Whether a length-encoded mark begins at zero

Whether the finding survives its honest rendering, and whether there is a finding at all

Sections 12.6, 12.7

Whether the grain of one mark was stated and confirmed

What grain the question requires

Sections 3.3, 12.10

Whether a reference is on the chart and labeled

Which reference the reader needs, and whether it was declared before the results were inspected

Sections 2.7, 12.9

Whether an assertion title is verifiable from the marks beneath it

Which of the defensible sentences to write, and therefore what the reader takes away

Sections 13.6, 13.15

Whether the removals were listed

Whether each removal is a simplification or the deletion of a qualification

Section 13.15

Whether uncertainty is present and in decision units

How much uncertainty the decision can carry, and what would reverse the recommendation

Sections 2.7, 13.9

Whether a reader’s takeaway was measured

What to do when the reader’s takeaway is not the intended one

Sections 13.2, 13.10

Whether every AI-touched sentence was verified against a view

Whether to keep the sentence

Sections 12.12, 13.11

Pass

The check was run and the artifact satisfies it.

Nothing further. Record it as run; a check recorded as passed without being run is the failure this appendix cannot detect.

Caution

The check was run, the artifact does not satisfy it, and the analyst has decided to accept the defect. Available on Level 2 checks only.

One sentence naming the defect and the reason for accepting it, in the sign-off record. An accepted defect is legitimate; an unrecorded one is not.

Stop

The check was run and the artifact fails it. Mandatory on any failed Level 1 check.

Repair before circulation, a repair note per Section E.6, and a re-run of the checks the repair touched.

Python chart inside a working notebook (Chapters 4–11)

E.5.1

E.1 (1.1, 1.3), E.2, E.3, E.4, E.5, E.6, E.7

E.13 if AI-assisted; E.14 if exported or submitted

Under two minutes

Tableau analytical view or pre-read (Chapter 12)

E.5.2

E.1, E.2, E.3, E.4, E.5, E.6, E.7, E.11 (11.5, 11.11), E.14

E.9 if the workbook carries controls; E.13 if AI-assisted

Ten to fifteen minutes

Executive dashboard or decision artifact (Chapter 13)

E.5.3

E.1 to E.8, E.10, E.11, E.12, E.14

E.9 if interactive; E.13 if AI-assisted

Thirty to forty minutes, plus the reader test

Recommendation memo paired with a visual artifact

E.5.4

E.1 (1.1, 1.2, 1.6), E.2 (2.13, 2.15), E.6, E.10, E.11 (11.1, 11.2), E.12 (12.12)

E.13 if AI-assisted

Ten minutes, plus the third-reader test

1.1

1

The decision the artifact supports is written in one sentence naming the decision-maker, the choice, and the deadline. If the sentence needs an “and,” the artifact is being asked to support two decisions and one of them belongs elsewhere.

1.2

2

The primary reader is named, together with how long she will spend, what she already knows, and what evidence would change her mind.

1.3

1

The artifact’s dominant job is declared — analytical or explanatory — and the design matches the declaration rather than contradicting it.

1.4

2

The form is the right one for that decision and that reader: chart, table, pre-read, dashboard, report, data story, or memo. A one-time decision made in a meeting by a reader new to the material is a data story with a memo attached, whatever was requested.

1.5

1

The intended three-second takeaway is written down and sealed before any test is run. This is the same obligation as check 12.1 and is recorded once.

1.6

2

The action or next question that should follow the reading is stated.

1.7

1

An explanatory artifact was rebuilt rather than inherited. A view carried over from the analytical version keeps every default the analytical version was entitled to and the explanatory version is not.

1.8

2

Where the artifact serves two audiences at once, the dominant job is named rather than averaged, and the checks that follow are run against that job.

1.9

2

An analytical artifact circulated to other readers carries a written reading guide naming the question each view answers, the grain of one mark in each, and the window they share, so that a reader who disagrees knows what she is disagreeing with.

2.1

1

The source table or certified extract is named on the artifact, not only in the analyst’s memory.

2.2

1

The date window is stated, and its start is defensible rather than inherited from a filter nobody set.

2.3

1

Every filter and exclusion, including defaults, is stated in words on the artifact. A control showing a filter’s state is a fact about an interface; a caption stating it is a fact about the evidence.

2.4

1

What one mark represents was written down before the view was inspected, and the view’s dimensions and the measure’s aggregation produce exactly that.

2.5

1

Every rate names its denominator, and the denominator is the eligible base rather than the convenient one.

2.6

1

Any per-order, per-customer, or per-store metric has been verified by hand on one group. Average line value and average order value differ by a factor no software will warn you about.

2.7

1

Totals reconcile to the certified source wherever reconciliation is expected, and any filtered total that no longer matches carries a stated filter.

2.8

1

Distinct counts use the correct key.

2.9

1

A join did not duplicate records: pre- and post-join row counts were compared and the cardinality was enforced.

2.10

1

Absent observations and zero observations are distinguished. A week in which a store was open and sold nothing is a zero; a week before it opened is a null; a line drawn straight across either one is a claim the data does not make.

2.11

1

Missing values were not silently omitted in a way that changes the claim.

2.12

1

A regenerable view has not been replaced by a stale screenshot where current data is required, and the refresh date is visible.

2.13

1

Any reference value is entered as a constant from a certified source rather than computed from the displayed marks. A pooled rate is computed on the pooled denominator and is not the unweighted mean of the displayed rates; a reference computed from the view also moves when the view is filtered, which makes the sentence above it false.

