AI Model Benchmarks
Compare AI model success rates across 5 production coding tasks before and after applying Chemical X standards.
Empirical Agent Benchmarks: Monolith vs Chemical X
Pass rates are grounded in published research: Liu et al. 2023 "Lost-in-the-Middle" (Stanford/Berkeley), the Chroma 2025 context-rot study, and SWE-bench Verified leaderboard data for Claude 3.7 Sonnet and GPT-4o. We mapped those findings onto 5 real-world coding agent tasks to show how monolithic vs. atomic context changes agent outcomes.
Task A: URL Param Filter Persistence
Objective: Synchronize category filter bidirectionally with URL search parameters on reload without infinite re-render cycles.
Buried reactive logic mid-context caused >30% accuracy drop per "Lost-in-the-Middle" (Liu et al., 2023). Agent missed the hydration guard entirely.
Surgical patch applied cleanly in isolated molecule under 100 lines. Zero regressions.
npm run audit -- --json
Static AST linter scans every AST node for Line Budget violations (>500 lines), Hook Saturation (>5 hooks/scope), and un-factored ternaries.
GitHub Discussions: Live Audit Category
Real-world codebase audit scorecards and before vs after transformation reports published by engineers using npx chemx audit --share.
Chemical-X-Protocol/starter-kit
@xopherdeepHalloftheGods/elysium
@xopherdeepChemical-X-Protocol/starter-kit
@xopherdeepChemical-X-Protocol/starter-kit
@xopherdeepChemical-X-Protocol/starter-kit
@xopherdeep