MaxGhenis · Sep 30, 2026 · Sep 30, 2026 · Sep 30, 2026 · Sep 30, 2026 · Sep 30, 2026
diff --git a/bundles/analysis-tools.json b/bundles/analysis-tools.json
@@ -2,7 +2,7 @@
   "name": "analysis-tools",
   "description": "Policy analysis and research - impact studies, dashboards, notebooks, and visualizations",
   "category": "analysis",
-  "version": "4.0.2",
+  "version": "4.0.3",
   "keywords": [
     "analysis",
     "research",

diff --git a/bundles/api-development.json b/bundles/api-development.json
@@ -2,7 +2,7 @@
   "name": "api-development",
   "description": "API development - Flask endpoints, caching, services, and integration patterns",
   "category": "development",
-  "version": "4.0.2",
+  "version": "4.0.3",
   "keywords": [
     "api",
     "flask",

diff --git a/bundles/app-development.json b/bundles/app-development.json
@@ -2,7 +2,7 @@
   "name": "app-development",
   "description": "React app development - app-v2, standalone tools, Vercel deployment, and PolicyEngine-specific patterns",
   "category": "development",
-  "version": "4.0.2",
+  "version": "4.0.3",
   "keywords": [
     "react",
     "frontend",

diff --git a/bundles/complete.json b/bundles/complete.json
@@ -2,7 +2,7 @@
   "name": "complete",
   "description": "Complete PolicyEngine knowledge base - all agents, commands, and skills for users, analysts, and contributors",
   "category": "complete",
-  "version": "4.0.2",
+  "version": "4.0.3",
   "keywords": [
     "complete",
     "all",

diff --git a/bundles/country-models.json b/bundles/country-models.json
@@ -2,7 +2,7 @@
   "name": "country-models",
   "description": "Complete multi-agent workflow for implementing government benefit programs in PolicyEngine country packages",
   "category": "development",
-  "version": "4.0.2",
+  "version": "4.0.3",
   "keywords": [
     "tax",
     "benefits",

diff --git a/bundles/dashboard-builder.json b/bundles/dashboard-builder.json
@@ -2,7 +2,7 @@
   "name": "dashboard-builder",
   "description": "Orchestrated AI workflow for creating PolicyEngine dashboards from natural-language descriptions",
   "category": "development",
-  "version": "4.0.2",
+  "version": "4.0.3",
   "keywords": [
     "dashboard",
     "calculator",

diff --git a/bundles/data-science.json b/bundles/data-science.json
@@ -2,7 +2,7 @@
   "name": "data-science",
   "description": "Data analysis, survey enhancement, imputation, calibration, and microdata utilities",
   "category": "data",
-  "version": "4.0.2",
+  "version": "4.0.3",
   "keywords": [
     "data",
     "microdata",

diff --git a/bundles/essential.json b/bundles/essential.json
@@ -2,7 +2,7 @@
   "name": "essential",
   "description": "Essential PolicyEngine knowledge for all users - understanding the platform, using the web app, and basic concepts",
   "category": "getting-started",
-  "version": "4.0.2",
+  "version": "4.0.3",
   "keywords": [
     "beginner",
     "users",

diff --git a/skills/policyengine-design/SKILL.md b/skills/policyengine-design/SKILL.md
@@ -149,8 +149,17 @@ automatically. Do not claim built-in dark mode at this version.
   PolicyEngine-contributed; see the policyengine-tools skill for the version gate and the
   `domain={["auto", "auto"]}` pairing). Conventions: gains `--chart-1`; losses `--chart-5`
   or `--destructive`; neutral `--border`.
-- **Plotly (Python maps/charts)** — house style: white `plot_bgcolor`/`paper_bgcolor`, Inter
-  font, teal series, logo bottom-right. Reference tokens by hex with a CSS-var comment.
+- **Posts on policyengine.org:** draw charts as ```` ```chart ```` blocks, which the website
+  renders with ui-kit's chart defaults, Inter and the PolicyEngine watermark (Recharts for
+  `bar` and `waterfall`, a tile grid for `stateMap`). See policyengine-writing, "Figures,
+  covers and replication in posts", and app-v2's `chart-standards.md` under
+  `docs/engineering/`. The website builds Tailwind with the `tw:` prefix, so ui-kit
+  components that rely on unprefixed utility classes render unstyled there; style blog chart
+  parts inline with the `:root` token variables. Never embed matplotlib PNGs or Plotly JSON
+  in new posts.
+- **Plotly (Python notebooks and standalone outputs, not posts)** — house style: white
+  `plot_bgcolor`/`paper_bgcolor`, Inter font, teal series, logo bottom-right. Reference
+  tokens by hex with a CSS-var comment.
   Sketch:
   ```python
   def format_fig(fig):

diff --git a/skills/policyengine-writing/SKILL.md b/skills/policyengine-writing/SKILL.md
@@ -445,6 +445,80 @@ The reform dramatically slashes child poverty
 This is Montana's largest income tax cut in history
 ```
 
