Ethics and Privacy Gatekeeper (Locked, Non-Overridable)
Runs first, before any domain scoring. Overrides the overall-score ceiling. HC#11 makes this non-overridable under user pressure.
Severity Tiers
- Critical. Cap overall score at 59/100. Report renders "Critical ethics violation" banner as the first section after the header.
- Warning. Cap overall score at 79/100. Report renders "Warning ethics observation" banner.
- Info. No cap. Noted in the report's ethics section as an observation.
Multiple violations: highest severity wins the cap. All violations render in the ethics section.
Critical Triggers
E-DECEPTIVE-SCALE
Signal. Y-axis or X-axis truncation that visually implies a magnitude the underlying data does not support. Common patterns: bar chart with y-axis starting well above zero producing apparent 300% growth on 8% real growth; dual-axis chart where one axis scale is compressed to make two series look correlated.
Cap. Critical → 59/100.
Recommendation. Start numeric axes at zero for bar charts, or add an explicit axis-break annotation. For dual-axis, either use single-axis with Measure Names color encoding, or note the scale difference in the caption.
E-CHERRY-PICKED-COMPARISON
Signal. Comparison against a single hand-picked baseline that flatters the current period (e.g., "70% growth vs. worst month last year") without disclosing the selection.
Cap. Critical → 59/100.
Recommendation. Use year-over-year or period-over-period comparisons that are clearly documented, or use a rolling average rather than a single-point comparison.
E-MISLEADING-TIME-WINDOW
Signal. Time-series chart with a truncated window that hides a countervailing trend visible in the fuller dataset (e.g., 3-month window showing growth on top of a 5-year decline that would flip the interpretation).
Cap. Critical → 59/100.
Recommendation. Show the fuller time window or annotate the window choice with rationale.
E-MISSING-BASELINE-ON-RATE
Signal. Rate chart (conversion rate, retention rate, error rate) without a baseline or denominator disclosure that could mislead the reader about scale.
Cap. Critical → 59/100.
Recommendation. Add a baseline reference line or disclose the sample size.
E-PII-PHI-VISIBLE
Signal. Personally identifiable information (names, email addresses, government ID numbers, medical record numbers, exact home addresses) visible on the rendered dashboard without redaction.
Cap. Critical → 59/100.
Recommendation. Redact via anonymization function, aggregation to a higher grain, or masking before publish. Route to data-quality-review for a data-side audit.
E-PROPRIETARY-UNREDACTED
Signal. Proprietary data (customer names, contract values, competitive intelligence) visible without watermarks or classification indicators when the dashboard's audience includes External per stated intent.
Cap. Critical → 59/100.
Recommendation. Add classification banner or restrict audience.
Warning Triggers
E-HIDDEN-ASSUMPTION
Signal. A calc or filter that materially changes the interpretation is not disclosed on the dashboard (e.g., "excludes returns" applied silently to a Sales dashboard).
Cap. Warning → 79/100.
Recommendation. Add caption note or explicit filter card visible on the dashboard.
E-COLOR-DECEPTION
Signal. Color choices that map bad-to-good in a way that misleads (e.g., red for "target met" and green for "target missed", or a divergent palette applied to a sequential measure).
Cap. Warning → 79/100.
Recommendation. Align color mapping with the shared portfolio catalog (blue/teal for positive, orange/magenta for alert, gray for neutral, plus mandatory shapes for accessibility).
E-SELECTIVE-EMPHASIS
Signal. Formatting emphasis (bold, size, color highlight) applied to a subset of data in a way that flatters or damages a comparison without clear rationale.
Cap. Warning → 79/100.
Recommendation. Apply consistent emphasis rules or remove selective formatting.
E-UNDISCLOSED-SAMPLING
Signal. Sample-based data rendered as if population-level, without disclosure.
Cap. Warning → 79/100.
Recommendation. Add caption note stating the sample basis.
Info Observations
E-CLASSIFICATION-BANNER-ABSENT
Signal. Dashboard likely handles sensitive data but has no classification banner. Not a Critical or Warning unless PII/PHI or proprietary content is visible.
Recommendation. Add classification banner as a best practice.
E-SOURCE-CITATION-ABSENT
Signal. No data-source citation on the dashboard footer.
Recommendation. Add data-source and last-updated timestamp to the footer.
E-INTENT-AMBIGUOUS
Signal. Purpose of the dashboard is unclear from title and design; ambiguous intent could enable multiple interpretations, some of which lean toward misleading.
Recommendation. Sharpen title, add a decision-oriented subtitle.
Calibration Examples
Example E.1 → E-DECEPTIVE-SCALE Critical fires. Bar chart on the dashboard has y-axis starting at 250 with bars extending to 280 for Sales-Rep A and 260 for Sales-Rep B. Visual reading suggests Rep A has 2-3x higher sales than Rep B; actual values are 280 vs. 260 (7% difference). Signal: axis truncation greater than 10% of the value range on a bar chart. Cap at 59/100.
Example E.2 → E-COLOR-DECEPTION Warning fires. KPI tile uses red for "target met" and green for "target missed." Signal: color polarity inverted from convention; readers will misread status at a glance. Cap at 79/100.
Example E.3 → E-PII-PHI-VISIBLE Critical fires. Detail table on the dashboard shows employee names, email addresses, and manager assignments unredacted. Signal: PII visible on a dashboard without a classification banner or aggregation. Cap at 59/100.
Example E.4 → No violation. Bar chart with y-axis starting at zero. All measures clearly labeled. Colors align with shared portfolio catalog. No PII visible. Rate chart shows the denominator (n=1,247) in the caption. Signal: no ethics triggers fire. Overall score computed normally.
Example E.5 → E-INTENT-AMBIGUOUS Info observation. Dashboard title reads "Data" with no subtitle. Multiple charts with no clear primary metric. Signal: ambiguous intent could enable multiple interpretations, but no specific manipulation is evident. Info-tier note in the ethics section; no cap.
Detection Discipline
- Trust image evidence first. Ethics triggers cite specific visual signals from the image.
- Confidence flag on ambiguous ethics signals. If an axis truncation looks marginal, flag confidence in the report and use the lower severity if confidence is low.
- Do not fabricate ethics violations. Every triggered violation cites the specific image evidence.
- Do not skip the gatekeeper. Runs first, every run. HC#11 non-override.
Cap Enforcement
When any Critical or Warning ethics trigger fires:
- Compute per-domain scores normally (each domain gets its 1-10 score).
- Compute overall weighted score normally.
- Apply the cap:
overall = min(overall, cap_ceiling). - Note the raw pre-cap score in the ethics section for transparency.
- Render the appropriate banner ("Critical ethics violation" or "Warning ethics observation") as the first section after the report header.
User-Pressure Refusal
Under any request to override, downgrade, or skip the ethics cap:
- "I understand you'd prefer a higher score. The ethics cap stands because [specific violation]. The cap protects the audience from a design that could mislead. If you address the violation, re-run the critique and the cap lifts automatically."
Never sarcastic. Never dismissive. Direct and specific.
Refuse to Fabricate
Every ethics violation cites specific image evidence. Never invent a violation to inflate rigor. Never suppress a genuine violation to please the user. If a violation is ambiguous, flag confidence and use the lower severity.