statement to evaluate: 'P is more profitable than Q.' The correct response is cannot say. Profit needs cost, and the table has no cost column; a product with a 17 pound price and a 16 pound unit cost is less profitable than one priced at 23 with a cost of 9, and nothing here rules that out. Contrast with 'Q generated more revenue per unit than P', which the table fully supports and which is true. The habit worth building is to finish every cannot-say judgement with the missing input named out loud ('cannot say, because unit cost is not given') because that forces you to distinguish a genuinely unanswerable item from one you simply have not worked hard enough on.

How to practise this skill

• Read the title, the units line, the row and column headers and any footnote before you look at a single value. Roughly the first fifteen seconds of an item should contain no arithmetic at all, and that fifteen seconds is what prevents the thousand-fold and percentage-point errors.
• For every item, write down which of the three quantities is wanted (share of total, relative change, or change in percentage points), before computing. Most wrong answers on this construct are correct arithmetic applied to the wrong quantity.
• Whenever two groups differ in size, ask what the denominator should be. If a question compares sites, teams, countries or periods of unequal length and you have not divided by something, you are almost certainly answering the wrong question.
• Practise the cannot-say items separately and force yourself to name the missing variable each time. Candidates who train only on computational items reliably over-answer inference statements under time pressure.
• Do not redraw or re-scale charts in your head. Locate the gridline values or the data labels, and if neither exists, interpolate between two labelled gridlines and state the bound rather than guessing a point value.
• Time yourself per item rather than per section. Data interpretation sets share a stimulus, so the first item costs the reading time and the rest should be fast; if item four takes as long as item one, you did not build a mental map of the table.

Glossary

Units multiplier

A scaling note in a table title or column header, such as (£000), (millions) or (per 1,000 population). It applies to every value in scope and is the single most common source of order-of-magnitude errors.

Index number

A series rescaled so a chosen base period equals 100. Changes between two non-base periods must be divided by the earlier value, so a movement in index points is not a percentage change except when measured from the base.

Rebasing

Restating an index against a new base period. Segments of a series with different bases cannot be compared directly, and a table that rebases partway down is testing whether you read the headings.

Truncated axis

A chart whose value axis does not start at zero, which exaggerates the apparent size of differences between bars or points. Legitimate for showing small movements in a large quantity, misleading if read as area or height.

Cumulative series

A line showing a running total rather than each period's value. It can only go up or stay flat, so a flattening cumulative line means the periodic figure is falling, not that the total is.

Weighted average

An average in which each value is multiplied by the size of the group it represents. Averaging two group percentages directly is only correct when the groups are the same size, which in these tables they rarely are.

Normalisation

Dividing a raw figure by an exposure measure (headcount, units sold, population, days open), so groups of different size can be compared. The denominator usually lives in a second table.

Cannot say

The verdict when a statement is neither supported nor contradicted by the data supplied. A correct cannot-say answer can always be defended by naming the specific missing variable.

Study this next

• Data interpretation skill hub
• Data interpretation practice in the Cognitive Skills Lab
• Numerical reasoning lesson
• Financial data interpretation suite
• Scientific data interpretation suite
• Data interpretation lesson on Novus Learn

Where this material comes from

• Every table, index series and chart described above was constructed for Novus Learn, and each figure was verified by recomputing the totals and the reverse calculation.
• Definitions of index numbers, rebasing and weighted averages cross-checked against standard public references such as the Wikipedia articles 'Index (economics)' and 'Weighted arithmetic mean'. Terminology only; no data or item text is taken from any source.
• Novus Learn aptitude construct registry (catalog seed) for construct scope and suite mapping.
• Public educational framing only: not affiliated with any official exam board, publisher or employer, and no copyrighted test item is reproduced.

