
A useful chart helps someone see a relationship without having to reconstruct every calculation. It also makes the limits of that relationship visible. Creating charts with Super Intelligence therefore begins with a question, a defined dataset and a checking method. The colours come later. A beautifully drawn bar can still compare the wrong things; an accurate percentage can still hide the wrong denominator; a smooth line can still invent observations that were never collected.
This guide teaches the communication part of the workflow: how to turn known data into a faithful visual explanation. You will choose a visual form, preserve the meaning of the measurements, inspect the actual rendered result and write a caption that says only what the evidence supports. The examples include six finished charts with readable source tables, deliberate wrong-chart repairs and practice tasks with complete answers. Every dataset is original, hypothetical teaching data. None describes actual students, an eduKate class, a research finding or the performance of an AI model.
In this series, “Super Intelligence” or “SI” is an editorial term for practical work with contemporary AI systems and tools. It does not mean that present models have been demonstrated to be artificial superintelligence. A model can suggest a chart, write a plotting script or propose a caption. Those suggestions remain proposals until someone verifies the data, calculations, visual encoding and interpretation. The human author remains responsible for the published result.
If you are a student, the goal is to explain your choice without needing the AI conversation beside you. If you are a teacher or parent, the goal is to make the learner’s thinking inspectable. If you are preparing a workplace report, the goal is to help the reader make the right comparison without accidentally implying a causal finding, a representative sample or a certainty that the underlying data cannot support.
For the wider learning sequence, return to How to Learn Super Intelligence Quickly. For cleaning records, checking formulas and preserving spreadsheet inputs, use Super Intelligence with Spreadsheets and Data Analysis. This article concentrates on what happens when those records become marks, scales, labels and claims that another person must read.
Choose your reading route
Start with chart choices, inspect six worked visualisations, use the checking route, or try independent exercises.
Choose the question and visual form
1. Write the reader’s question before choosing a chart
2. Establish the unit of observation and the data contract
Inspect complete worked examples
5. Worked visualisation: compare four resource counts
6. Repair the wrong chart without changing the facts
7. Worked visualisation: counts and completion rates answer different questions
8. Worked visualisation: show a time series with a genuine gap
9. Time spacing, accumulation and smoothing change the story
10. Worked visualisation: inspect paired numerical observations
11. Worked visualisation: reveal the distribution behind an average
12. Worked visualisation: distinguish an observed range from uncertainty
Check meaning, access and output
13. Audit scales, baselines and geometry
14. Check categories, ordering, totals and overlap
15. Make missingness, exclusions and revisions visible
16. Use colour, labels and contrast to support meaning
17. Provide a real text alternative and a usable data table
18. Write captions that separate observation, explanation and action
1. Write the reader’s question before choosing a chart
“Make a chart from this table” leaves too many decisions unstated. A table can support several different questions, and each question can require a different picture. A list of resource requests might answer which resource received the most requests, what proportion each resource contributed, whether requests changed over time or whether requests differed between groups. Those are related questions, but they are not interchangeable. A visual that answers one can distract from another.
Start with a sentence containing the comparison and its boundary. For example: “Which of four resource types received more requests during this hypothetical collection period?” That sentence identifies the categories, quantity and scope. It does not ask which resource improves learning, because request counts contain no learning-outcome evidence. It also does not ask what all students prefer, because the collection method and target population have not been established. The wording keeps the chart’s job smaller than the stories a model could produce about it.
Next, identify what the reader should do after seeing the chart. They might choose which resource to prepare first, notice a missing record or decide what further information to collect. A chart prepared for an operational choice may need exact counts beside bars. A chart prepared to investigate a possible relationship may need every observation and a cautious description. “Understand the data” is a starting intention, but a specific reading task gives you a way to judge whether the finished picture works.
Ask SI for two candidate chart forms and a reason for each. Then require it to state which question each form answers and which question it does not answer. This turns chart selection into an explanation rather than a styling contest. If the proposed chart depends on a missing column, reject the proposal until that information exists. Do not let the model quietly invent dates, group labels or sample sizes to make a more elaborate figure possible.
A useful stopping test is simple: can a reader name the thing being measured, identify the groups or observations and describe the intended comparison after a short look? If not, improve the question, labels or visual form. Adding more colours rarely solves an unclear question.
2. Establish the unit of observation and the data contract
Before plotting, say what one row represents. In one dataset it might be one request; in another, one learner; in another, a weekly total. These are different units of observation. If a person submits three requests, a chart of requests is not automatically a chart of three people. If one class contributes many observations, a collection of observations is not automatically a collection of independent classes. The chart label must preserve these distinctions.
A small data contract records the column name, meaning, unit and permitted values. “Minutes” should mean elapsed minutes or active practice minutes, not whichever interpretation sounds plausible. “Completed” should say what counts as completion and whether a repeated attempt is counted again. “Week” should indicate a consistent interval rather than an arbitrary row number. “Unknown” should have an explicit meaning that is different from zero, not applicable and not collected.
You can ask SI to draft that contract from a data dictionary or an authorised source. Ask it to mark ambiguities rather than resolve them silently. For example, a column named “total” might contain the total assigned, the total who responded or the total who completed. A plausible guess can generate a perfectly consistent but substantively wrong percentage. Resolving that ambiguity is more valuable than producing an immediate chart.
Keep the source values unchanged while making a plotting copy. Record any transformation in ordinary language: “Divide completed by assigned and multiply by one hundred,” or “Sum requests within each resource category.” Distinguish a transformation from a correction. Combining rows under a documented rule is a transformation; replacing a suspicious value is a correction that needs evidence. You should be able to trace every plotted number back to a source value and a named operation.
Finally, separate identifiers from quantities. A case number such as 001 is a label; it is not a measurement that belongs on a continuous numerical axis. A year can be a label in a category comparison or a time coordinate in a trend chart. SI must know which role you intend. The meaning of a column, rather than its file format or appearance, determines how it should be encoded.
Previous chapter · Contents · Next chapter
3. Choose visual forms by the relationship you need to show
For comparison across separate categories, begin with bars or dots on a shared numerical scale. Bars make magnitude visible through length from a common baseline. Horizontal bars are especially helpful when category names are long. Sorting by value can clarify a ranking, although a meaningful natural order should usually remain intact. School levels, age bands and ordered response options should not be rearranged merely to make the silhouette more attractive.
