The dashboard derives three different fractions from six fictional records. Overall error is 3/6, or 50%. Compact has two errors in two records, but contributes only two of the three errors. Its filter exposes R1 and R2 alongside error rate 100%, coverage 66.67%, and representation 33.33%. Every value is calculated from the authored table, not copied from a live model assessment.
Keep target and prediction together
The constant-B baseline predicts B for every row. R1/R2 are compact with target A; R3/R4 are roomy with target B; R5 is roomy with target A; R6 is roomy with target B. Comparing target and prediction produces errors R1,R2,R5. Keep fixture V1 and baseline labels attached to this record set. Do not maintain a separate manual error flag that could survive a later prediction change and make the dashboard stale while still looking coherent.
Compact error rate divides two errors by its two records: 2/2. Coverage divides those errors by all three errors: 2/3. Representation divides its two records by all six evaluable records: 2/6. Roomy similarly has 1/4 error rate, 1/3 coverage, and 4/6 representation. Microsoft's error-analysis concepts distinguish these population questions. This example explicitly expresses representation as a share of the evaluable population, keeping the count and denominator next to every displayed percentage.
Filter rows without changing the question
Selecting Compact changes the visible rows, not the full-population denominator. Overall records stay six and all errors stay three. Select Roomy to see R5, the error that disproves “Compact contains every error.” Clear the filter to restore all six. A summary should name the relevant IDs as well as the cohort label. Without that trace, changing the display filter can accidentally change the reported question from coverage of all errors to coverage of only visible errors.
Add the separate R7 scenario to inspect missing truth. R7 has prediction B but no target, so the page reports six evaluable of seven supplied and lists it as unscorable. Under this declared policy it contributes to none of the three metric denominators. Missing truth is not a correct prediction and not an error. A real reporting workflow can choose a different population policy, but should label it and keep linked fractions consistent rather than silently mixing supplied and evaluable counts.
Preserve undefined values
The zero-error scenario replaces predictions with the supplied targets. Overall error becomes zero, but coverage now divides by zero total errors and is N/A. An empty cohort also makes its error rate N/A because it contains no evaluable records. Neither result should become an invented 0/0 percentage or a success badge. The controls let reviewers inspect these boundary cases directly, with the record table remaining available to explain why the requested percentage has no denominator.
Use the error IDs to begin review, not a responsibility score. In this fixture, constant-B predictions and A targets explain the authored mismatches. Layout does not establish causation, and six records cannot certify fairness or future performance. Maintain native inputs, prediction exports and evaluation revisions as the actual evidence. The useful result is a review in which someone can name each erroneous record and explain exactly what each displayed percentage divides.
Make your own working example
Build an original error-cohort dashboard. Fixture V1 has R1/R2 compact target A prediction B; R3/R4 roomy target B prediction B; R5 roomy target A prediction B; R6 roomy target B prediction B. Derive errors from target-versus-prediction, not stored manual flags. Label the baseline authored constant-B, not a model run.
Provide All, Compact, Roomy, Empty cohort, Add missing-target R7, Zero-error scenario, and Reset buttons. Display overall error 3/6=50%; compact rate 2/2=100%, coverage 2/3=66.67%, representation 2/6=33.33%; roomy 1/4=25%,1/3=33.33%,4/6=66.67%. Keep population denominators unchanged when filtering and expose the filtered record IDs.
R7 has missing truth and prediction B; report 6 evaluable of 7 supplied and exclude it from every metric denominator. The active scenario label must describe target-matching predictions during Zero-error and return to authored constant-B on Reset. Zero errors makes coverage N/A; empty cohort makes error rate N/A. Reset restores the six original rows and predictions. Keep fractions and sample rows together near the top, with a static explanation of all three denominators and N/A for a zero denominator. Do not imitate vendor widgets, certify fairness, assert causation, or invent production statistics.
Use one self-contained HTML file with no remote dependency. Keep the initial sample, output, and labeled controls near the top. Use at least 16px text, strong contrast, visible focus, and a stacked layout at 360px without horizontal overflow. Use explicit type=button click handlers and explicit DOM selectors; do not depend on native form submission. Clear stale feedback when task inputs change. Render entered text safely. Keep practice data local to the current page; do not imply saving, sending, or production integration. Return only the finished HTML.
Annotate the cohort discussion
Keep prediction exports in the native assessment workflow. For the HTML explanation, NOS lets you revise eligible static notes visually without asking AI to regenerate the page. Clarify the denominator a reviewer questioned while keeping computed metrics tied to their source records.