v2.28.0 — LLM-as-a-Judge QA + 85-Round Flywheel

What's New

LLM-as-a-Judge QA Scoring System

Replaced 120+ regex false-positive patterns with Gemini 2.0 Flash semantic classification. The judge classifies QA issues into 4 categories:

• genuine_bug — real functional/visual defects
• design_opinion — subjective design preferences, not bugs
• screenshot_artifact — compression artifacts misread as issues
• mock_data — placeholder content flagged as incomplete

Expanded Deterministic Metrics (Layer 1)

7 → 12 boolean Playwright checks:

• no_layout_shift (CLS post-hydration, Google 0.25 threshold)
• no_404_resources (static asset 404s)
• no_slow_resources (>5s load time)
• no_mixed_content (HTTP on HTTPS)
• viewport_meta_ok (viewport meta validation)

3-Layer Scoring Rebalance

• Layer 1 (60%): 12 deterministic boolean metrics — zero variance
• Layer 2 (30%): Severity rubric on LLM-judged genuine issues only
• Layer 3 (10%): Legacy P-level taste deductions

QA Score Journey: R50→R85

R50: 56/100 (regex-only era, 120+ patterns)
R78: 85/100 (first LLM judge run)
R82: 82/100 (Layer 1 permanent failures fixed)
R84: 97/100 (judge tuning converging)
R85: 100/100 ✅ (perfect score — all 3 layers maxed)

148-Screenshot Gallery

Full responsive matrix: dark desktop, light desktop, dark mobile, light mobile — covering all 37 routes.

UI Polish (200+ components)

• Motion-safe animations (111 files)
• Dark mode gap fixes (research, analytics, narrative, onboarding)
• Jargon purge (ISO dates, label consistency)
• Accessibility (focus rings, ARIA labels)

Packages

• nodebench-mcp@2.28.0 — npm
• @homenshum/convex-mcp-nodebench@0.10.1 — npm

Key Files

• scripts/ui/runDogfoodGeminiQa.mjs — Complete scoring engine rewrite
• .claude/rules/gemini_qa_loop.md — QA loop integration rule
• public/dogfood/qa-results.json — 85-round score history
• public/dogfood/screenshots/ — 148 screenshots

Full Changelog: v2.27.0...v2.28.0