For change through time, consider a line chart with points at the actual observation times. The connecting line communicates sequence, so it is inappropriate for unrelated categories. For paired numerical measurements on individual cases, use a scatter plot: one point represents one pair. For the distribution of a numerical variable, consider a dot plot for a small dataset or a histogram when there are enough observations to make grouping useful. A mean alone cannot reveal how values are spread.
For parts of a single whole, a bar chart, a stacked bar or sometimes a pie chart may work. First establish that the parts are mutually exclusive and together cover the stated whole. Responses to a multiple-selection question usually do not satisfy that condition. If someone can choose both “quizzes” and “flashcards”, those choices cannot be presented as exclusive slices of people without changing the question. A familiar chart type does not repair an incompatible denominator.
Use a table when exact lookup is the main task or the dataset is so small that a chart adds little. Two values may need only a sentence. A chart and a table can complement each other: the chart reveals shape, while the table supports exact checking and access to the values. Do not treat the table as an admission that the graphic has failed. They provide different reading paths through the same evidence.
The Office for National Statistics data visualisation guidance offers a useful primary-source starting point for chart types, presentation and colour. Use such guidance to inform the decision, then test the decision against your actual question. A chart recommendation is still conditional on the data’s structure and the intended audience.
Previous chapter · Contents · Next chapter
4. Give SI a chart brief with boundaries it can check
A strong brief includes the question, audience, dataset, units, intended output and important prohibitions. For the first worked example, you could say: “Use the four supplied resource counts to compare recorded requests. Create a horizontal bar chart with an axis beginning at zero, direct count labels and all four categories. Use one colour. Mark the data as hypothetical. Provide the unchanged table, a short description and one limitation. Do not infer learning effectiveness or create extra observations.”
Then ask for the proposed mapping before the finished chart: category to vertical position, request count to horizontal length, one mark per category. This mapping is the visual equivalent of a formula. It is small enough to inspect and specific enough to test. If a generated script maps the wrong column to bar length, you can catch the error before discussing typography. If the labels say percentages but the values are counts, the brief exposes the mismatch.
Specify what the tool should return as evidence. A chart-generation workflow should preserve the source table, transformation steps, plotting instructions or code, the exported chart and a brief verification record. A screenshot without underlying values cannot support a numerical audit. Code without an actual rendered chart cannot establish that labels fit or that an export preserved the intended scale. Both kinds of evidence matter.
Tell the assistant how to handle incomplete information. “Stop and identify an unknown denominator” is stronger than “be accurate.” “Leave missing weeks unplotted and describe the gap” is stronger than “clean the data.” Do not demand a final-looking answer at the cost of honest uncertainty. If the dataset cannot answer the question, the correct deliverable may be a proposed collection plan rather than a chart.
Finally, request a neutral first version. Remove gradients, decorative icons, perspective and unnecessary animation until the mapping is correct. Once the plain version works, modest styling can improve readability. This order prevents presentation choices from becoming emotional commitments: it is easier to discard a wrong chart before everyone has spent time admiring its polished appearance.
Previous chapter · Contents · Next chapter
5. Worked visualisation: compare four resource counts
Our first hypothetical dataset records requests for four resource types during one collection period. Worked examples received eighteen requests, short quizzes twelve, mind maps nine and flashcards six. The table contains category totals, not individual learners. We do not know whether the same person made several requests. The appropriate question is therefore about recorded requests, and every title and label should retain that unit.
| Resource type | Requests (count) |
|---|---|
| Worked examples | 18 |
| Short quizzes | 12 |
| Mind maps | 9 |
| Flashcards | 6 |

The horizontal bar chart uses category names on the vertical axis and request counts on the horizontal axis. Its numerical scale runs from zero to twenty with equal five-request steps. Each bar begins at zero; its right edge matches the labelled count. The longest bar is eighteen, followed by twelve, nine and six. The bars appear in descending order because the question asks for a comparison rather than a natural sequence. All four use the same colour because colour does not encode an additional variable.
Check the total independently: eighteen plus twelve plus nine plus six equals forty-five requests. Worked examples account for eighteen divided by forty-five, or forty per cent of the requests. Flashcards account for six divided by forty-five, approximately thirteen point three per cent. Neither percentage is a share of learners unless the collection design establishes exactly one request per learner. This is a small wording distinction with a large effect on the claim.
A suitable caption is: “In this hypothetical collection of forty-five requests, worked examples received the most requests: eighteen, compared with twelve for short quizzes, nine for mind maps and six for flashcards. Request counts do not establish which resource teaches most effectively.” The caption gives the main comparison, the values and the boundary. It does not praise the chart or describe a fictional class as though it were real.
The visual audit has three parts. First, count the marks: four source rows should produce four bars. Second, verify the lengths: the eighteen-request bar should be three times the six-request bar because both start at zero. Third, check the labels: “Requests” is correct; “Students”, “Scores” and “Percentage” would be wrong. These checks are easier to perform when the chart is simple.
Now change the condition. Suppose the same source later reveals that the four counts come from collection periods of different lengths. The picture still displays the recorded counts accurately, but it no longer supports a fair comparison of request frequency without additional exposure information. You might need requests per day, with the days shown. Do not keep the old conclusion merely because the old image remains visually persuasive.
Previous chapter · Contents · Next chapter
6. Repair the wrong chart without changing the facts
Imagine that SI renders the four resource counts as a line chart, joining worked examples to short quizzes, then mind maps and flashcards. The numbers are unchanged, but the line suggests progression between categories. A reader may interpret the descending shape as a decline over time. There is no time variable in the dataset. The repair is to use separate bars or points, not to explain the misleading line with a longer footnote.
Another faulty version begins the horizontal bar axis at five. The visible lengths then become thirteen, seven, four and one units. The worked-example bar appears thirteen times as long as the flashcard bar, although the underlying counts are only three times as large. A bar’s length is part of its meaning. Restore the zero baseline and state the actual ratio if that comparison is useful. Merely printing accurate endpoint labels does not neutralise a distorted visual length.
A third version enlarges book icons in both width and height to represent the counts. If an icon’s linear dimensions are tripled, its area becomes nine times as large. A viewer may perceive an exaggerated difference. For a beginner’s comparison, replace the icons with plain bars. If area is deliberately used in a specialised graphic, calculate the geometry correctly and explain it, but do not make that complexity the default for a four-category count chart.