Worked examples and pitfalls

The four transcription error families: Source reference: INV-2024-08871. Four corrupted copies, one of each family. INV-2024-08571 is a substitution: a single character changed, 8 to 5. INV-2024-0871 is an omission: one character dropped, and the string is now a character short. INV-2024-088711 is a duplication: a character repeated, one character long. INV-2024-08817 is a transposition: the final 7 and 1 have swapped places, and crucially the string is the same length and contains exactly the same characters. That last property is why transposition is the hardest of the four to see and the most dangerous in practice. Anything that checks length catches omission and duplication. Anything that compares character sets or sums the digits catches substitution but not transposition, because 0+8+8+7+1 and 0+8+8+1+7 both come to 24. Reading the reference aloud catches substitution reliably and transposition poorly, because the ear is comparing sounds and both versions sound similar at speed. The only defences against transposition are a positional comparison and a weighted check digit.

A fixed scan order finds the one field that moved: Source record: Name Priyanka Ramaswamy / Account 4471-9026-3358 / Date of birth 14/03/1991 / Postcode SW1A 2AA / Phone 0161 496 0872. Copy: Name Priyanka Ramaswamy / Account 4471-9026-3358 / Date of birth 14/03/1991 / Postcode SW1A 2AA / Phone 0161 469 0872. The difference is in the phone field, where 496 has become 469: a transposition inside the middle block. Most candidates find it eventually; the ones who find it in eight seconds are the ones running a procedure. Always compare in the same field order, top to bottom, never skipping a field because it 'looks fine'. Compare long strings in the printed blocks rather than as a single run (4471, then 9026, then 3358) because short-term memory holds about four items and a twelve-digit run does not fit. Say the block silently, look, compare, move on. On a same-or-different item you may stop at the first mismatch; on a 'how many fields differ' item you must complete every field, and the instruction wording is what tells you which regime you are in. Getting this wrong in either direction costs marks: stopping early on a count item, or exhaustively checking a same-or-different item you had already resolved.

Modulus 11: why a weighted check digit catches a transposition: The ISBN-10 scheme multiplies the ten digits by descending weights 10, 9, 8 down to 1 and requires the total to be divisible by 11. Take 0-306-40615-2. Working left to right: 10x0 = 0, 9x3 = 27, 8x0 = 0, 7x6 = 42, 6x4 = 24, 5x0 = 0, 4x6 = 24, 3x1 = 3, 2x5 = 10, and the check digit 1x2 = 2. The sum is 0 + 27 + 0 + 42 + 24 + 0 + 24 + 3 + 10 + 2 = 132, and 132 = 12 x 11 exactly, so the number is valid. Now transpose the 1 and the 5 to give 0-306-40651-2. The digits are identical, so a plain digit sum is unchanged at 27 either way and would report no problem. The weighted sum, however, becomes 0 + 27 + 0 + 42 + 24 + 0 + 24 + 15 + 2 + 2 = 136, and 136 - 132 = 4, so it is not a multiple of 11 and the record is rejected. That is the entire reason check digits are weighted rather than plain: position has to matter, or the commonest human error passes straight through.

Luhn: the number that fails, and the one case it misses: The Luhn check, used on payment and many membership numbers, doubles every second digit counting from the right, subtracts 9 from any doubled result above 9, sums everything, and requires a total ending in zero. Take 4539 1488 0343 6467. The doubled positions contribute 3, 3, 8, 0, 7, 2, 6 and 8, totalling 37; the undoubled positions contribute 7, 4, 3, 3, 8, 4, 9 and 5, totalling 43. The grand total is 80, which ends in zero, so the number passes. Now transpose the last two digits to 4539 1488 0343 6476. The doubled contributions become 5, 3, 8, 0, 7, 2, 6, 8 = 39 and the undoubled become 6, 4, 3, 3, 8, 4, 9, 5 = 42, giving 81. Not a multiple of ten, so the mistyped number is rejected at the point of entry. Luhn catches every single-digit substitution and almost every adjacent transposition: with exactly one blind spot. Swapping an adjacent 0 and 9 leaves the total unchanged, because doubling 0 gives 0 and doubling 9 gives 18 which reduces to 9, so both digits contribute the same amount whether doubled or not. A form that validates with Luhn will happily accept 90 where you typed 09. Knowing the blind spot is the point: a check digit narrows the space of undetected errors, it never closes it.