2.14

2

Where a reference legitimately is a descriptive summary of the plotted marks — a median, a mean, a prior period — it is labeled as one, so that a reader knows the line summarizes the marks rather than standing independently of them.

2.15

1

Where an artifact carries figures describing different populations, each figure’s population is named beside it and no figure is used as though it described both.

2.16

1

Every derived date field is reproducible: the window is anchored to the extract rather than to the system clock, and the week-start convention is declared rather than inherited from a data source’s configuration.

3.1

1

The question the view answers is written in one sentence.

3.2

1

The comparison the reader must make is explicit, and it is the comparison the question requires.

3.3

2

The analytic purpose is named — distribution, comparison, trend, relationship, composition, or geographic — before the form was chosen.

3.4

2

The chart form makes that comparison easy, and the wrong-form column of Table 12.1 has been read for this purpose.

3.5

2

A table is used where precise lookup matters more than visual pattern.

3.6

2

Categories are sorted in the order the comparison needs, and the sort was chosen rather than left at the software’s default.

3.7

1

Time runs chronologically, and a line is not drawn across categorical positions, which invents a continuity the data does not have.

3.8

2

Where a time series is drawn as bars, discrete periods are the intended reading. Bars are appropriate for discrete monthly or annual comparisons and discard the continuity the eye reads best when the movement between periods is the finding.

3.9

2

Composition is not shown in a form whose middle categories cannot be ranked, and shares are not compared across segments whose baselines float.

3.10

2

Small multiples were considered wherever an aggregate could hide heterogeneity, and the panel order is reproducible from a field rather than arranged by hand.

3.11

1

The view answers the question it was built for, rather than a different question the software aggregated into.

3.12

1

The units placed on a common scale are comparable. A shared axis is an implicit claim that the things it holds can be compared, and that claim is made by the layout rather than by the data.

3.13

2

Where a purpose required more than one view, the additional views were built rather than the purpose narrowed to fit one.

3.14

1

Where a view carries a claim — an assertion title, a finding in accompanying prose, a recommendation — the question it answers was declared before the view was built. Exploratory views legitimately generate questions rather than answer declared ones; what is not available is producing several views, selecting the one that says something, and presenting it as though its question had been asked in advance, which is the visual form of the multiple-comparisons problem of Section 11.8.

3.15

2

Where a view bins a continuous variable, the bin width was chosen and stated. Binning is a transformation applied before encoding, so it changes the picture without changing the data.

4.1

2

The mark type’s implied reading is the intended one: bars read as discrete comparisons, lines as continuous trends.

4.2

1

The most important comparison is carried by the highest-ranking channel practical — position on a common scale, then position on identical non-aligned scales, then length.

4.3

1

Angle, area, volume, and color intensity are not carrying a magnitude the reader must decode precisely.

4.4

1

Size encodings scale area rather than radius, map zero to zero area, and carry a size legend with labeled reference values.

4.5

2

Shape encodings remain distinguishable at the size the artifact will actually be seen.

4.6

2

Multiple encodings do not contradict one another, and no field is encoded twice without a reason that can be stated.

4.7

1

Three-dimensional effects are absent. There is no analytical case for them.

4.8

2

Legend categories match the field’s level of measurement, and direct labels replace the legend wherever the lookup would cost the reader more than the labels cost the canvas.

4.9

2

Faceting’s cost was accepted deliberately: every panel shares one scale, and the trade of one rung of the hierarchy for the ability to see many subsets at once was the right one for this question.

4.10

1

Where a magnitude comparison and a spatial question were both wanted, they were split across two views rather than forced into one. A map answers where; it does not answer how much.

4.11

1

A filled or shaded map encodes a rate over a defined area rather than a raw count. A choropleth of counts mostly draws a population map, and a map of a per-store quantity mostly draws where the stores are.

5.1

1

Ordinary magnitude bars begin at zero. A bar encodes value in length and length is read as a proportion, so a truncated bar does not exaggerate, it makes a false statement in the channel readers trust most.

5.2

1

Floating, interval, Gantt, and waterfall bars begin at their explicitly defined reference, and the reference is stated on the chart.

5.3

1

Stacked bars and areas begin at zero, and it is stated whether the segments are values or shares.

5.4

1

Every position-encoded axis range was chosen deliberately and is stated where a reader might be misled. A restricted line-chart axis is honest only if the restriction is visible and defensible; “the software chose it” is not a defense.

5.5

2

A logarithmic scale is used only where the question is multiplicative — a growth rate, a ratio, a proportional difference, or a comparison across orders of magnitude — every plotted value is positive, and the scale type is noted for a reader who will not check it.

5.6

1

Axis units are visible, and the aggregation is named where the number is a sum, an average, or a rate.

5.7

1

Panels meant to be compared carry identical scales, and any dual axis in a comparison view is synchronized.

5.8

1

Dual axes are absent unless the two scales are fixed by definition. Two independent ranges manufacture the coupling the reader will read as a relationship.

5.9

1

The displayed window covers at least one full seasonal cycle in a seasonal business, or the shortfall is disclosed and no seasonal claim is made on the strength of it; and the pre-period is shown wherever an event is being credited with what followed it.

5.10

1

The baseline, target, prior period, control, comparison group, or benchmark the reader needs to interpret the values is on the chart. Where there is nothing to compare against, the title contains no comparative word.

5.11

1

Every reference line is labeled, and an externally supplied decision threshold was declared before the results were inspected. A threshold chosen after the fact is a description, not a benchmark.

5.12

2

A decision threshold, a forecast trigger, or an experimental cut is not presented as a neutral constant.

5.13

1

Every axis was re-read after the last change to the view. Ranges recompute silently when a filter is applied, a measure is swapped, or a table calculation is added.