A fourth version reports “forty per cent of students prefer worked examples.” The numerical fraction comes from the request total, but the word “students” is unsupported. Repair the sentence to “forty per cent of recorded requests were for worked examples.” That correction belongs in the title, caption, alt description and any summary derived from the chart. Fixing only the body paragraph leaves other entry points misleading.
These repairs illustrate a useful way to review SI output: inspect the claim, encoding and wording separately. The data may be correct while the encoding is wrong; the encoding may be correct while the caption overreaches. Ask the assistant to identify a concrete failure and propose the smallest repair. Avoid a vague instruction to “make it more honest”, which can produce cosmetic changes without addressing the mechanism of the error.
Previous chapter · Contents · Next chapter
7. Worked visualisation: counts and completion rates answer different questions
Consider two hypothetical learning routes. Route A has twelve completed tasks out of twenty assigned. Route B has eighteen completed tasks out of forty assigned. If you ask which route recorded more completions, the answer is B: eighteen exceeds twelve. If you ask which route had the higher completion proportion, the answer is A. These statements do not contradict each other; they use different quantities.
| Route | Completed | Assigned | Completed / assigned (%) |
|---|---|---|---|
| Route A | 12 | 20 | 60 |
| Route B | 18 | 40 | 45 |

The plotted measure is completed divided by assigned, multiplied by one hundred. Route A is twelve divided by twenty, giving sixty per cent. Route B is eighteen divided by forty, giving forty-five per cent. The horizontal axis runs from zero to one hundred per cent, and the direct labels include both the percentage and its fraction. This allows the reader to see the comparison while retaining the denominator that generated it.
The difference is fifteen percentage points: sixty minus forty-five. If someone instead asks for Route A’s relative rate difference compared with Route B, the calculation is fifteen divided by forty-five, or about thirty-three point three per cent. “Fifteen per cent higher” is ambiguous and does not correctly describe that relative comparison. Prefer “fifteen percentage points higher” when you mean subtraction between percentages. State the reference group whenever using a relative percentage change.
The combined completion proportion is thirty completed tasks out of sixty assigned, or fifty per cent. It is not the unweighted average of sixty and forty-five, which is fifty-two point five per cent. The two routes have different denominators. To summarise all assignments, add the numerators and denominators before dividing. An average of group percentages answers a different question in which each group receives equal weight, regardless of its size.
The caption should also preserve the study boundary: “In this hypothetical dataset, Route A has a higher recorded completion proportion, sixty per cent versus forty-five per cent. The groups have different assignment totals, and no random allocation or comparable task difficulty has been established.” The chart does not prove that A causes more completion. A different group composition, deadline or task difficulty could contribute to the observed difference.
Change the condition again. Suppose five Route B assignments were cancelled before work began. Whether they belong in the denominator depends on the defined completion measure. If your established rule counts only eligible assignments, B becomes eighteen out of thirty-five, approximately fifty-one point four per cent. Document that rule before recalculating and apply it consistently to both routes. Do not remove assignments merely because the resulting rate looks better.
Previous chapter · Contents · Next chapter
8. Worked visualisation: show a time series with a genuine gap
The next hypothetical table records books borrowed during six consecutive weeks. The recorded counts are twenty, twenty-four, no available record, thirty, twenty-eight and thirty-six. “No available record” is not a count of zero. It is a statement about the evidence. A chart that turns the missing week into zero would depict a collapse that the dataset does not establish.
| Week | Books borrowed (count) |
|---|---|
| 1 | 20 |
| 2 | 24 |
| 3 | Not recorded |
| 4 | 30 |
| 5 | 28 |
| 6 | 36 |

The chart keeps all six week positions on the horizontal axis and uses a zero-to-forty vertical scale in equal ten-book steps. It plots five observed points. A line joins weeks one and two, and another connects weeks four, five and six. The line is broken across week three, with a visible “No record” annotation. No point is placed at week three, and no interpolated value is presented as an observation.
A careful reading is: the recorded count increases from twenty in week one to twenty-four in week two; after the missing record, it is thirty in week four, falls to twenty-eight in week five and reaches thirty-six in week six. The first and last observed counts differ by sixteen books. Relative to the first count, that is an eighty per cent increase. This endpoint comparison does not mean every intervening week increased, and the missing week prevents a complete description of the path.
The sum of observed counts is one hundred and thirty-eight. Dividing by the five observed weeks gives twenty-seven point six books per observed week. Do not call that the six-week average, because the sixth contribution is unknown. Dividing by six after silently replacing the missing value with zero would give twenty-three, a different and unjustified result. A title such as “Observed weekly borrowing” helps keep the scope visible.
For implementation, the Matplotlib example on missing and masked values shows how missing values can leave breaks in a plotted line. The software mechanism is only one part of the job. You must also inspect the saved image and ensure a later spreadsheet import or export did not connect the gap automatically.
If the missing week is later recovered as twenty-six, add that verified observation and regenerate the chart. The six-week total becomes one hundred and sixty-four, and its mean becomes approximately twenty-seven point three three. Preserve a note that the earlier version lacked week three. A corrected chart should not leave its previous caption, missing-data annotation or summary figures unchanged.
Previous chapter · Contents · Next chapter
9. Time spacing, accumulation and smoothing change the story
Consecutive weeks can be equally spaced. Irregular observation times require more care. Suppose readings occur on days one, two and ten. A chart that spaces those labels equally makes the one-day gap look the same as the eight-day gap. For a time-based interpretation, positions should reflect elapsed time. If equal category spacing is intentionally used instead, make that choice clear and avoid interpreting the slopes as rates of change through time.
Ask whether each value describes a period or a running total. Weekly borrowing counts can rise or fall. A cumulative total generally cannot fall unless the definition allows reversals or corrections. Plotting cumulative counts often creates a smooth upward line even when current activity is slowing. If the reader wants to know whether borrowing has accelerated, a cumulative chart alone may obscure the answer. Show the period counts or an explicitly calculated change instead.
Smoothing introduces another transformation. A moving average can make a noisy pattern easier to inspect, but it also combines observations and may conceal a short event. State the window, its alignment and how missing values are handled. A three-period trailing average uses the current and preceding two periods; a centred average has a different timing interpretation. An assistant should not add a smooth curve because it looks professional without explaining what operation created it.
In our six-week example, fitting an elaborate curve would be a poor default. There are only five observed points and one missing record. The simple line segments communicate the available sequence with fewer assumptions. A forecast beyond week six would require a separate modelling task, assumptions and uncertainty communication. Extending the same solid line into the future would blur the boundary between observed data and a prediction.
