kapso Files

Results

• 🏆 IOAI² Grand Master Trophy at IOAI 2026: competing fully autonomously in the AI Model Track of the International Olympiad in AI, Kapso surpassed the best human contestant and ranked top 3 among all AI system participants, a field spanning major AI labs and startups. Details: benchmarks/ioai2026/.
• Beats the best foundation model on RelBench: on Stanford's benchmark for predictive ML over enterprise data, Kapso passes KumoRFM-v2 in outcome prediction and forecasting, and the best reported results in recommendations. It also beats a frontier coding agent under an identical budget. Details: benchmarks/relbench/.

Platform

• One task-agnostic platform: each benchmark plugs in as a handler plus a knowledge bank plus config declarations, with a thin harness runner.
• Generic search decomposed into focused modules: ideation, lens planning, expansion lanes, implementation, registered evaluation, feedback flow.
• Handler contract v2: stop authority (honor_agent_stop), insured finalization reserve, run-selection stamping, and a declared tree-search surface.
• Config layering: benchmark modes deep-merge over platform defaults; copy-paste drift eliminated.
• Behavior knobs: implementation_web, ensemble_time_split, per-benchmark gate and web policies.
• Evaluation governance: score-of-record manifests, rescore-based final selection, evaluator integrity fingerprints, non-fatal maintainer triage.
• New coding agents: Codex CLI and Claude Code on OSS endpoints; selector hardening and crashed-lane fallback retry; ensemble ideation with a task-aware lens planner and replanner.

Benchmarks

• benchmarks/ioai2026: the IOAI 2026 AI Model Track pack (task preflight, submit-and-learn lanes, kernel slot management, knowledge bank) plus the past_learning harvest factory.
• benchmarks/relbench: sanitized-cache protocol, provided official evaluator, campaign driver, and a reproducible Claude Code baseline harness.