5.14

1

An index or percent-change view draws its base — 100 or zero percent — as a labeled reference line, and names the base period explicitly.

6.1

2

Analytical views carry descriptive titles appropriate to inquiry; explanatory views carry assertion titles, and only where the assertion is fully supported.

6.2

1

Every assertion title is verifiable from the marks in the view directly beneath it, in that view’s own units, with no other evidence admitted. Cover what you know, look only at the marks, and confirm a reader could do it. Every number in an assertion title is therefore visible in the view or directly derivable from what is visible.

6.3

1

The verb is one the evidence licenses. Rose, is, reached, differs, and is associated with are descriptive verbs a view can support; drives, causes, proves, and will are claims about mechanism or the future that a descriptive view cannot carry however strong the analysis behind it.

6.4

1

Every number in accompanying prose — a subtitle, an annotation, a source-band line, a memo sentence — appears in a view or in the certified metrics file the views are built from, and is traceable to whichever it came from. The stricter rule of check 6.2 governs assertion titles, which a reader verifies against one view and cannot check against a file she does not have.

6.5

1

The population in the sentence is the population in the view. Four stores are four stores and not a format; a statement about physical stores is not a statement about the chain.

6.6

1

Percentage change and percentage-point change are distinguished wherever both are available.

6.7

1

Units appear in the axis, the labels, or the title, and every abbreviation is defined.

6.8

2

Labels are direct wherever they remove a legend lookup, and the labeled set is small enough to remain readable.

6.9

2

Annotation explains a relevant event, break, threshold, or exception rather than decorating the view, and the number of annotations is small enough that they retain their force.

6.10

1

The footer carries source, window, filters, grain, and refresh date whenever the artifact leaves the workbook — under every view on an analytical artifact, once for the screen on a dashboard, with per-view grain moved into tooltips only where nothing load-bearing goes with it.

6.11

1

A metric definition accompanies any number a reader could otherwise misinterpret.

6.12

1

A title over a parameterized or filterable view remains true in every state the control can produce, or the title is built dynamically so that it tracks the control.

6.13

2

The claim-making sentences were counted — every title, subtitle, annotation, source-band line, and line of accompanying prose — and that count is the denominator reported against check 6.2.

7.1

1

Palette type matches the field’s level of measurement: categorical for unordered groups, sequential for ordered ones and for quantities, diverging only where a meaningful midpoint exists and both distance and direction from it are the point.

7.2

1

Color is not the only visual means by which any essential distinction is conveyed. Position, ordering, direct labels, shape, line pattern, or lightness carries it as well.

7.3

1

The grayscale screening has been run: convert the artifact and confirm the distinctions survive. Collapse means rebuild; survival means one test passed rather than all of them, because grayscale exposes reliance on hue and weak lightness contrast and does not simulate every deficiency.

7.4

1

The artifact has been inspected through a color-vision-deficiency simulator. Roughly one in twelve men of Northern European descent has some form of red–green deficiency, and the deficiency is invisible to everyone else in the room including the person who chose the palette.

7.5

1

Contrast meets a stated criterion at the condition the artifact will actually be read in — a projector in a lit room, a phone, a printed page — rather than on the analyst’s monitor. This guide adopts WCAG 2.2 Success Criterion 1.4.3 for text, at 4.5 to 1 against its background, and 1.4.11 for graphical objects and interface components, at 3 to 1 against adjacent colors (W3C, 2024).

7.6

2

Red-to-green is not used as an unexamined good/bad shorthand, and no unordered rainbow scale carries an ordered quantity. An ordered quantity takes a perceptually ordered palette — sequential where it has no meaningful center, diverging where it does — chosen for the variable and for the comparison the reader must make.

7.7

2

Neutral categories are not assigned unintended positive or negative valence.

7.8

2

Emphasis color is spent rather than spread: one accent for the subject, gray for the context, and an emphasized set small enough that the reader does not have to search inside it.

7.9

2

Background, gridlines, borders, and decoration are de-emphasized or removed, and every retained non-data element is doing a job that can be named. Reference lines, direct labels, annotation, and the footer are non-data ink and are kept.

7.10

2

A brand palette is used for identity and accent, with any sequential ramp derived by varying lightness within one brand hue and a neutral gray as the workhorse. A diverging palette takes two accessible hues around a neutral midpoint, since it cannot be built inside a single hue. Where a brand guideline and an encoding requirement conflict on an analytical artifact, the encoding wins.

7.11

2

A categorical palette carries no more than seven or eight hues. Beyond that, color alone stops being a reliable identifier and the distinction moves to direct labels or to another channel.

8.1

1

The KPI hierarchy was written down before the canvas was opened: one or two primary views, a supporting tier, a diagnostic tier, each defined by the reader’s question it answers.

8.2

1

Prominence matches the hierarchy. The view the decision turns on is the largest object and sits on the reader’s entry path, even when it is the plainer object.

8.3

2

Position and prominence agree rather than fight; a large high-contrast object elsewhere on the canvas does not capture the entry fixation.

8.4

2

Reading order matches the sequence in which the reader must acquire the argument.

8.5

2

Views in the same tier are adjacent and share their formatting; views sharing a scale are aligned, and their axes were confirmed identical rather than assumed.

8.6

1

The number of findings the reader must hold at once was stated before layout and is small. Everything else is available in sequence rather than simultaneously.

8.7

2

Whitespace separates functions rather than filling unused space, and padding was set rather than left at zero.

8.8

1

Every view can be justified by its tier and by the point in the reader’s sequence at which she needs it. A view with no justification comes off the canvas and goes into the memo’s appendix; it does not become a smaller view.