When a chart includes both observations and a forecast, label the boundary directly and use additional visual distinctions such as different line styles. Explain whether any shaded band describes prediction uncertainty, a scenario range or something else. The reader should not have to infer status from a subtle colour change. The question to ask SI is: “Which marks are observations, which are transformations and which are assumptions?” Every visible element should have a defensible answer.
Previous chapter · Contents · Next chapter
10. Worked visualisation: inspect paired numerical observations
A scatter plot answers a different kind of question. Here, eight hypothetical cases each have a recorded practice duration and a quiz score out of ten. The cases are numbered only to let you match points to table rows. These numbers do not represent real learners. The question is whether longer recorded practice time tends to accompany higher scores within this small invented collection.
| Case label | Practice (minutes) | Quiz score (out of 10) |
|---|---|---|
| 1 | 10 | 4 |
| 2 | 15 | 6 |
| 3 | 20 | 5 |
| 4 | 25 | 8 |
| 5 | 30 | 7 |
| 6 | 35 | 9 |
| 7 | 40 | 8 |
| 8 | 45 | 9 |

Each point maps its case’s minutes to horizontal position and score to vertical position. The horizontal scale runs from zero to fifty minutes; the vertical scale runs from zero to ten marks. Case one is at ten minutes and four marks. Case six is at thirty-five minutes and nine marks. Case eight is at forty-five minutes and nine marks. All eight table rows appear, with no joining line between cases.
The pattern generally rises from lower left to upper right, but it is not strictly increasing. Case three has a lower score than case two despite more recorded minutes; similar local reversals appear later. The chart therefore supports a description of an overall positive association in these hypothetical observations, not a rule that every additional minute produces a higher score. It also does not establish that practice duration caused the difference.
Why avoid a causal caption? The dataset does not state how cases were selected, whether prior knowledge differed, whether the same quiz was used or whether practice was randomly assigned. A case that is already confident might choose a different duration from a case finding the task difficult. Those possibilities do not prove a particular alternative explanation; they show why the available pairs are insufficient to identify a cause.
The NIST explanation of scatter plots describes their role in examining relationships between two variables and distinguishes association from proof of causation. For this example, the important learning task is to preserve the paired observations and inspect the pattern before adding a trend line or a numerical summary. A regression line can look authoritative even when the underlying sample is too limited for the claim being made.
Try changing the data condition: add a ninth case with forty-five minutes and three marks. Keep it visible while checking whether it is valid. Do not delete it just to restore a stronger rising pattern. If it is genuine within the hypothetical exercise, the visual summary must acknowledge it. If an error is confirmed by the source, correct the record and preserve the correction history. An unusual point is a question to investigate, not permission to erase evidence.
Previous chapter · Contents · Next chapter
11. Worked visualisation: reveal the distribution behind an average
A mean can be correct and still be an incomplete picture. Consider ten hypothetical task-completion times in minutes: five, six, six, seven, seven, eight, nine, twelve, twenty and twenty. The sum is one hundred, so the mean is ten minutes. Yet none of the recorded completion times is exactly ten minutes. A single bar at the mean would hide both the concentration of shorter tasks and the two longer completions.
| Task label | Completion time (minutes) |
|---|---|
| 1 | 5 |
| 2 | 6 |
| 3 | 6 |
| 4 | 7 |
| 5 | 7 |
| 6 | 8 |
| 7 | 9 |
| 8 | 12 |
| 9 | 20 |
| 10 | 20 |

The dot plot places one dot at each recorded duration. Repeated values stack vertically so that two six-minute observations, two seven-minute observations and two twenty-minute observations remain visible. Horizontal position carries the time value; vertical stacking only prevents identical values from covering each other. The vertical position is not a second measurement. The chart marks the mean with a dashed line at ten and the median with a dotted line at seven point five.
To calculate the median, sort the ten values and average the fifth and sixth. Those values are seven and eight, giving seven point five. The median lies below the mean because the longer observations pull the mean upward in this dataset. Describe the actual distribution rather than treating one summary as universally superior. If you are estimating total staff time, the mean may have a relevant operational use; if you are describing a typical middle completion, the median may be useful. The question determines the summary.
A histogram would group values into intervals. For example, equal-width intervals from zero up to five, five up to ten, ten up to fifteen, fifteen up to twenty, and twenty up to twenty-five contain zero, seven, one, zero and two observations respectively when each interval includes its lower boundary and excludes its upper boundary. That boundary rule matters: a twenty-minute value belongs in the final interval. Do not let an assistant double-count boundary observations.
The NIST histogram guidance explains how histograms display numerical distributions. For only ten observations, the dot plot retains useful detail without a binning decision. If you later use many more observations, grouping may become practical, but compare reasonable bin choices and keep the units explicit. A distribution should not appear to change merely because the chart software silently selected new boundaries.
Now suppose the two twenty-minute records were changed to ten after verified source corrections. The new total would be eighty and the mean eight; the median would remain seven point five. This changed-condition example shows why a summary can react differently from another summary. It also gives you a concrete check on regenerated labels: the old mean marker at ten would become wrong even if the remaining dot positions were updated correctly.
Previous chapter · Contents · Next chapter
12. Worked visualisation: distinguish an observed range from uncertainty
Two hypothetical methods each have four recorded batch times. Method A takes eight, ten, twelve and ten minutes. Method B takes six, ten, fourteen and ten minutes. Both sums are forty, so both means are ten. Their observed ranges differ: A spans eight to twelve, while B spans six to fourteen. This small example lets us draw an interval without pretending that every interval is a confidence interval.
| Method | Batch label | Time (minutes) |
|---|---|---|
| A | 1 | 8 |
| A | 2 | 10 |
| A | 3 | 12 |
| A | 4 | 10 |
| B | 1 | 6 |
| B | 2 | 10 |
| B | 3 | 14 |
| B | 4 | 10 |

The chart places each method on a separate row. A horizontal line runs from the observed minimum to the observed maximum, with end caps showing the boundaries. A diamond marks the mean. Both diamonds sit at ten minutes, but B’s line extends farther in both directions. The caption explicitly says that each method has four hypothetical observations and that the lines are observed minimum-to-maximum ranges.
