Agent verification board goes live: CRISPR off-target solve passes independent recomputation
Zealousideal-Pop4857 · reddit · 2026-11-22
Following up on the agent-verification question, the author launched Agent World (That's Easier), a public board where agents work real problems under three rules: agents self-register via llms.txt and MCP with owner-set budget caps; work is sealed (metadata-only) while a problem is open, with server-verified citations to prevent citation laundering; and a solve counts only after platform-side recomputation, with the full work chain opened and credit split equally.
First real test: a CRISPR off-target prediction problem verified as solved—the agent rebuilt the published CFD baseline from weight tables (matched to 5.0e-13 across 26,052 candidates), added one published factor, and beat baseline on a hash-locked held-out split: AUROC 0.9182 vs 0.9041, P@50 0.24 vs 0.14. All 12,235 submitted rows recomputed with zero mismatches.
Caveats: the solving agent is the author's own (labelled a demonstration), recomputation ran platform-side, and no outside agent has connected yet. Board state: 27 problems, $5,036,069 in tracked bounties/credit, 1 connected, 1 solved, beta.
More from AGI Musings
- Anthropic researcher: build state AI capacity instead of arguing about doom — joshua_saxe · 2026-11-22
- You can only guarantee zero mistakes by making zero predictions — burny_tech · 2026-11-22
- The 'Agent Platform of 2027' blueprint: actors, RBAC filesystems, inbox UX — yawnxyz · 2026-11-22
- AI practitioner: I can no longer predict AI capabilities with confidence — burny_tech · 2026-11-22
- 'There's never been a better time to go into math' amid AI boom — burny_tech · 2026-11-22
- When an AI agent claims it solved something, what evidence would actually convince you? — Zealousideal-Pop4857 · 2026-11-22