8.9

1

The case’s weakest facts are on the canvas rather than in a backup. A reader who discovers a weakness herself concludes that the analyst either missed it or hid it, and both conclusions are expensive.

8.10

2

The sizing mode matches the consumption condition — fixed at a known projection resolution, device-specific for unknown displays — and the choice was recorded rather than defaulted.

8.11

1

The squint test was run: blur the artifact until the text is unreadable and confirm the intended object still dominates. If it does not, the hierarchy is in the analyst’s head and the repair is size and position rather than a bigger title.

8.12

1

The primary message survives at the intended display size, and no essential evidence sits below the fold or behind an unexplained interaction.

8.13

2

The artifact can be read without the analyst narrating every element.

8.14

2

Nothing was solved by shrinking. A canvas of twelve small views is not a decluttered canvas of twelve views; it is an illegible one.

8.15

2

Where the authoring tool offers layout containers — as Tableau does — the canvas is built from nested containers rather than floating objects, so that the tiers hold their relationships when the artifact is resized or something is added late. In a tool without them, the equivalent obligation is that the layout survives a change of size or a late addition without being rebuilt.

8.16

2

The title band carries the decision the screen supports and the date rather than a restatement of the data, and the analyst of record is named on it.

9.1

1

Every finding the artifact must communicate is visible in the default state. Interactivity may add depth to a finding that has already landed; it may not deliver one.

9.2

1

Each control exists to answer a reader question the analyst can name.

9.3

1

The state the reader meets is disclosed in words on the canvas rather than shown only in a control. Where a screen that could carry controls carries none, it says so, because a screen with nothing to adjust otherwise makes no claim about its own completeness.

9.4

1

No control can falsify a sentence on the same screen. Where no safe scope exists, the control comes off and the state is declared in words instead.

9.5

1

Filter scope was chosen deliberately and stated. A filter applied to every worksheet on a data source will move views whose titles quote chain-wide figures and whose references were computed on another population.

9.6

1

A filter written against a field that only one kind of row possesses does not silently delete the other kind.

9.7

1

No filter or action combination produces an empty, partial, or nonsensical state without explanation — a view emptied of its marks, a frame whose count no longer matches the sentence above it, or a region deleted while its caption remains.

9.8

2

The reader can return to the default state, and the route back is discoverable.

9.9

1

Cross-filtering does not silently change a denominator.

9.10

2

Where one view’s selection drives another — a Tableau highlight action, a cross-filter, a linked selection — the interaction runs from the smaller value set to the larger, every target contains the values the source passes it, and each was tested and observed to fire. An interaction that does nothing reads as a broken artifact rather than as an empty match.

9.11

1

Nothing load-bearing lives in a tooltip. A qualification that makes a headline defensible belongs on the canvas.

9.12

2

Navigation actions carry a visible cue that the deeper view exists.

9.13

1

The static export still contains the essential claim and its context.

9.14

2

The interaction inventory is complete: for every interaction, the reader question it answers, what a reader who never triggers it misses, and confirmation that it fires. Where the answer to the second names a finding, check 9.1 has failed and the finding moves onto the canvas.

10.1

1

Material uncertainty is present wherever the decision depends on it. A point estimate presented alone transfers risk from the analyst’s credibility to the organization’s balance sheet.

10.2

1

The interval is named correctly — prediction interval, confidence interval, simulated range — and its level is stated.

10.3

1

The interval is expressed in decision units, and the conversion is shown on the canvas in steps with its denominator named rather than asserted as a result.

10.4

1

Where a conversion would require an assumption that cannot be established, what is measured or forecast was changed so the assumption is not needed; where that is impossible, the assumption is disclosed where a reader will find it rather than buried in an arithmetic step.

10.5

1

Where both framings are used, both are given: the end-to-end width of the range and the spread around the central estimate. Quoting the first alone overstates how uncertain the estimate is.

10.6

1

Interval bounds exist as values at the display’s own grain, computed upstream rather than requested from the drawing tool. A region whose edges move across the horizontal axis is a shape built from those values — in Tableau a polygon over computed vertices, since a reference band renders at constant height; in a plotting library, a fill between two series.

10.7

1

An interval over an aggregate is not produced by summing the bounds of its component periods, which overstates the width. Where the interval comes from a simulation, the aggregate is formed on each path and the quantiles are taken from the aggregated totals; where it comes from another method, the aggregate interval is derived by that method’s own rule and the rule is named.

10.8

1

Periods that have already resolved are shown as actuals outside the interval, with the seam marked and the forecast origin stated. Presenting elapsed months inside a band prices risk that no longer exists.

10.9

2

False precision has been removed and every assumption is named.

10.10

1

The range is not visually compressed — by an over-wide axis, a long empty tail, or a scale chosen after the interval was known — so as to appear more certain than it is.

10.11

1

The artifact does not imply that an observational relationship is an experimental effect.

10.12

1

The complete decision artifact carries all four parts: the recommendation, the evidence, the uncertainty, and the conditions under which the recommendation would change.

10.13

2

Where the cost of being wrong is asymmetric, the asymmetry is stated in comparable units, with both subtractions shown rather than two figures set against each other that are not comparable.

10.14

1

Every assumption supplied rather than measured — a planning figure, a cost, a capacity — is attributed to its source on the artifact rather than presented as something the data produced.

10.15

1

A recommendation states the rule before any count, and any count it carries is reconcilable to a table the reader can be shown. A count that cannot be reproduced from a stated rule is an opinion with a number attached.