Negative marking: 44 right can beat 46 right: Many clerical checking sections score correct minus incorrect, with omissions neutral. Under that rule, guessing between two remaining options has an expected value of 0.5 x (+1) + 0.5 x (-1) = 0, exactly neutral, and guessing blindly among four options has an expected value of 0.25 - 0.75 = -0.5, clearly negative. Concretely: on a 60-item section you attempt 48, get 44 right and 4 wrong, and score 40. A colleague attempts all 60, gets 46 right and 14 wrong, and scores 32. More correct answers, a lower score. Reverse the scoring rule to plain raw-correct with no penalty and the ordering flips: 46 beats 44 and leaving a blank becomes strictly irrational. Neither strategy is universally right, which is why the instruction screen is not optional reading. Find the sentence that says whether wrong answers are penalised, and if it is absent, assume raw scoring and answer everything.

Reconcile the batch: consistent is not the same as correct: An invoice lists three lines: 12 at 14.25 = 171.00; 7 at 33.60 = 235.20; 3 at 128.00 = 384.00. The stated subtotal is 709.20, VAT at 20 percent is stated as 141.84, and the total is stated as 851.04. Every downstream figure checks out against the one above it: 709.20 x 0.20 = 141.84 and 709.20 + 141.84 = 851.04. Nothing in the tax or total column is inconsistent. But recompute the lines: 171.00 + 235.20 + 384.00 = 790.20, not 709.20. The subtotal is a transposition of the correct figure, and because every later number was calculated from the wrong subtotal, the document is perfectly self-consistent and perfectly wrong. The true VAT should be 158.04 and the true total 948.24, a shortfall of 97.20. This is the single most valuable habit in the construct: never verify a total against the number printed above it, always recompute it from the underlying items. Internal consistency proves that one person did the arithmetic carefully; it proves nothing about whether they started from the right number.

How to practise this skill

• Drill transposition specifically. Generate pairs where the only difference is two adjacent characters swapped, because that is the family your eye is worst at and the family a naive check will not catch.
• Fix a scan order and never vary it: field by field, top to bottom, and within a long value block by block. Free-roaming comparison feels faster and reliably misses one field per record.
• Learn one check-digit scheme by hand, modulus 11 is the easiest, until you can run it in about twenty seconds. It converts a whole class of 'does this record look right' items from judgement into arithmetic.
• Read the scoring rule before the first item and decide your guessing policy then, not at minute seven when you are behind. Under penalty scoring, an omission is a legitimate answer; under raw scoring it is a wasted mark.
• When you are checking financial or tabular documents, recompute the lowest-level figures first and work upward. Errors introduced at a subtotal propagate perfectly and are invisible from below.
• Log the errors you miss, not the ones you find. A miss log after four sessions usually shows a personal signature (always the middle of long numeric strings, or always the second occurrence of a repeated field), and that signature is what to drill.

Glossary

Transposition error

Two adjacent characters exchanged, as in 496 typed for 469. Length and character content are unchanged, so length checks and plain digit sums both pass it; only positional comparison or a weighted check digit detects it.

Substitution error

One character replaced by another, as in 08571 for 08871. It is the family most reliably caught by reading a value aloud against the source, and by any check digit scheme.

Check digit

An extra digit computed from the others so that a corrupted record fails an arithmetic test. It detects errors; it never corrects them and never proves the record refers to the right person or item.

Modulus 11

A weighted check scheme in which digits are multiplied by descending position weights and the total must divide by 11. Used in ISBN-10 and several banking and national identifier formats.

Luhn algorithm

A modulus 10 check that doubles alternate digits from the right and requires the sum to end in zero. It catches all single-digit errors and most adjacent transpositions except an adjacent 0 and 9.

Miss and false alarm

A miss is a genuine discrepancy passed as correct; a false alarm is a difference flagged where none exists, usually a formatting variant such as a trailing space or a differently written date. Misses are the costlier of the two in records work, which is why procedures normalise formatting first and then bias toward escalating doubt.

