Site tier-1: open-set banner, sensitivity selector, bootstrap intervals

[MaxGhenis]

Summary

Implements the tier-1 leaderboard improvements from docs/site_improvements_scope.md, plus a shared sticky header that the /paper page now reuses (matching the home page chrome, without the Global/US/UK view selector).

Header

• Extract SiteHeader from Hero. The new component owns the sticky brand + nav + view-selector + action-link layout and supports an alwaysExpanded mode for pages that don't drive their own collapse.
• Hero refactored to wrap SiteHeader and pass the country-aware subtitle, stat strip, and snapshot pill as expandedContent. Drops the Top score stat and the Leading: <model> sidebar — the leaderboard itself is the canonical source for both.
• /paper uses SiteHeader with alwaysExpanded, no view selector, and a Benchmark action link. The page body keeps its eyebrow/buttons/iframe; the inline H1 is gone since the header carries the brand.

Open-set banner and snapshot pill

• Warning-tinted note above the leaderboard: "Open-set leaderboard. The public scenario explorer exposes prompts and PolicyEngine reference outputs, so future model releases or fine-tunes could learn from the released cases…".
• Snapshot date pill (Snapshot 2026-05-01) in the hero stat row on the home page and next to the Manuscript eyebrow on /paper.

Sensitivity-view selector

• Segmented control above the leaderboard table: Main / Amount only / Binary only / Positive cases / Zero cases. Selecting a view rescores models client-side from scenarioPredictions and reorders the table; the description for the active view appears inline.
• New utilities under app/src/lib/:
  • scoring.ts ports score_single_prediction (mean of exact / within-1% / within-5% / within-10% for amount outputs; classification accuracy for binary; output-group resolution for person-expanded variables). Verified against canonical analysis.py on the snapshot for both US and UK headline scopes.
  • sensitivity.ts builds the per-row score table from a DashboardBundle and aggregates output-group means → country → global, preserving country-equal weighting. Sensitivity views filter rows before aggregation.

Bootstrap rank intervals

• bootstrap.ts implements the household-resampling bootstrap with a deterministic mulberry32 RNG (seed 42, 400 draws). For each model in the active sensitivity view, it reports the 95% score interval and the rank range.
• ModelLeaderboard renders Rank N(-M) · 95% L-U next to each model's point estimate. Sample output for the main global view: Rank 1 has 95% CI 79.8-83.5; Rank 2-3 cluster at 77.4-81.8; the tail spreads Rank 8-11 / Rank 9-11. Tooltip names the bootstrap parameters.

Repo

• Move the Python wheel-artifact lib/ rule in .gitignore to /lib/ and /lib64/ (top-level only) so app/src/lib/ is tracked.

Verification

• bun run lint — clean
• bun run build — clean (Next.js 16 production build)
• bun run start — SSR render of / contains the open-set banner, the snapshot pill, the five sensitivity selector chips, and per-model Rank/95% interval rows for all 12 models. /paper renders SiteHeader with the snapshot pill and Benchmark action link, no view selector.
• Scoring math reconciled against analysis.py for both countries' top-5 models (gpt-5.5: US 89.20 / UK 81.64; grok-4.20: US 88.71; gemini-3.1-pro-preview: UK 79.61, etc.).

Test plan

• CI passes
• Visit / — confirm header drops "Top score" stat and "Leading: GPT-5.5" pill, and the snapshot pill is visible
• Switch sensitivity view to Positive cases — leaderboard reorders and ranks/intervals refresh
• Switch to Zero cases — same
• Visit /paper — confirm the same sticky header style without a view selector
• Mobile width — confirm bootstrap interval line wraps under each model row

Follow-ups (not in this PR)

• Tier-1 per-model deep-dive page (/model/[id])
• Tier-2 cross-country compare and cost surfacing — see docs/site_improvements_scope.md

🤖 Generated with Claude Code

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
policybench	Ready	Preview, Comment	May 6, 2026 10:38am