10.16

2

The reversal conditions are a small, enumerable set — two or three is the working shape — each with a date or a threshold attached, and at least one names what would be re-estimated when the underlying data next refreshes.

10.17

2

Where part of a forecast horizon has already resolved, the model’s prediction for those periods was compared against what was reported, and the comparison is disclosed as evidence for or against the range on the periods still open.

11.1

1

The counterfactual has been asked and answered in writing: would this reader revise her conclusion if she saw everything the analyst saw? A yes means the compression or the emphasis requires repair or disclosure.

11.2

1

Every removal made for the audience is listed — views cut, categories grouped, windows shortened, filters defaulted or scoped, qualifications moved into tooltips — and each is sorted into a legitimate simplification or the deletion of a qualification.

11.3

1

Grouping into “other” does not hide a category whose individual value would change a ranking or a conclusion.

11.4

1

The quantity is presented in a form that does not select the reader’s reaction, and an equally true alternative framing was considered rather than avoided.

11.5

1

The views shown represent the evidence rather than the favorable part of it, and inconvenient findings remain available or are disclosed.

11.6

1

A title does not select the favorable reading where the same view equally supports an unfavorable one that the decision turns on.

11.7

1

Axis, canvas, and scale choices do not minimize uncertainty or adverse outcomes.

11.8

2

The artifact would survive its own audience being fully informed: hand the reader the full evidence and an afternoon, and expect the same decision. An artifact that depends on the reader not having that afternoon is an argument requiring an uninformed audience.

11.9

2

The standard applied was did I check, which leaves evidence, rather than did I intend, which cannot be audited by anyone.

11.10

2

Emphasis was not abandoned in the name of neutrality. A screen with no hierarchy is not a neutral screen; it is one whose hierarchy was set by the order the views were built in, which still produces a conclusion in the reader’s head and is unaccountable for it.

11.11

2

The views that were built and rejected are listed, with one line apiece on why, so that a reader can see the analyst’s search rather than only its result.

11.12

1

A case singled out for attention was selected by a rule declared in advance rather than chosen after inspection and then explained, which is post hoc hunting in a different costume.

12.1

1

The intended takeaway was written down and sealed before the test. A takeaway written afterward is not a prediction and the protocol has no value without one.

12.2

2

Three further predictions were sealed: what the reader’s eye lands on first, the first question she asks, and the element most likely to be misread.

12.3

1

The reader is representative and has not seen the artifact. A colleague who built the same thing last week knows where to look and is a test of nothing.

12.4

1

The artifact was shown for approximately three seconds, removed, and the reader’s takeaway recorded verbatim rather than paraphrased.

12.5

1

A ninety-second task followed — “tell me what you would decide and why” — with reading order, questions, hesitations, and the stated decision recorded.

12.6

1

The analyst said nothing during the task: no explanation, no guidance, no answering of questions until the time was up.

12.7

2

A mismatch in the subject of the reader’s sentence was investigated as a hierarchy failure first, since size and position are the cheapest repair and the most common cause.

12.8

2

A mismatch in the verb was investigated as a title or emphasis failure first, and repaired in language and color before anything structural was changed.

12.9

2

Views the reader skipped entirely were distinguished from views that failed to persuade. A skipped view was never read, and the two have different repairs.

12.10

1

Material mismatches produced a revision and a retest on a fresh reader; accepted defects were recorded with one sentence of reason each.

12.11

1

The result is reported as a formative check on one or two readers, which finds gross hierarchy and language failures, rather than as validation that the artifact works for its actual audience.

12.12

2

The memo alone, without the artifact, was handed to a third reader who could state the recommendation and the basis for it.

12.13

2

Where a peer critique was run, it followed the protocol in order — report the apparent takeaway, ask what decision the artifact supports, identify the specific elements that impede reading, and only then propose changes — with the author listening without defending until the sequence is complete.

12.14

2

Every piece of feedback declined is recorded with one sentence of reason, alongside the changes that were made and the evidence that prompted each.

13.1

1

The faults were predicted from the specification, in writing, before the view was rendered. A rendered chart is persuasive in a way its specification is not, and the human tendency is to evaluate whether it is attractive rather than whether it is right.

13.2

1

The specification was built exactly as written, faults included, so that the faults can be audited rather than corrected invisibly. This applies in an instructional or otherwise safe environment; a specification that would create a privacy, security, cost, or irreversible operational exposure is audited on paper and never executed.

13.3

1

No field, filter, value, or business fact was invented. Every column referenced exists in the schema that was supplied.

13.4

1

The grain and the aggregation were stated by the analyst rather than chosen by the assistant.

13.5

1

Software and assistant defaults were inspected individually: axis range, sort order, palette, aggregation, filter state, mark type.

13.6

1

Candidate titles were verified against the built view one sentence at a time and rejected on first failure rather than repaired.

13.7

1

No causal or predictive verb was introduced at the narration step, where assistants reintroduce it even when the numbers are clean.

13.8

1

No interval, threshold, qualification, or reversal condition was removed during a rewrite. The two versions were diffed clause by clause rather than the new one read.

13.9

1

A proposed layout implements the analyst’s KPI hierarchy, supplied as an input, rather than a conventional grid the assistant has no basis for.

13.10

1

No connective sentence asserts a relationship between two views that no view establishes. It reads as synthesis; it is confabulation, and only the traceability rule catches it.

13.11

2

Alt text, footer content, contrast, and filter-state disclosure were requested explicitly, because nothing in a request for a headline or a layout will produce any of them unasked.

13.12

1

The prediction was graded: both the faults missed and the faults predicted that did not occur are recorded, since a systematic error in the analyst’s model of the assistant is worth more than any single caught title.