+## Figures, covers and replication in posts
+
+**Charts.** Draw every chart in a policyengine.org post as a ```` ```chart ```` block: a
+JSON spec that the website renders with ui-kit's chart defaults, Inter and the PolicyEngine
+watermark. `bar` and `waterfall` draw with Recharts; `stateMap` draws a US tile grid. In
+policyengine-app-v2, the header comments of `BlogChart.tsx`, `BlogWaterfallChart.tsx` and
+`BlogStateMap.tsx` (under `website/src/components/blog/`) document each spec, and the
+engineering chart standards (`chart-standards.md` under `docs/engineering/`) cover the
+details. Do not put matplotlib PNGs or Plotly JSON in new posts; when a post needs a chart
+type the website lacks, add the type with tests first. Keep each chart's numbers identical
+to the table or text beside it.
+
+**Charts elsewhere.** For a newsletter or social image, screenshot the chart as the live post
+renders it (headless Chromium at 2x). A re-plot drifts from the post's labels, source line
+and styling.
+
+**Covers.** Give research posts a photographic cover: a still life of relevant objects on a
+light wooden table in soft natural light, with PolicyEngine teal (`#319795`, `#285E61`) only
+in the objects, and nothing legible or recognizable: no text, numbers, currency, logos,
+charts, portraits or faces. Generate it with an image model, inspect every object for stray lettering, and save
+it as an 1800×840 WebP beside the post's other assets. Never use a chart or an infographic
+as a cover. A product or tool launch may use a 1200×630 typographic card with no data
+graphic instead.
+
+Prompt adapted from the one used for the graded 2025 SPM prediction post (swap the objects
+for the post's subject):
+
+> Photorealistic editorial still life, high three-quarter angle, on a light oak table beside
+> a window, soft diffuse morning light from the left, shallow depth of field. Objects grouped
+> on the right two-thirds, leaving calm empty wood on the left: [objects]. Muted natural
+> palette of warm wood, white paper and soft gray, with teal (#319795 and #285E61) appearing
+> only in [two or three objects]. No legible text, numbers, currency, logos, charts or faces
+> anywhere. Wide 2.14:1 landscape (1800×840).
+
+**Replication detail.** Name the model and its version in the body, as in the
+methodological transparency example below. Put other package versions, script paths, commit
+hashes and rerun steps in footnotes, so the body states what the numbers show and the
+footnotes let a reader reproduce them.
+
+**Post dates.** Write `posts.json` dates with a time of day (`"2026-09-29 08:00:00"`). The
+article page and research list parse a bare date as UTC midnight and show it as the previous
+day to readers in US time zones.
+
+## Grading forecasts and explaining misses
+
+When a post grades a registered prediction or explains why the model missed:
+
+- **Grade what was registered.** Compare the registered quantity with the published figure
+  at the precision the source publishes, and give the published margins of error. When the
+  registration builds a level from a modeled change, grade the change as well. When the
+  published change is not statistically significant, say so.
+- **Reruns show which assumptions matter.** A rerun that matches the published figure does
+  not identify the right inputs; say so beside the result.
+- **Every rerun is a full model run** with the changed assumption entered as a model input.
+  Never estimate a rerun by scaling an earlier run's output or by adding effects from
+  separate runs. Name the run or runs behind each figure.
+- **Lead with totals.** Splitting a change into parts (thresholds and resources, or one
+  income source at a time) can give different answers depending on the order of the steps.
+  Decompose each rerun from its own runs, report each order or the average over all orders, and do not
+  assign a remaining gap to one part on the strength of the baseline decomposition.
+- **Compare survey years with care.** The CPS reweights to new population controls, so the
+  weighted count of people in a group, or of benefit recipients, can jump for reasons that
+  have nothing to do with the economy. Compare amounts per recipient, hold the age mix fixed
+  when it shifts, and treat two survey years as separate cross-sections, not a panel.
+- **Show sampling error before claiming a pattern.** A gradient across deciles or groups
+  needs a bootstrap or replicate-weight standard error. When the pattern sits within its
+  error, present a rerun built on it as a sensitivity check.
+- **Name each rate exactly**: "CPI-U, annual average" or "January over January"; "per
+  recipient aged 66 and over, at the 2024 age mix". Use the publisher's label for a data
+  series; change it only after checking the primary source.
+- **Review before publishing.** Have an independent reviewer check every number and
+  mechanism claim against the result files and the code, and review again after each round
+  of fixes.
+
 ## Structure and Organization
 
 Follow a clear hierarchical structure with key findings up front.
@@ -747,6 +821,10 @@ Before publishing, verify:
 - [ ] Introduce finished work fresh; keep development history in supporting records
 - [ ] Remove self-referential clauses, empty contrasts, and unintroduced names or roles
 - [ ] Check mechanism claims against the code, test, or primary source
+- [ ] Draw charts as ```` ```chart ```` blocks; no chart PNGs or Plotly JSON in new posts
+- [ ] Use a photographic cover with nothing legible, never a chart
+- [ ] Name the model and its version in the body; put other package versions, script paths and rerun steps in footnotes
+- [ ] For graded forecasts: totals first, order-dependent splits labeled, sampling error shown for any pattern
 
 ## Resources
 

diff --git a/targets/claude/marketplace.template.json b/targets/claude/marketplace.template.json
@@ -1,7 +1,7 @@
 {
   "$schema": "https://claude.ai/schemas/plugin-marketplace.json",
   "name": "policyengine-claude",
-  "version": "4.0.2",
+  "version": "4.0.3",
   "description": "Complete PolicyEngine knowledge base - for users, analysts, and contributors across the entire PolicyEngine ecosystem",
   "owner": {
     "name": "PolicyEngine",