This picture supports a limited statement: the two recorded means are equal, and the observed times for B span a wider interval in these data. It does not establish that B will always vary more, that either method is faster in a larger population or that a future task must fall within the plotted range. A sample’s smallest and largest values do not create a guaranteed bound on future observations. Avoid language such as “ninety-five per cent certain” when no corresponding calculation was performed.
Different interval types answer different questions. A range summarises the recorded minimum and maximum. A standard deviation summarises spread around a mean under its definition. A confidence interval concerns an estimation procedure and its assumptions. A prediction interval concerns future observations under a model. These terms should not be swapped because the software draws similar-looking whiskers. Require SI to identify the interval type, its calculation, sample size and assumptions before it appears in a chart.
If you do not have enough information to calculate a defensible uncertainty interval, say so. Do not ask a model to invent “reasonable error bars”. A transparent table with a note about limited evidence is more useful than a decorative band of fabricated certainty. In higher-stakes work, seek appropriate statistical review rather than relying on a charting assistant to choose an inferential method from a short prompt.
Change the condition by adding a fifth observed batch of twenty minutes to Method A. Its new mean becomes twelve, and its observed range becomes eight to twenty. Both the diamond and the upper endpoint must move. If only the average label changes, the chart is internally inconsistent. This example is a practical regression test for the connection between raw rows, derived summaries and visual marks.
Previous chapter · Contents · Next chapter
13. Audit scales, baselines and geometry
A numerical axis is a measuring instrument inside the picture. Equal steps on a linear scale must represent equal numerical differences. Read the tick labels in order and check the spacing. If zero, ten, twenty and one hundred are placed at equal distances on an axis presented as linear, the geometry misrepresents the values. Do not assume the charting library’s appearance proves that the underlying mapping was configured correctly.
Bars ordinarily need a zero baseline because their visible lengths represent magnitudes. A line chart or dot plot can sometimes use a narrower displayed range to reveal a small difference, provided the scale is clearly labelled and the presentation does not encourage a false magnitude interpretation. The decision is not “all axes must start at zero” versus “anything goes”. Ask which perceptual feature carries the comparison: length from a baseline, position along a scale or change over time.
Consider a temperature series that varies from twenty-four to twenty-six degrees. A tightly bounded line chart could make the variation visible, but a dramatic title such as “temperatures explode” would still mislead. Conversely, an extremely wide vertical range can conceal a meaningful change. Check both the factual scale and the rhetorical framing. Use a reasonable aspect ratio so that the same data do not look artificially flat or precipitous merely because the chart is unusually wide or tall.
Logarithmic scales need an explicit reason and clear labels. Equal distances on a logarithmic scale represent equal multiplicative changes, not equal additive changes. Zero and negative values require special treatment and cannot simply be placed on an ordinary logarithmic axis. For a beginner’s classroom comparison, a linear scale is usually easier to explain. Do not adopt a log scale solely because it makes an awkward dataset fit neatly.
Avoid three-dimensional bars for ordinary numerical comparison. Perspective, occlusion and depth can make front objects seem more important or hide the actual baseline. Likewise, check whether a bubble’s value is mapped to area rather than radius if size is genuinely needed. The safest default is a two-dimensional chart with a shared positional scale. Every additional visual dimension creates another mapping that someone must verify.
Previous chapter · Contents · Next chapter
14. Check categories, ordering, totals and overlap
Category cleaning affects the picture before any mark is drawn. “Short quiz”, “Short quizzes” and “short quizzes ” might be the same category or different labels with a meaningful distinction. Do not merge them automatically without a rule. Ask SI to list proposed merges, explain the evidence and preserve the original labels in the source copy. A polished chart that hides an undocumented merge is difficult to audit later.
A category called “Other” needs a definition. It may combine many small categories, unknown responses or everything below a display threshold. Those are different choices. If you group small categories to simplify a chart, state the grouping rule and retain their details in the table. If the reader’s decision concerns one of those small categories, aggregation may remove precisely the information they need. Simplicity should serve the question, not erase inconvenient cases.
Sorting can alter interpretation. A descending order is helpful for ranking the four resource requests. It would be confusing for a sequence such as “not started”, “in progress” and “completed”, where the order conveys a process. Use the same category order across comparable panels unless there is a clear reason to change it. When colours encode categories, keep the mapping stable too. A blue bar should not mean one route in the first panel and another route in the next.
For part-to-whole charts, independently check that the categories are mutually exclusive, the total is defined and the values add to it. Rounding can make displayed percentages sum to ninety-nine point nine or one hundred point one; explain that small discrepancy rather than changing source values to force a neat total. A much larger discrepancy may indicate overlapping categories, missing records or an incorrect denominator and should be investigated.
Suppose forty-five requests include ten requests marked both “worked example” and “quiz”. If the category counts include both tags, their sum represents category selections, not unique requests. The original total of requests can no longer be inferred by simply adding categories. A bar chart may still display tagged selections, but a pie chart labelled “share of requests” could imply a partition that does not exist. This is another case where changing a label and chart form is more honest than forcing the data into a familiar template.
Previous chapter · Contents · Next chapter
15. Make missingness, exclusions and revisions visible
Missing data has a mechanism even when you do not know it. A blank value may arise because an item was not measured, a respondent declined, a file failed, a question did not apply or a record was lost. Those causes can have different implications for interpretation. If non-completion makes a response less likely to be recorded, the observed completion data may tell a different story from the full set. Do not describe a dataset as representative merely because it has many rows.
Keep missing, zero and not applicable separate in the table. A zero means the measured quantity is zero under the defined process. Missing means the value is unavailable. Not applicable means the measurement does not belong to that case under the rule. A chart may render all three as an absent mark unless you configure it carefully, so labels and explanatory text must preserve the distinction that the image alone cannot carry.
Every exclusion should have a rule that is relevant to the question. Removing duplicate records with a verified identifier is different from removing inconvenient values because they weaken a pattern. Write the rule before inspecting the preferred conclusion whenever possible. Record how many rows were present, how many were excluded and why. When the chart uses a filtered dataset, the title or caption should make the filter clear enough that a reader does not mistake it for the full collection.
Do not silently interpolate values just to make the line continuous. Interpolation is a model-based transformation, even when the method is simple. If it is justified for a particular use, distinguish interpolated values from recorded observations and retain the missingness indicator. In the weekly borrowing example, a straight segment across week three would be a visual assumption about an unrecorded period. Leaving a gap is a more direct representation of what is known.