13.13

1

The exchange is documented per Appendix D, including the output that was rejected.

13.14

1

The analyst can explain and defend every encoding and every sentence in the artifact. An assistant can write a sentence; only the analyst can sign it.

13.15

1

Every candidate sentence arrived with its own verification line naming the marks a reader would have to look at to confirm it. A candidate whose verification line is vague is a candidate that cannot be verified.

13.16

2

The surviving title was shown on its view to someone who had seen neither, and her paraphrase was the claim the analyst intended rather than the stronger one next door.

14.1

1

Final dimensions match the delivery channel, and text is readable at actual size in the room, on the page, or on the device where it will be read.

14.2

1

The static export preserves labels, legends, reference lines, annotations, and the footer.

14.3

1

Interactive links open with the intended permissions and in the intended default state.

14.4

2

The workbook, notebook, or query accompanies the artifact wherever reproducibility is required.

14.5

1

The file name and version are unambiguous, and the version on the artifact matches the version in the sign-off record.

14.6

1

The data refresh date is visible wherever recency matters.

14.7

1

Alternative text describes the view’s purpose, population, units, main pattern, and any material exception — not the field names, which is what the software supplies by default.

14.8

1

Every recipient is authorized for everything the delivery carries. Labels, tooltips, annotations, file names, and the extract behind the artifact expose no protected field and no more detail than the decision requires, permissions on any interactive link match the intended audience, and nothing travels that the analyst is not authorized to disclose to the people who will receive it.

14.9

2

A decision-supporting artifact travels with its written recommendation rather than alone.

14.10

2

The sign-off record of Section E.7 is complete, and the analyst of record is named on the artifact or travels with it.

What decision or question does this view support, and is its job analytical or explanatory?

1.1, 1.3

What question does this view answer, and what comparison must the reader make?

3.1, 3.2

What does one mark represent, and do the dimensions and aggregation produce that?

2.4

Is the channel carrying the main comparison the best one available?

4.2

Does every length-encoded mark begin at zero, and was every axis range chosen and stated?

5.1, 5.4

Is the reference the reader needs on the chart, is it labeled, and was it entered from a certified source?

5.10, 5.11, 2.13

Does the palette match the field, and does the chart survive grayscale?

7.1, 7.3

Does the title claim only what the marks support, with its units and its verb?

6.2, 6.3, 6.7

Does the footer carry source, window, filters, grain, and refresh date?

6.10

Was the sort chosen rather than inherited?

3.6

Is the artifact’s job declared as analytical, and does its density serve inspection rather than speed?

1.3

Is there a written reading guide naming the question each view answers, its grain, and the shared window?

1.9

Does every view state its own grain, window, filters, and source?

2.1, 2.2, 2.3, 2.4, 6.10

Do the totals reconcile, and has one per-unit metric been hand-checked?

2.6, 2.7

Are the units placed on any shared scale genuinely comparable?

3.12

Are the pre-periods, comparison groups, and full seasonal cycles present?

5.9, 5.10

Were small multiples considered wherever an aggregate could hide heterogeneity?

3.10

Are definitions and scales consistent across views, and every reference a certified constant?

2.13, 5.7, 6.11

Are the titles descriptive rather than assertive, so that the reader forms her own conclusion?

6.1

Is the inconvenient view present, and are the charts that were built and rejected listed?

11.5, 11.11

Does the delivered export carry the labels, references, and footers the workbook shows?

14.2, 14.6

Is the decision written in one sentence, with the reader, the choice, and the deadline?

1.1, 1.2

Was every view rebuilt rather than inherited from the analytical version?

1.7

Is the KPI hierarchy written down, and does prominence match it under the squint test?

8.1, 8.2, 8.11

Does the title band carry the decision and the analyst’s name?

8.16

Is every sentence on the screen verifiable from the marks beneath it, in its own units, with a licensed verb and no imported number?

6.2, 6.3, 6.4, 6.5

Are the case’s weakest facts on the canvas?

8.9

Is the uncertainty present, in decision units, with the conversion shown and its denominator named?

10.1, 10.3

Are the assumptions attributed to their sources, and the reversal conditions stated with dates or thresholds?

10.14, 10.16

Is the default state disclosed in words, and can no control falsify a sentence on the same screen?

9.3, 9.4

Is every finding visible without interaction, and does no control produce an unexplained empty state?

9.1, 9.7

Was the three-second test run on a fresh reader, and did her takeaway match the sealed one?

12.1, 12.4

Is the removal list written and sorted?

11.2

Does the export carry labels, alt text, refresh date, and no restricted data?

14.2, 14.7, 14.8

Does it open with a recommendation — an action or a rule — rather than a finding wearing a recommendation’s clothes?

10.12

Is the recommendation a rule before it is a count, and can any count be reconciled to a table in the appendix?

10.15

Does every evidence sentence name the view it comes from and state its finding in decision units?

6.2, 10.3

Is every verb one the evidence licenses, and every population the one the evidence describes?

6.3, 6.5

Do the memo’s numbers agree with the artifact’s, sentence by sentence?

6.4

Is the uncertainty in the reader’s units, with the arithmetic and its denominator shown?

10.3, 10.5

Are the scope limits stated — the populations the evidence does not cover, and the false precision removed?

2.15, 10.9

Is every assumption supplied rather than measured attributed to its source?

10.14

Are there two or three reversal conditions, each with a date or a threshold attached?

10.16

Are the removal list and the Appendix D documentation in the appendix?

11.2, 13.13

Handed to a third reader without the artifact, can she state the recommendation and its basis?

