Brings README.md to launch quality and full parity with what the app actually does. README.md only — every claim verified against the code.

What changed

• Top-of-README pitch — scannable one-paragraph pitch, a short feature-bullet list, and a live-demo CTA so a social visitor gets it in ~10 seconds. Keeps the honest "open-source re-implementation, not affiliated with Google" framing and Apache-2.0.
• Full workflow documented — the README previously undersold the engine. "What it does" now walks the real loop: generate → reflect → Elo tournament → multi-round evolution → recurring self-critique → stress-test the top 3 → apply fixes + re-rank → meta-review. The most important stage, adversarial stress-testing of the top 3 (contradicting-evidence search, per-citation verification, feasibility, prototype-scale experiment), is featured.
• HIGH RISK mode — documented (composer toggle + [run] high_risk).
• Architecture diagram — redrawn from the 6-agent linear box to the full loop, ASCII, consistent with the existing style.
• Keyless literature — names openalex_search / pubmed_search / arxiv_search / europe_pmc_search (all keyless, always on) with per-citation verification; web_search only with a Tavily/Brave key.
• Routing — updated to the new landing-at-/, chat-composer-at-/chat flow ("Launch the demo"), matching the sibling routing change. GitHub Pages base stays /Co-Scientist/.
• Citation verified — the Nature 2026 Gottweis DOI (10.1038/s41586-026-10644-y) confirmed via Crossref: "Accelerating scientific discovery with Co-Scientist," Juraj Gottweis et al., Nature, 2026-05-19. Added the verified title.

Verification

• Confirmed every command/path referenced still exists: pip install -e ".[dev]", python -m webapp.server --port, cd frontend && npm run dev, docker compose up, CLI run/serve/report (with --n), config/default.toml, frontend/.env.example, deploy/oracle/README.md, .github/workflows/deploy-pages.yml, LICENSE, the /s/<id>/site microsite route and "View as website" button.
• RUN_PRESETS numbers (Quick 5 / Standard 15 / Deep 50; budgets 1M/2M/5M) match frontend/src/types.ts — kept as-is.
• HIGH RISK toggle confirmed in frontend/src/pages/Chat.tsx and [run] high_risk in config/default.toml.
• Stress-test stage confirmed in co_scientist/agents/supervisor.py (_run_stress_stage) and stress_test_top_k = 3.
• README references no docs/superpowers path, so the sibling docs/design/ rename needs no link change here.

No invented capabilities — every claim maps to real code.

🤖 Generated with Claude Code