Correction for guessing

A scoring rule that subtracts a fraction or multiple of the wrong answers from the correct ones, making blind guessing negative in expectation and turning omission into a rational choice.

Reconciliation

Recomputing an aggregate from its components rather than accepting the printed figure. It is the only check that catches an error introduced at summary level, where everything below and above it still agrees.

Study this next

• Error detection skill hub
• Error detection practice in the Cognitive Skills Lab
• Data checking and clerical accuracy lesson
• Clerical checking suite
• Copyediting and proofreading suite
• Error detection lesson on Novus Learn

Where this material comes from

• The ISBN-10 modulus 11 worked example (0-306-40615-2, weighted sum 132) and the Luhn worked example (4539 1488 0343 6467, sum 80) were computed by hand for this lesson and each was re-verified digit by digit, including the corrupted variants.
• Algorithm definitions cross-checked against the public specifications described in the Wikipedia articles 'International Standard Book Number' and 'Luhn algorithm', including the documented 09/90 transposition blind spot. Records, invoices and reference numbers above are invented.
• Novus Learn aptitude construct registry (catalog seed) for construct scope and suite mapping.
• Public educational framing only: not affiliated with any official exam board, publisher or employer, and no copyrighted test item is reproduced.

Worked examples and pitfalls

Worked scenario: an unverified safety concern: A colleague reports a possible equipment fault while a deadline is approaching. First distinguish the known fact, the report, from the unverified cause. A strong response protects people and affected work, checks the concern through the right channel, tells the relevant lead, and records what was done. Ignoring the report underreacts; shutting down unrelated work or accusing someone before checking the facts overreacts.

Method: facts, duties, risks, response: Write four short notes before ranking options: what is known, who may be affected, which duty or boundary applies, and what safe next step is available now. Prefer an action that addresses the immediate issue and creates useful follow-through. Do not reward an option merely because it sounds decisive.

How to practise this skill

• Answer the question asked: best first action, worst action, or complete response are different tasks.
• Check whether an option acts within the person's authority and escalates only as far as the risk requires.
• When two options look reasonable, prefer the one that gathers missing facts and communicates ownership.

Glossary

Proportionality

Matching the urgency and scope of a response to the evidence, likely impact, and authority available.

Role boundary

The limit of what a person may decide or do without approval, specialist help, or escalation.

Follow-through

Confirming ownership, recording the decision, and checking that the issue was actually resolved.

Study this next

• Situational judgement test
• Workplace ethics
• Decision-making
• Situational judgment lesson on Novus Learn

Where this material comes from

• Novus Learn original situational-judgment suite scenarios and published construct mapping.
• Novus educational framework: facts, duties, risks, role boundaries, proportional action, and follow-through.

Worked examples and pitfalls

The negation test finds assumptions; nothing else does: Argument: 'The council should fit bin sensors across the district. In the trial depot they cut collection trips by a fifth.' Candidate assumption A: 'The trial depot's waste pattern is broadly representative of the rest of the district.' Negate it - suppose the trial depot is nothing like the rest of the district - and the argument collapses, because a result that does not transfer supports nothing about the district. So A is a required assumption. Candidate assumption B: 'Bin sensors are the cheapest available technology for this purpose.' Negate it - suppose they are the most expensive - and the argument still stands as given, since it argued from a reduction in trips, not from cost. B is not required; it would strengthen a cost-based argument that was never made. Candidate C: 'Fewer collection trips reduce total operating cost.' Negate it and the recommendation loses its point, so C is required too, even though the argument never mentions cost. That last case is the one candidates miss, because the required assumption is what bridges the evidence to the recommendation, and bridging assumptions are by definition unstated. The failure mode across this whole item type is picking statements that support the conclusion instead of statements the argument cannot survive without.