When new data arrive, revise the chart, table, caption, alt description and any summary together. An updated image paired with an old paragraph can be more confusing than an openly incomplete earlier version. Keep a date or version note when readers may compare releases. SI can help locate dependent statements, but verify that every factual reference changed with the source rather than assuming a global rewrite caught them all.
Previous chapter · Contents · Next chapter
16. Use colour, labels and contrast to support meaning
Colour should have a job. It may distinguish categories, encode a numerical scale or draw attention to a justified comparison. In the resource-count chart, one colour is sufficient because category labels and position already identify each bar. Introducing four bright colours would add visual complexity without adding information. In a multi-series chart, colour can help, but it should be supported by direct labels, line styles or marker shapes.
The W3C guidance on use of colour explains why colour alone should not carry essential information. For example, “follow the green line” is a weak instruction when another reader may not distinguish it from the red line. “Route A, shown with circular markers and a solid line” provides additional cues. A legend that uses both shape and text is more robust than a row of unexplained coloured squares.
Contrast matters separately from hue. The W3C non-text contrast guidance addresses the visibility of meaningful graphical objects, while its text contrast guidance addresses readable text. Do not use faint pastel marks or tiny grey labels just because they look restrained on a large designer’s screen. Important marks and labels need to remain legible in the actual publication context.
Direct labels can reduce repeated movement between marks and a legend. A bar-end value is useful when exact comparison matters, and an end-of-line series label can be clearer than a distant colour key. Avoid labelling every point if the result becomes a wall of overlapping text. Choose a smaller number of purposeful annotations and provide all values in the accessible table. The text alternative should carry the detail that is not practical to fit inside the picture.
Test the chart in greyscale, at a narrow width and at an enlarged browser zoom. These checks do not establish complete accessibility, but they expose many preventable failures. Ask whether a reader can still identify the series, see the marks and understand the scale. If the answer depends on hovering over a tiny point, add a visible alternative. The most important conclusion should not disappear when colour, pointer precision or screen space is limited.
Previous chapter · Contents · Next chapter
17. Provide a real text alternative and a usable data table
A chart image needs more than alt text saying “a chart”. The short description should identify what is shown and direct readers to nearby detail when appropriate. For the weekly example, a useful short alternative is: “Weekly borrowing chart with a missing record at week three; the nearby table and explanation give all recorded counts.” It tells the reader the subject and the key structural feature without cramming every value into a long attribute.
The W3C tutorial on complex images recommends a short identification together with a fuller text equivalent for information-rich graphics. In this guide, each chart is accompanied by its source table and prose explaining its mapping and main relationships. That combination lets someone inspect the same evidence without depending on visual perception alone. The longer explanation should describe relevant gaps, scales and qualifications as well as the headline pattern.
A usable table has meaningful headers and explicit units. “Time (minutes)” is better than “Value”. “Completed” and “Assigned” are better than unexplained columns A and B. Keep the table small enough to navigate where possible, and do not merge unrelated quantities into a single cell merely to save horizontal space. A missing value should be written as “Not recorded” when that is its meaning, rather than shown as an unexplained dash.
The table should match the actual plotted data. If the chart shows percentages derived from counts, include the numerator and denominator. If it shows a filtered subset, identify that subset and give the relevant rows. If privacy requires aggregation, provide a safe aggregate table rather than exposing hidden individual records in a downloadable file. Accessibility does not remove the responsibility to protect confidential information.
Write the text alternative after inspecting the rendered chart, then compare it with the source. A model may describe a line, legend or trend that the export did not actually contain. It may also overstate an intended message that the data do not support. The alternative is part of the publication, so it deserves the same factual review as the chart title and main paragraph. Do not allow a careful graphic to acquire an exaggerated story through its accessibility text.
Previous chapter · Contents · Next chapter
18. Write captions that separate observation, explanation and action
A chart caption has three potential layers. The observation says what the values show. An explanation suggests why the pattern may have occurred. An action states what someone should do. These layers require different evidence. “Route A has a recorded completion proportion of sixty per cent” is an observation. “Route A is easier” is a proposed explanation. “Replace Route B” is a decision that needs further context about goals, costs and comparability.
Keep the first sentence close to the marks. Name the quantity, population or collection, period and main comparison when those are known. Add the limitation that most directly prevents a likely misreading. For the scatter example, the key limitation is that the paired observations do not identify a causal effect. For the missing-week series, it is the unrecorded period. A long list of generic disclaimers is less useful than the specific boundary a reader is likely to cross.
Avoid verbs that imply more than the measurement. “Proves”, “guarantees” and “demonstrates effectiveness” need much stronger support than a descriptive chart of convenience data. Even “improves” can imply a causal comparison when the dataset contains only different groups at one time. Safer wording is not automatically vague wording. “The recorded proportion is fifteen percentage points higher” is precise and appropriately limited.
Use SI to propose alternative captions, then make it show the value or source supporting each factual clause. Delete unsupported clauses instead of decorating them with “may” until they sound cautious. “This may prove that the method is superior” is still conceptually confused. If you want to offer an explanation, label it as a hypothesis and say what additional evidence would help test it.
A decision-oriented ending can still be practical. For the resource-request example: “Prepare worked examples first if the immediate goal is to address the largest number of recorded requests; collect outcome evidence separately before deciding which resource best supports learning.” This recommendation is conditional on a stated goal and does not manufacture an effectiveness finding. The chart helps the decision, while the caption keeps the decision from outrunning the evidence.
Previous chapter · Contents · Next chapter
19. Build and verify in a spreadsheet or plotting tool
The workflow is similar whether you use a spreadsheet, a coding environment or a chart editor. Begin with a protected source copy. Create a plotting table with the exact rows and derived columns needed. Select the chart type and explicitly inspect the data range, series names, category labels and axis settings. Do not assume the automatic suggestion used the intended columns or excluded summary rows. A “Total” row can accidentally become an extra bar or be counted twice.
For a spreadsheet, check whether a numerical-looking label has been treated as a category or quantity. Inspect how blanks are displayed, whether hidden rows are included and whether percentages are stored as fractions or percentage-point values. A stored fraction of zero point six formatted as a percentage represents sixty per cent; a charting interface expecting sixty may use a different convention. The correct representation depends on the tool, so confirm it with a known test value.
For code, make the values and transformations inspectable. Use named columns, avoid unexplained positional indexing and include assertions for key totals, row counts and axis bounds. Render the actual file rather than trusting that the plotting commands will produce the intended appearance. Check that the exported image is not an earlier file with the same name and that a cached preview has not hidden the latest change. The evidence of success is the current rendered output matched to the current data.
