Dashboards as Decision Surfaces
A dashboard is not a wall of charts. It is a recurring decision surface: a compact view that tells a named owner whether the evidence is ready, what changed, where to diagnose it, and which reversible action is safe. A metric with no possible decision belongs in exploratory analysis or a reference report, not in the primary decision view.
Prerequisites
Complete attribution and its causal limits. Keep its distinction between descriptive credit and causal effect when a dashboard compares periods.
Learning goals
- Start from a decision and work backward to evidence.
- Show metric version, population, time window, close status, uncertainty, and data quality beside the value.
- Assign one accountable owner without hiding supporting teams.
- Define both a safe action and an explicit no-action condition.
Decision contract
- Decision: Continue, hold, or revise the staged report-export rollout.
- Owner:
core-product; data quality and economics have separate named owners. - Threshold rationale: Investigate an absolute adoption change of at least 0.20 because that is the predeclared triage sensitivity for this small fixture, not because it proves practical importance or impact.
- Diagnostic path: Validate schema, consent, duplicate handling, and close status; compare first close with restatement; then inspect support and privacy guardrails.
- Safe action: Hold automatic expansion while evidence is unresolved. Preserve the current cohort; do not roll back solely because one descriptive metric moved.
- No-action condition: Keep observing when the threshold is not crossed, the reporting window is incomplete, or a metric-neutral duplicate fully explains the alert.
Design from the decision backward
Put the decision, due date, owner, and metric version before any trend. Follow them with a small comparison table whose row and column headers are explicit in markup. Label first-close values as provisional, preserve restated values, and expose counts with rates so a stable percentage cannot hide a collapsing denominator.
The diagnostic path should be shallow. A reader should be able to move from the headline to data-quality status, denominator changes, lifecycle stage, economic assumptions, and guardrails without guessing. Color may reinforce status, but words such as “provisional,” “pass,” and “follow-up required” must carry the meaning.
Worked example
import json
from pathlib import Path
contract = json.loads(Path("tests/fixtures/product-analytics/product-review-contract-v1.json").read_text())
dashboard = contract["dashboard"]
print(f"decision={dashboard['decision']} owner={dashboard['owner']}")
print(f"threshold=adoption delta >= {contract['alerts']['adoption_absolute_change']:.2f} safe_action=hold expansion no_action=keep observing")
Expected output:
decision=continue staged report-export rollout owner=core-product
threshold=adoption delta >= 0.20 safe_action=hold expansion no_action=keep observing
Limitations and common mistakes
- A dashboard refresh does not make source data complete or correct.
- A threshold is an attention rule, not a null-hypothesis test, causal estimand, or universal business boundary.
- Aggregates can hide subgroup harm, changing eligibility, and identity drift.
- More panels increase search cost; they do not substitute for a declared diagnostic path.
- A dashboard cannot resolve a decision whose owner, alternatives, or guardrails remain unspecified.
Practice
Take an existing chart and write its decision, owner, threshold rationale, diagnostic path, safe action, and no-action condition. Remove it from the primary view if no credible action changes across its plausible range.
Next steps
Metric alerts and anomaly triage turns the same contract into a low-noise investigation workflow.
Primary sources
Checked 2026-10-19: Google SRE, Monitoring Distributed Systems; W3C WAI, Tables Tutorial; UK Analysis Function, Performance measurement.