Base rates: 90 percent accurate, 15 percent right: A production line runs an automated defect test. Defects occur in 1 percent of units. The test flags 90 percent of genuinely defective units, and wrongly flags 5 percent of good units. A unit has just been flagged - how likely is it to be defective? Work in whole units rather than probabilities. Take 10,000 units: 100 are defective and 9,900 are good. Of the 100 defective, the test flags 90. Of the 9,900 good, it wrongly flags 5 percent, which is 495. Total flags: 90 + 495 = 585. Of those, 90 are genuine, so the answer is 90 / 585 = 15.4 percent. Better than the 1 percent base rate, and nowhere near the 90 percent that intuition offers. The error has a name - confusing the probability of a flag given a defect with the probability of a defect given a flag - and it has a practical consequence: the rework queue has to be sized for 585 units a batch, not 100, and 495 of the 585 units sitting in it are perfectly good. Whenever a test, screening rule or model accuracy figure appears in an item, ask how many of the negatives there are, because with a rare condition the false positives from a large clean population swamp the true positives from a small affected one.

Three alternative explanations, and the comparison that settles it: Claim: 'Employees who attended the optional resilience workshop took 30 percent fewer sick days last year, so the workshop works.' Alternative one, selection: attendance was optional, so the people who signed up may have been the healthier and more engaged to begin with, and would have taken fewer sick days regardless. Alternative two, reverse direction: employees who were frequently unwell were the least able to attend a full-day workshop, so illness determined attendance rather than the other way round. Alternative three, a common third factor: the workshop ran on Thursday afternoons at head office, so attendance largely tracks being an office-based rather than shift-based worker, and shift workers take more sick leave for reasons that have nothing to do with resilience training. What would settle it: randomly assigning the offer to half the workforce and comparing the two groups. Failing that, the cheapest useful check already exists in the payroll data - compare attendees' and non-attendees' sick days in the year before the workshop. If attendees were already 30 percent lower, selection is doing the entire job. The habit worth building is to name the comparison group before you accept any before-and-after number.

Deduction: affirming the consequent versus modus tollens: Rule: 'If an invoice is over 10,000 euro, it requires two signatures.' Argument A: 'This invoice has two signatures, therefore it is over 10,000 euro.' Invalid - this is affirming the consequent. The rule makes two signatures necessary for large invoices; it says nothing that prevents a cautious manager from double-signing a 400 euro invoice, so a two-signature invoice of any size is consistent with the rule. Argument B: 'This invoice has only one signature, therefore it is not over 10,000 euro.' Valid - this is modus tollens, denying the consequent to deny the antecedent, and it holds assuming the rule was followed. Argument C: 'This invoice is not over 10,000 euro, therefore it does not have two signatures.' Invalid again, denying the antecedent. Two further points that deduction items test directly. First, validity is about form alone: 'All banks close on Sundays; Riverton Mutual is a bank; therefore Riverton Mutual closes on Sundays' is perfectly valid even if the first premise is false, and a valid argument with a false premise can deliver a false conclusion. Second, in these items you must accept the premises as given even when you know them to be untrue, because the question is whether the conclusion follows, not whether it is true.

Interpretation: what a survey figure does and does not license: Data: all 1,200 employees of a firm were surveyed. 62 percent said they would accept a four-day week at 90 percent pay. Of those who said yes, 71 percent were under 35. Conclusion one: 'A majority of this firm's employees would accept the trade.' This follows - 62 percent of a complete census of the firm is a majority, and the conclusion is properly limited to this firm. Conclusion two: 'A majority of the firm's under-35 employees would accept the trade.' This does not follow, and the arithmetic shows why. Sixty-two percent of 1,200 is 744 accepters, and 71 percent of 744 is about 528 accepters aged under 35. That is the composition of the accepters, not the acceptance rate within the under-35 group, and without knowing how many under-35s the firm employs the rate is undetermined: if there are 900, the rate is 528 / 900 = 59 percent; if there are 600, it is 88 percent; if there were 1,100 the rate would be below half. Conclusion three: 'The firm should move to a four-day week.' This does not follow either - a stated preference about pay says nothing about output, shift cover, customer hours or cost. Notice that conclusion two is the base-rate error from the screening example, wearing survey clothing.

