[vahid-ahmadi]

Acts on a fresh independent review of the article. Mechanical and clear-cut fixes only — the editorial calls are listed at the bottom, untouched.

Wrong or misleading

• "They fall into one of five types" — plural subject, singular predicate. The Markdown copy had it right; the page had regressed.
• "without switching tools" contradicted the section immediately above it, which is entirely about the model switching between 21 tools. The point is that the reader does not leave the conversation.
• The worked-example chip did not match the request the prose quotes. It omitted "Child Benefit" and the requested output, while stage 1 claimed to extract "budgetary impact" from "the user's exact words" and stage 2 bound gov.hmrc.child_benefit.amount.eldest from wording that was never shown.
• gateway route and full lifecycle rendered as <code> chips styled identically to validate_reform and run_household_simulation, so they read as public tool names. They are not tools; the row is now omitted when a question kind has none.
• Tax revenue rising on a benefit increase reads as a sign error. Now states that the High Income Child Benefit Charge is why.
• content/current-article.md had drifted badly — missing the PolicyBench paragraph, "The boundary", and "What you can ask" entirely, and carrying a different results table — while the README called it "the canonical editorial reference". Regenerated from the page. The README now says the page is canonical, because a hand-synced duplicate will drift again.

Accessibility

• --text-light was #7a9494, 3.2:1 on white, under the 4.5:1 minimum, and it carried the results-table sign-convention caveat plus every card meta line. Darkened at the token, which fixes six call sites at once.
• The six stages were <span>s, so a screen-reader user could not navigate them by heading, and the sections had no accessible name. Now <h3> with aria-labelledby; the CSS already gave .step-title its own size, so nothing moves visually.
• The boundary diagram rendered its labels at about 3.5px on a phone — max-width: 100% prevented .agent-flow-container's overflow-x: auto from ever engaging. A min-width lets it scroll.
• Keyboard focus on a tool chip was styled identically to selection, so two chips read as active while tabbing.
• The question panel's aria-live region re-announced what the tab change had already announced.

Also

Capped the measure at 78ch (running text was ~140 characters a line), added a publication date to a piece whose figures are pinned to a dataset release, dropped the worked-example timeline's hover affordances since it has no click handler, set lang="en-GB", and removed instructional filler ("Explore these categories in the interactive below", "(more on those later)") and meta-commentary about the article's own structure.

Deliberately not changed

• The CTA host still points at policyengine-uk-chat.vercel.app.
• No evaluation of UK Chat itself. PolicyBench measures unaided US models — the premise, not the conclusion — and Limitations does not acknowledge the gap.
• The gate runs on a model-generated plan. A confident misreading with the right parameter path passes cleanly and returns a correct number answering the wrong question. The closing caveat covers ambiguity, not confident misinterpretation.

The reviewer also could not verify the article's claims about the chat service — the 30-round cap, the five gate outcomes, the SSE event types, "no generic agent framework in the request path" — against policyengine-uk-chat, which is a separate repo. Worth checking before publication.

bun run check passes.

[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 12, 2026 1:48pm