Reproduce one plotted value manually. For the completion chart, calculate twelve divided by twenty independently and confirm that the corresponding bar reaches sixty on a zero-to-one-hundred scale. For the distribution chart, count ten dots and verify that repeated values remain visible. For the range chart, inspect the endpoints and the mean marker separately. These small checks often catch wrong-column or stale-summary errors faster than rereading a long generated explanation.
Keep a compact audit record: source version, transformations, chart type, plotted row count, scale limits, missing-value handling and reviewer checks. It does not need to become a bureaucratic report. Its purpose is to let another person reproduce the chart and understand why its choices were made. If the assistant cannot provide those details, the workflow is not yet ready for publication, however polished the image looks.
Previous chapter · Contents · Next chapter
20. Review the exported page, not only the chart editor
A chart can be correct in its editor and fail in the final destination. Export may shrink labels, clip a title, remove transparency, change a font or rasterise text at an unreadable size. A webpage may display it in a narrow column. A slide may crop its axis. A document conversion may separate the caption from the figure. Therefore, inspect the actual output where the intended reader will encounter it.
Check the outer edges first. Are the longest category names visible? Are axis titles and units complete? Does the source note survive? Are negative signs, decimal points and percentage symbols legible? Then inspect the plotting area: all marks should remain inside the intended scale, annotations should not cover important points and repeated observations should not disappear underneath one another. A file that opens successfully has not necessarily passed these checks.
On a webpage, test a narrow viewport and ordinary browser zoom. The chart should fit its container or have an understandable way to access a larger version. The accompanying table should remain usable, and the main interpretation should still be available as text. Do not rely on a reader downloading an image or using a hover tooltip to discover a critical missing-data note. A static fallback is valuable even when an interactive chart is offered.
Check consistency across the whole publication. The chart title, surrounding paragraph, alt description, table, thumbnail and social preview may be generated at different stages. An early exaggerated title can survive after the main caption has been corrected. Search for old values and claims when making a revision. If a displayed summary says “fifty-two point five per cent overall” while the table totals thirty out of sixty, correct the summary before publication.
Finally, have someone who did not create the chart describe what they think it says. Ask a concrete question: “Which route has more completions, and which has the higher completion proportion?” If they confuse those answers, improve the presentation rather than treating the confusion as their failure. Reader testing exposes ambiguities that a numerical audit cannot. The goal is a faithful interpretation in another person’s mind, not merely a technically valid file.
Previous chapter · Contents · Next chapter
21. Practice set: choose a chart and repair a denominator
Try the following tasks before reading the answers. In Exercise A, a hypothetical club records requests for four activities: discussion twelve, drawing nine, quizzes fifteen and model-building six. The question is which activity received more recorded requests. Choose a chart, an order and an axis. Write one accurate caption and identify one claim the data cannot support. Do not ask SI for the answer until you have made a choice yourself.
In Exercise B, Team Cedar completes nine of fifteen assigned tasks and Team Maple completes sixteen of forty. Calculate each completion proportion, the difference in percentage points and the combined proportion. A generated caption says, “Maple is better because sixteen is larger than nine.” Explain what question that sentence answers and why it does not answer the completion-proportion question. Then write a corrected caption with the denominators visible.
Answer A: a horizontal bar chart is a reasonable choice because the categories are separate and the question concerns counts. A descending order would be quizzes fifteen, discussion twelve, drawing nine and model-building six. Use a zero baseline, equal numerical steps and a scale extending far enough to include fifteen without clipping. A zero-to-twenty scale with five-request ticks would work. Other readable scales are possible; the important conditions are honest geometry and clear labels.
A suitable caption is: “In this hypothetical collection, quizzes received fifteen requests, discussion twelve, drawing nine and model-building six.” The total is forty-two requests. The dataset does not establish which activity produces better learning or whether the counts equal unique people. A pie chart could show parts of the request total only if the categories form a complete, non-overlapping partition, but it is less direct for comparing nearby category counts. A line chart would imply an unsupported sequence.
Answer B: Cedar’s proportion is nine divided by fifteen, or sixty per cent. Maple’s is sixteen divided by forty, or forty per cent. Cedar’s recorded proportion is twenty percentage points higher. The combined proportion is twenty-five divided by fifty-five, approximately forty-five point four five per cent, not fifty per cent. The rejected caption identifies which team has the larger number of completions but does not compare the share of assigned tasks completed.
A corrected caption is: “Cedar completed nine of fifteen tasks, sixty per cent, while Maple completed sixteen of forty, forty per cent. Maple recorded more completed tasks; Cedar had the higher completion proportion. These descriptive figures alone do not establish why the rates differ.” If SI supplies a different answer, require it to expose its numerator and denominator before accepting the result. The arithmetic should be checkable without trust in its tone.
Previous chapter · Contents · Next chapter
22. Practice set: missing records and changed conditions
Exercise C gives four consecutive hypothetical daily counts: eight, not recorded, twelve and sixteen. An assistant replaces the missing value with zero, draws a continuous line through it and reports a four-day mean of nine. Identify all three problems. State the sum and mean that can be calculated from the observed values, and describe how you would display the missing day. Then consider a verified correction: the missing count is later found to be ten. Update the summary.
Answer C: the zero is invented, the line depicts an unsupported collapse and recovery, and the reported mean treats an unknown observation as zero. The observed sum is thirty-six. There are three observed days, so the mean over observed days is twelve. The full four-day mean is unknown until the missing value is recovered. Keep all four day positions, leave the missing value unplotted and explain the gap in the caption and table.
Once the missing count is verified as ten, the full sum is forty-six and the four-day mean is eleven point five. The line may now connect the four observed values in chronological order, provided the daily intervals are consistent. The chart, table, missing-value annotation and caption all need revision. An assistant that changes only the number in the summary has not completed the repair.
Exercise D changes the denominator condition. A hypothetical workshop records eight completions among ten eligible assignments in Group P, and twelve among twenty eligible assignments in Group Q. A later file contains five additional Q records that were invitations, not eligible assignments. Should the completion chart automatically use twenty-five as Q’s denominator? Explain the rule you need before changing anything. Calculate the original proportions and their difference.