12.12

5.1

Redrew the quarterly bars from a zero baseline.

A reader now sees a quarter that moved slightly, rather than a quarter that appeared to jump twelvefold.

5.8

Replaced the dual axis with two panels sharing the time axis, and added the urban series as the comparison group.

A reader now sees two series she must compare deliberately, rather than one apparent relationship she was invited to accept.

3.9

Rebuilt the seven-category pie as a sorted bar, then as a 100 percent stacked bar split by metro.

A reader can now rank the middle categories, which the pie made impossible, and can see the share comparison the pie was pretending to answer.

7.1

Replaced the red-to-green tier palette with a single-hue sequential ramp ordered None to Backstage.

A reader now sees four ordered tiers, rather than two tiers she cannot distinguish and a good/bad judgment the tiers do not carry.

2.4

Repointed the sheet at the order-grain extract and relabeled the measure.

A reader now sees average order value, rather than average line value under a label that promised the first.

6.2

Replaced an assertion title with the sentence the view’s marks support.

A reader now takes away a claim she can verify from the chart in front of her, rather than one she must take on trust.

Artifact title

The name a reader will see, not the file name.

File and version

The delivered file, its version, and the format it was delivered in.

Data refresh date

The date of the extract or the certified source behind the artifact.

Population and window

The population every store-level or unit-level figure describes, and the window, in the words used on the artifact.

Instrument run

Whether this was the baseline audit — the applicable master checkpoints run in full — or a short-form re-run against an existing baseline, and which of the four forms of Section E.5 was used as the front end.

Completed checklist

The file name, spreadsheet tab, or record identifier of the completed checklist the counts below are taken from. A sign-off without this line asserts its counts rather than evidencing them.

Checkpoints run

The checkpoints run, and the checkpoints recorded as not applicable with a three-word reason each.

Unique applicable check IDs

How many distinct checks applied to this artifact, given the routing of Section E.3 and the checkpoints marked not applicable. Counted once for the artifact, not once per view: the individual instances live on the completed checklist.

Unique check IDs passed

How many distinct applicable checks were run and satisfied on every instance.

Unique Level 2 check IDs accepted

How many distinct Level 2 checks were accepted as Cautions. Must equal the number of entries in the accepted-defects field.

Unique Level 1 check IDs repaired

How many distinct Level 1 checks failed anywhere in the artifact and were repaired everywhere they failed. Must equal the number of entries in the stop-ship field. One check failing on four views is one identifier, four instances, and potentially four repair notes.

Unique Level 1 check IDs unresolved

How many distinct Level 1 checks failed and were not repaired. Any number above zero means the artifact has not been released; if it was released anyway, the override and the name of the person who authorized it go here.

Stop-ship failures found

Every Level 1 failure by identifier, with its repair note per Section E.6.

Communication revisions

Every Level 2 change made, by identifier.

Accepted defects

Every Level 2 check recorded as Caution, by identifier, with one sentence of reason.

Sentence count

For an explanatory artifact: how many claim-making sentences were checked against 6.2, and how many were changed.

Removal list

The C3 disclosure of check 11.2: what came off, and whether each removal is material to the decision.

Usability evidence

The sealed intended takeaway, the reader’s verbatim three-second takeaway, the reading order observed, and the revisions that followed.

AI-use record

The Appendix D record identifier for every exchange that touched the artifact, and what was rejected.

Remaining limitations

What the artifact still does not establish, in the analyst’s own words.

Analyst of record

Name and date. The person who circulates the artifact owns every encoding in it.

Reviewer

Name and date, where a second reader ran the checklist.

Artifact title

Suburban store performance — the four views reproduced and repaired in Lab 12.2

File and version

suburban_deck_v1, four Tableau worksheets, delivered as PNG exports. Two further views travelled in the same email and are not covered by this record

Data refresh date

Certified extract, line grain and order grain. No refresh date appeared on any of the four as delivered

Population and window

Suburban and resort store types. Two most recent quarters on view one; full certified window on views two through four. Neither the population nor the window was stated on any of the four

Instrument run

Baseline audit: the applicable master checkpoints run in full on each of the four views. Front-end form: E.5.1, run once per view

Completed checklist

appendix_e_signoff_suburban_deck_v1.xlsx, tabs V1 through V4

Checkpoints run

E.1, E.2, E.3, E.4, E.5, E.6, E.7, E.11, E.13, E.14. E.8 not applicable: separate views, no canvas. E.9 not applicable: static exports, no controls. E.10 not applicable: no forecast, estimate, or threshold. E.12 not applicable: working exchange between colleagues, no reader test required

Unique applicable check IDs

89, after the four not-applicable checkpoints and the per-view exclusions recorded on each tab. Counted once for the artifact; the 4 views produce 274 individual check instances, which live on tabs V1 through V4

Unique check IDs passed

71

Unique Level 2 check IDs accepted

1

Unique Level 1 check IDs repaired

17

Unique Level 1 check IDs unresolved

0

Stop-ship failures found

1.1 and 1.3 — no decision sentence, and the views’ job never declared before they were forwarded as decision support. 2.1, 2.2, and 2.3 — source, window, and filters stated on none of the four. 2.4 and 2.6 — a per-order claim carried on line-grain marks, unverified by hand. 2.12 — no refresh date. 4.3 — the pie puts a magnitude the reader must rank onto angle. 5.1 — view one’s axis runs from $240,000, drawing the second bar at roughly twelve times the first against a value ratio of 1.046 to one. 5.8 — view two’s dual axis carries two independently scaled ranges. 6.10 — no footer on any of the four. 6.11 — the covering email calls view four “average order value” while the worksheet computes an average of line revenue. 7.1 and 7.2 — a red-to-green palette on an ordered field, with hue carrying the distinction alone. 14.2 and 14.6 — the delivered exports carry neither footers nor a refresh date. Seventeen distinct checks, failing across the four views on 31 instances in all; every instance repaired before recirculation, with its repair note on tabs V1 through V4

