Act on the structural review, and show the product

[vahid-ahmadi]

Acts on a fresh structural review, plus adds the product screenshot. Not for merging yet — the structural review also raised five judgement calls that are deliberately not in here, listed at the bottom.

The finding that drove most of this

.narrative-step { min-height: 90vh } across six stages made "Under the hood" five to six viewport heights — roughly a quarter of the article by scroll — for content that fills about a third of a screen per stage. The sticky diagram it drives changes only which of six boxes is highlighted.

This inverted the assumption that the interactives were the weight problem. They are not: the tool explorer, the call trace, and the question picker are each bounded to about one screen by progressive disclosure. The tool explorer authors 21 tools and costs one panel of scroll. Reduced the minimum to 58vh, recovering about two screens with the diagram still tracking.

Seven changes, all flagged high confidence

1. Deleted "The language model proposes a plan". The three-part architecture — language model, gateway, tools — was stated four times: the introduction, the close of "The boundary", this section, then again through the stage segment tags. This was the weakest instance, sitting between the section that had just made the point and the tool explorer the reader was trying to reach. Its one non-duplicative sentence, that the gateway may ask for clarification, is already covered by the Gate stage's needs plan outcome.
2. Deleted "What's next", folding its roadmap sentence into the Limitations opener — which already said reform coverage is incomplete and gateway behaviour will move — and its ecosystem sentence into "Try it yourself". The article now closes on two beats, caveats then the ask, instead of three.
3. Cut the paragraph narrating the trace's simulation step. It read aloud exactly what the component shows with a tick or cross against each field. Kept the one claim not in the component: that the analysis tools are the only route from a stored simulation to a number.
4. Cut the worked example's closing clarification paragraph. It ended the payoff section on a hypothetical caveat immediately after the £0.9bn result, and the Gate stage already owns the clarification path.
5. Moved the computed result directly under the trace, so five calls resolve into their answer with nothing between them. It was sitting at about 90% depth behind two paragraphs that narrated the component above it; the commentary now follows a result the reader already has.
6. Stated the request once. It appeared in stage one's example and then again as a fresh "Consider: …" a section later, which read as a restart. It now opens "Under the hood" as the shared premise for both sections, and the worked example opens as a continuation — including an explicit note that the loop re-derives what the gateway approved, and that the match is the check.
7. Reduced .narrative-step min-height to 58vh.

Nine H2s, down from eleven.

Also: the product screenshot

Between the beta note and the call to action, linking into the chat. Two separate reviews observed that a launch article for a chat interface showed no chat interface — no rendered answer, nothing a user would see. This is the real capture, fitted to 1700px and 106 kB.

Deliberately not included — the reviewer's judgement calls

Held for a decision rather than applied:

• Move "What you can ask" from position 7 to position 5. It is the only section telling a prospective user whether the tool is for them, and it currently arrives near the end. Strong lean from the reviewer, but it would need its opening rewritten and would put tool names in front of the reader before the tools are introduced.
• Swap "Separating calculation" and "How policy analysis is done today", so the problem claim sits next to its evidence. The current placement is deliberate.
• Merge scrolly stages 2–4 into one "Gateway" stage, so the deep section confirms the three-part frame rather than complicating it with six. Costs one of the three stage examples.
• Surface the 80.3% PolicyBench figure in the lede.
• Add a section showing a rendered answer, beyond the screenshot.

bun run check passes, and content/current-article.md is in step.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
policyengine-uk-chat-article	Ready	Preview	Aug 14, 2026 10:59am

[anth-volk]

Thanks, Vahid. PR #10 now integrates many of the structural changes developed here, including removing the redundant “The language model proposes a plan” section, folding the roadmap material into Limitations and Try it yourself, removing redundant worked-example and clarification copy, and tightening the beta framing.

PR #10 takes the reorganisation further by replacing the separate static stage walkthrough and call-by-call section with one six-stage interactive reconstruction of UK Chat. It carries the catalogue evidence, reform validation, public tool sequence, and final answer within their corresponding stages, so the same Child Benefit example is no longer explained twice.

I’m closing this PR in favour of #10. Its remaining differences—such as the standalone product screenshot and shorter legacy scrollytelling layout—have not been copied wholesale because #10 uses the interactive chat reproduction instead.