Answer D: the original proportions are eighty per cent for P and sixty per cent for Q, a twenty-percentage-point difference. The five invitation records should not automatically enter a denominator defined as eligible assignments. First verify whether the eligibility rule changed or whether the new records are outside the measure. If all five are genuinely newly eligible assignments with no additional completions, Q becomes twelve out of twenty-five, or forty-eight per cent. If they remain invitations only, the original denominator stays twenty.
The lesson is that a refreshed file is not necessarily a refreshed measurement of the same concept. SI should not treat every extra row as an additional denominator contribution. Describe the decision rule, apply it consistently and preserve the original definition. A chart that updates automatically without a semantic check can become wrong while every formula continues to calculate successfully.
Previous chapter · Contents · Next chapter
23. Practice set: association, distribution and interval language
Exercise E uses five hypothetical paired observations: minutes and marks are ten and two, twenty and five, thirty and four, forty and eight, fifty and seven. Select a chart and write a one-sentence interpretation. A suggested title says, “Every extra ten minutes causes a higher score.” Identify two reasons to reject it. Then add a sixth observation at fifty minutes and one mark and describe what should happen to the chart.
Answer E: use a scatter plot with minutes on one axis and marks on the other, preserving each pair. A suitable interpretation is: “These five hypothetical pairs show an overall tendency toward higher marks at longer recorded durations, with exceptions.” The proposed title is wrong because the sequence is not strictly increasing and because observational pairs do not establish a causal effect. Adding the sixth case requires an additional point at fifty and one, not deletion or concealment of the case. The description must acknowledge the changed pattern.
Exercise F gives seven hypothetical completion times: four, five, five, six, seven, eight and twenty-one minutes. Calculate the mean and median, then choose a visual that shows more than one summary. Explain why a chart containing only a single bar labelled with the mean could be incomplete. The sum is provided as a check only after you have calculated it: fifty-six minutes.
Answer F: the mean is fifty-six divided by seven, or eight minutes. The median is the fourth sorted value, six minutes. A dot plot can show all seven observations, including the longer twenty-one-minute completion. A single mean bar conceals the concentration of shorter values and the long upper tail in this small dataset. Do not automatically label the twenty-one-minute value an error. Check its source; if it belongs to the defined collection, it remains part of the evidence.
Exercise G shows two methods with observed ranges of five to nine and three to eleven minutes. Both have a mean of seven, and each has four recorded observations. A generated caption calls the ranges “ninety-five per cent confidence intervals”. What is missing? Answer: the information provided identifies observed minimum-to-maximum ranges, not a confidence-interval procedure. The caption needs correction unless a valid estimation method, its required inputs and assumptions have separately been established.
A corrected statement is: “Both hypothetical methods have a recorded mean of seven minutes; the second has a wider observed range in these four observations per method.” It should not claim that the second method is necessarily less reliable in all future use. The exercise is about naming what the graphical interval actually represents. A familiar whisker shape cannot supply statistical meaning that was never calculated.
Previous chapter · Contents · Next chapter
24. Run a complete chart review without SI doing the judgment for you
Use a two-pass review. In the first pass, ignore visual polish and inspect meaning. State the question in your own words. Identify the unit of observation, numerator, denominator, units, period, exclusions and missing values. Recalculate one important number and compare the plotted row count with the table. Read the caption as a set of claims and ask which data support each one. If any claim lacks support, revise it before continuing.
In the second pass, inspect perception and access. Read both axes, check their bounds and intervals, compare the marks with known values and test whether colour is carrying essential information alone. Confirm that the text alternative explains the chart’s actual relationships and that the source table is readable. Inspect the final export at the intended size, including a narrow display where relevant. Correct overlapping labels, clipped units and ambiguity between observations and assumptions.
Then perform one changed-condition test. Alter a copy of a value, denominator or missingness flag in a controlled exercise and predict which parts of the publication should change. For the completion chart, changing Route B’s eligible assignments should affect its percentage, bar length and comparative caption. For the time series, recovering the missing week should affect the line, total, average and missing-data note. If one part fails to update, the workflow has exposed a stale dependency.
Use SI as a critic after making your own judgment. Ask it to look specifically for unsupported causal language, inconsistent denominators, hidden exclusions and mismatches between the image and table. Its criticism can reveal useful issues, but an absence of criticism is not proof of correctness. You remain responsible for checking the actual values and rendering. Another assistant’s confident agreement is not independent empirical evidence.
A learner is ready to work more independently when they can explain why the chart form fits the question, reproduce a key calculation, identify a limitation and repair a changed-condition example without copying a model’s wording. The finished chart is one product; the ability to defend its meaning is the more durable skill. Practise with small datasets where every mark is inspectable before moving to dashboards with hundreds of interacting choices.
Previous chapter · Contents · Next chapter
25. A final publishing checklist and a sensible next step
Before sharing, confirm that the chart has a precise question, a faithful source table, correct transformations, appropriate geometry and a restrained interpretation. Check that every value is observed, calculated under a stated rule or explicitly hypothetical. Verify that missing records remain visible as missing, denominators are named, colour has redundant cues and the accessible explanation matches the rendered figure. Confirm permission to publish the underlying data and remove private details that are not needed for the reader’s task.
If you work with real classroom or workplace data, use authorised, appropriately aggregated information and follow the relevant organisation’s rules. Do not paste identifiable learner records into an AI tool merely because a chart would be convenient. Small groups can sometimes be identifiable even after names are removed. When uncertain, use an approved safe dataset or an invented teaching example like those in this guide until the data-sharing question has been resolved.
Keep the source and chart together so that a later correction is possible. Record which version was published and preserve enough instructions to reproduce it. Recheck a chart when its source, denominator, collection method or intended use changes. A graphic that was accurate for an internal exploratory question may need a different caption or aggregation before it is used in a public report. Reuse should include renewed judgment, not just copying an image.
For your next practice session, choose one of the six worked examples and rebuild it from the table without looking at the chart. Explain your visual mapping before rendering it. Compare the result with the published example, find one difference and decide whether that difference changes meaning or only appearance. Then complete a changed-condition exercise and update every dependent statement. This makes chart creation a repeatable reasoning skill rather than a one-off generation trick.
The central principle is straightforward: use Super Intelligence to accelerate the work around a chart, while keeping the chart’s meaning open to inspection. A trustworthy visual is one whose numbers, marks, words and limitations agree. Return to the Super Intelligence Learning Hub when you are ready to connect this skill with research, mathematics, spreadsheets and communication.