Communication revisions

3.9 — the pie replaced by a sorted bar and then by a 100 percent stacked bar split by metro. 3.6 — sort applied to the rebuilt bar. 7.8 — accent hue reserved for the series under discussion on the rebuilt trend view

Accepted defects

11.11 accepted as Caution: no list of charts built and rejected was kept, because the views were drafted in one pass and no alternatives were built. Recorded rather than repaired, and it is the reason the next exchange starts with a specification

Sentence count

Five claim-making sentences checked against 6.2 and 6.4 — four titles and the covering email. Two changed; the email’s “average order value” sentence withdrawn

Removal list

Nothing removed. The pie was replaced rather than cut, and every repair changed an encoding rather than the evidence on display

Usability evidence

Not applicable; a working exchange between colleagues, and the reader test of Checkpoint E.12 is reserved for decision artifacts

AI-use record

Appendix D record for the original request, the response, and the four repairs. Checkpoint E.13: 13.1 failed on the original exchange, since no faults were predicted before the specification was built, and passed on the graded re-run; 13.2 passed, the specification having been built as written with its faults intact; 13.5 failed on four inherited defaults; 13.12 and 13.13 completed

Remaining limitations

The repaired views establish that suburban and resort revenue rose 4.6 percent quarter over quarter. They do not establish that the movement is larger than ordinary quarter-to-quarter variation in this series, and the eight-quarter trend view built in the repair is the artifact that would answer it. Nothing here covers the two views that travelled in the same email

Analyst of record

[Name], [date]

Reviewer

[Name], [date]

2.7; 12.9

The baseline discipline, and its visual form as the reference line

E.5

3.3; 3.4; 12.10

Grain, levels of measurement, and aggregation in the view

E.2, E.7

12.1; 12.2

Visualization’s two jobs, and the drift from the first to the second

E.1

12.3; 12.4

Marks, encodings, and the perceptual hierarchy

E.4

12.5

Chart selection by analytic purpose

E.3

12.6; 12.7

Scales, the zero-baseline obligation by encoding, and the five mechanics

E.5

12.8

Color as an encoding system, and accessibility as correctness

E.7

12.9

Reference lines, annotation, and the footer discipline

E.5, E.6, E.14

12.11

Faceting and small multiples

E.3, E.4

12.12

The AI chart-drafting routine and its nine failure modes

E.13

12.16

The disclosure standard and the duty to inspect

E.11

13.2

Audience analysis and the three-second test

E.1, E.12

13.3

Exploratory and explanatory artifacts, and choosing the form

E.1

13.4; 13.5

The KPI hierarchy and its expression in layout

E.8

13.6

The assertion title

E.6

13.6

Narrative order across a set of views

E.8

13.7

Preattentive emphasis and data-ink decluttering

E.7, E.8

13.8

Interactivity that communicates and interactivity that conceals

E.9

13.5; 13.12.4

Dashboard sizing, containers, the title band, and alternative text

E.8, E.14

13.9

Uncertainty in decision units, and the four-part recommendation

E.10

13.10

The memo, the usability test, and the critique protocol

E.12

13.11

The AI communication routine and its eight failure modes

E.13

13.15

Emphasis, distortion, and the counterfactual test

E.11

Table 8.5 (three transferable points)

Frame, baseline, and error in decision units

E.1, E.5, E.10

Table 12.5 (V1–V3)

Grain and aggregation; scale and baseline honesty; encoding fits the purpose

E.2, E.5, E.3 and E.4

Table 13.6 (C1–C3)

Every sentence earned; the reader’s takeaway; what was removed is disclosed

E.6, E.12, E.11

Chart-selection theory, the perceptual hierarchy’s evidence, and the grammar of a chart

Chapter 12 owns all three. Checkpoints E.3 and E.4 enforce the conclusions and cite the sections; an appendix that re-argued them would be a second chapter and would drift from the first.

Upstream data-quality checks

Chapter 4 owns the verification log and the defect catalog. Checkpoint E.2 checks that the artifact is consistent with a certified source; it does not certify the source.

Metric definitions

Appendix B is the canonical dictionary. Checkpoint E.6 checks that a definition accompanies a number a reader could misread; it does not supply the definition.

Statistical guidance on choosing an interval

Chapters 10 and 11 own the intervals. Checkpoint E.10 checks that the interval that exists is displayed honestly and converted correctly.

Prompts for AI assistance

Appendix C owns the templates. Checkpoint E.13 is the audit that runs on their output, and Templates C.16 through C.18 route back here for the repair.

A pass/fail score, a weighting, or a grade

The checklist distinguishes hard errors from judgment calls and records both. Converting that into a single number would suppress the distinction, which is the appendix’s most useful property.

Project rubrics

The rubrics are a separate instrument and are released with the project briefs. This appendix is incorporated into Project #2’s submission workflow as the sign-off record of Section E.7, whatever the rubrics’ final home.

Screenshots of software interfaces

Numbered checks and build paths in the chapters are more durable and more accessible than an image of a menu that will be redrawn.

Vendor-specific rules

Every check is written to be tool-neutral. Where a Tableau behavior is named — reference bands, filter scope, alt text, action matching — it is because the behavior is what makes the check necessary, and the vendor’s documentation governs the current interface.