
K3
The 15 phase-1 tasks were run by Kimi Code CLI 0.42.0 at thinkingEffort `max`, one fresh independent single-turn session per task. Tool-call telemetry is complete (reasoning in clear text), so the runs went through the normal dimension-by-dimension audit (not the no-telemetry default).
Version basis: Version is recorded per run as a snapshot (`environment.reportedModelId`); no single version number is claimed.Selected works
Works first, numbers second.
Archive statistics
K3 · Phase 1
由 15 个 run 的逐题指标求和;每题一个独立单轮会话;任一题缺失的指标记 null,不按 0 补齐。
Phase progress
Model snapshots (harness-reported model id)
The model snapshot this run actually used; different snapshots of one model are not the same configuration.
k3Model id reported by the harness (alias `kimi-code/k3`) and recorded in each phase-01 submission.json; the organizer confirmed the provider is Kimi official, and the provider-side snapshot identifier is not independently confirmed.Phase assessment
93.3/100
AI 评价,非盲评(维护 agent 的 Codex 桌面会话;评委后端快照未提供,故不猜测)
Only one AI phase assessment (93.3/100) is archived for this model, so no average is computed; rubrics differ and are not compared across models.
All works









Run history
One row per archived run; open a row for the full record.
1afdc8cacbb8…