Strong and weak arguments on the same question: Question: should the agency publish raw inspection scores for every site? Strong argument for: 'In the neighbouring authority, publication was followed by a measurable fall in repeat violations, which is the stated aim of the inspection regime.' It is directly about this decision, it addresses the regime's own objective, and it offers evidence - and if true it would change your view. Weak argument for: 'Yes, because transparency is always good.' Sweeping, unsupported, and it would apply identically to publishing anything at all, which is the tell. Strong argument against: 'Raw scores omit the severity weighting, so a site with one critical failure can rank above a site with six minor ones, and a member of the public cannot see the difference.' Specific, mechanistic, directly about the proposal, and it would change your view. Weak argument against: 'No, because businesses will object.' It may be true and it may matter politically, but it does not bear on whether publication achieves the regime's aim, and no evidence is offered. Two tests separate strong from weak: is the argument about the exact question asked, and would accepting it change the decision? Note that you can hold a firm personal view on publication and still be required to mark the argument on the other side as the strong one - that split is what the item type is measuring.

How to practise this skill

• Negate every candidate assumption out loud. If the argument survives the negation, it was never an assumption, however supportive it sounds. This single habit converts the assumption item type from guesswork into a mechanical check.
• Write the conclusion in your own words before reading any option. Half of the wrong answers on inference and evaluation items are responses to a conclusion the argument never reached.
• Ask 'percentage of what?' on every percentage in the stimulus and reconstruct the denominator in whole units. Base-rate and composition errors both dissolve the moment you write out counts instead of rates.
• Deliberately practise marking arguments you disagree with as strong and arguments you agree with as weak. Assessments in this construct are built to catch agreement masquerading as evaluation, and the effect is largest on politically loaded stimuli.
• For any causal claim, write two alternatives - selection and a third factor - and name the comparison group that would rule them out, before you decide whether the evidence supports the claim.
• Log wrong answers by task type rather than by topic: inference, assumption, deduction, interpretation, evaluation. Candidates are rarely weak across all five, and the profile tells you where the next hour of practice belongs.

Glossary

Conclusion

The claim an argument is trying to establish. It is not always last, is often signalled by therefore, so, or should, and locating it correctly determines every subsequent judgement about the argument.

Assumption

An unstated premise the argument requires in order to work. Identified by negation: negate it and a genuine assumption brings the argument down, while a merely helpful statement leaves it standing.

Validity and soundness

An argument is valid when the conclusion cannot be false while the premises are true, which is a property of form alone. It is sound when it is valid and the premises are actually true.

Affirming the consequent

The invalid pattern 'if P then Q; Q; therefore P'. It is the most common deductive distractor because it differs from the valid modus tollens form by only the position of a negation.

Base rate

How common something is in the population before any test or evidence is applied. Ignoring it makes accurate-sounding tests appear far more informative than they are when the condition is rare.

Confounder

A third factor associated with both the supposed cause and the outcome, capable of producing the entire observed relationship on its own. Ruled out by randomisation or by an explicit comparison group.

Selection effect

A difference between groups created by how people entered them rather than by the treatment under study. Optional programmes and voluntary surveys are where it appears most often.

Falsifiability

The property of a claim that some observation could show it to be wrong. A claim compatible with every possible result carries no information, which is why 'what would change your mind?' is a diagnostic question.

Study this next

• Critical thinking skill hub
• Practise the critical thinking bank
• Verbal reasoning lesson
• Government audit suite
• Investigator and detective reasoning suite
• Critical thinking lesson on Novus Learn

Where this material comes from

• Worked items written for Novus Learn. The arguments, survey figures, defect-test numbers and evaluation options above are original and invented for this lesson; no published or copyrighted test item is reproduced.
• The five task types named in the summary - inference, assumption, deduction, interpretation, evaluation of arguments - describe a structure used publicly across several critical-thinking assessments; no affiliation with any publisher is claimed or implied.
• Terminology follows standard, widely published usage in introductory logic and research methods - validity, soundness, affirming the consequent, base rate, confounder, selection effect.
• Novus Learn aptitude construct registry (catalog seed) for the construct scope and the suite mapping shown in the related links.
• Public educational framing only - not affiliated with any official exam board, publisher or employer.