Dashboard review

Harden an already-built data visualization before it is shared. The goal is a dashboard a skeptical, informed viewer can trust at a glance.

Priority order — non-negotiable: accuracy first, clarity second, never reach for drama. If a figure is not accurate and does not contribute to clear messaging, it does not belong — bias toward removal. Pairs with the web-dev skill (which governs how it's built); this skill governs whether what's built is true, clear, and honestly framed.

Operating model

1. Inventory first. Enumerate every card/panel, every figure shown, every claim/headline, and every source link — and flag every figure that has no source link (that gap is itself a finding; see Dimension 4). You can't audit what you haven't listed.
2. Audit across the dimensions below, gathering findings — don't fix yet.
3. Produce a severity-ranked findings doc (see Output), each finding with: location, the problem, why it matters, and a proposed change. Rank by severity (accuracy > clarity), and within that by how badly it misleads.
4. Discuss, then apply on confirmation. Accuracy changes are high-stakes and often judgment calls — present findings and let the user decide. Do not silently rewrite numbers or claims. (Mechanical, unambiguous fixes like a dead link can be batched once confirmed.)
5. Verify after applying — rebuild, re-check that every shown figure still traces to a working source, and re-read the copy.

Scale effort to the ask: a quick "is this clear?" gets the high-confidence findings; "audit this before launch" gets all dimensions, with source figures re-verified against primary sources.

---

The dimensions

1. Primary-source verification

• Trace each figure to a primary source — the original measurement, not an intermediary that compiled or re-stated it. If a fact sheet/article cites a dataset, go to the dataset.
• Confirm the number actually matches the source (open the PDF page, run the query, read the table). Don't trust that a cited number was transcribed correctly.
• Tag each figure with source + year + denominator + measured-vs-modeled. Flag modeled/estimated/"capture-recapture"/self-reported figures as such; never let a modeled split read as a hard count.
• When only an intermediary exists and it's privately circulated or copyrighted, cite it by name but link the primary source (and don't redistribute the intermediary).

2. Denominator & population integrity (the highest-yield check)

• Every funnel stage, ratio, and comparison must use ONE consistent population. Walk each multi-stage visual: is stage 2 a true subset of stage 1, or a different universe?
• Tell: a big top-of-funnel number that's broader than what the lower stages can come from (e.g. "all drug users" → "opioid-Medi-Cal patients on medication"; "everyone in any care" → "the intensive tier"). The drop reads as attrition but is mostly definitional — that's a conflation. Start the funnel at the population the rest actually nests under, or drop the mismatched top entirely.
• A ratio's numerator and denominator must be the same universe. "X% retained" computed across two populations is misleading even when both numbers are real.
• Never sum overlapping attributes (someone can be unsheltered and using and mentally ill). Check nothing visually implies addition.
• If two panels measure genuinely different things, don't force them into the same shape to look symmetric — use the shape each one's data honestly supports (a funnel vs. a single stat).

3. Recency & internal consistency

• Date-stamp every figure. Flag mixed time-bases presented as comparable (e.g. a 2024 count next to a 2026 count, or an annual cohort vs. a point-in-time snapshot).
• Cross-check the same metric everywhere it appears — flag a value that differs between cards, between the dashboard and its spec, or between a figure and a footnote.
• For live/changing data, require an "as of " and name the dataset + id.

4. Source-link coverage & hygiene (verifiability is a core deliverable)

• Every figure a skeptic might doubt exposes its own source link — to the reader, in the published view. A figure with no way to check it is a finding in itself. Default fix: ADD the link, don't cut the figure — the bias-toward-removal default does not apply here. Reviewer-side verification (Dimension 1) is necessary but not sufficient: "I traced it to source" is not the same as "the reader can."
• Gold standard — round-trip verifiability. The link runs a live query that returns the displayed number (click "129,354" → the query returns 129,354). Freeze the exact window/filters into the URL so the count matches and doesn't drift. For derived or estimated figures that no single query reproduces, link to the method that produced them and label them as derived — don't fake an exact-match link.
• Working & resolvable. No local/relative file links that break when published; no links to files you can't redistribute; links open safely (new tab, rel="noopener") and the label says what they source.
• No dangling links (the other direction). If you remove a stat, fix or remove the source label that cited it — and never leave a source link pointing at a figure that isn't shown.

5. Chart-type & encoding honesty

• Match the visual to the data. Pie/stacked charts imply a clean partition — wrong for overlapping attributes. Funnels imply nested subsets — only valid for true subsets.
• No truncated or distorted axes, no 3D/area tricks; encoded length/width should be proportional to the value.

6. Headline–data fidelity

• Each title/label accurately describes what's actually shown. Flag a header that promises something the panel doesn't deliver (e.g. a title saying "how common" when only an overlap-within-a-subset is shown).
• Each stat's referent is instantly legible — a big number should say what it counts in prominent text, not bury it in a muted caption ("488 / in Full Service Partnerships", "437 / on the shelter waitlist").
• Define acronyms and avoid internal jargon ("persona-lens").

7. Tone — let the data speak for itself (preferred default)

• Read every header, summary/lede, kicker, and label and ask: is this making a dramatic claim, or stating the fact and letting the number carry it? Data-speaking-for-itself is the preferred state.
• Section titles must be descriptive, not editorial — "What we are trying to provide," not "…and how little of it we provide."
• Cut intensifiers and punch-lines — "only," "just," "even," "almost no one," "a tiny fraction," "the system isn't delivering it." State the number and the standard ("an effective course runs ≥12 months; 210 remain at 6") instead of the editorial conclusion.
• Soften advocacy framing in callouts (e.g. "Evidence:" rather than "✔ It works!"). The throughline can still be pointed — it should emerge from the data, not from adjectives.

8. "Earn its place" test (cognitive load)

• For every figure, caveat, and sentence, ask: does this add accuracy AND clarity? If not, cut it. Bias toward removal.
• Kill redundancy — the same explanation stated twice, a number restated across cards, a cross-reference that invites a misleading mental calculation.
• Put each caveat next to the figure it qualifies; don't pool caveats far from their subject.

---

Lighter appendix (check if time/scope allows)

• Number formatting: consistent thousands separators, rounding, and significant figures; ~ for estimates vs. exact counts; consistent units.
• Link/asset health on publish: no dead links, no references to removed files, build serves correctly under its final URL/base path.

---

Output

Write findings to a markdown doc (e.g. <name>-review.md) the user can discuss against:

• A ranked table: # | finding | dimension | severity | status.
• For each finding: location (file:line / card), the problem, why it misleads, proposed change, and a status that updates as decisions are made (Agreed / Resolved / Deferred / Done).
• Record resolved decisions and the reasoning inline as you go (this becomes a durable rationale log — useful when someone later asks "why isn't X on here?").
• Keep the spec / source-of-truth doc in sync with applied changes.

Reminders

• Don't change a number or claim without surfacing it first.
• Re-verify, don't assume — open the source, run the query, rebuild.
• Removing a figure is a valid and often correct outcome. The strongest dashboard is the one where every remaining number is true, sourced, and pulling its weight.