{ "$schema": "./dashboard.schema.json", "schema_version": "1.0.0", "competition_name": "ARC Prize 2026 / ARC-AGI-2", "updated": "2026-07-31T05:27:59Z", "system": { "status": "degraded", "status_label": "自动化运行中 · 发布受阻", "mode": "多 Agent 自动化 · 只读监控", "goal": "整合候选管线、SFT 产物与发布安全链。", "scope_note": "本地开发证据与 Kaggle 榜单严格分开；当前尚未提交，没有线上 LB 成绩。", "producer": "multi-agent public snapshot", "refresh_seconds": 60 }, "best": { "cv_score": 0.3903345724907063, "cv_correct_outputs": 420, "cv_total_outputs": 1076, "cv_solved_tasks": 404, "cv_total_tasks": 1000, "score_scope": "local-development-cv", "lb_direction": "maximize" }, "submission": { "status": "not_submitted", "label": "未提交", "message": "无线上成绩；本地开发 CV 不代表 Kaggle LB。" }, "agents": [ { "id": "orchestrator", "name": "Orchestrator", "avatar": "OR", "role": "协调与交付", "status": "running", "task": "整合候选管线、SFT 产物与发布安全链", "progress_label": "持续协调", "message": "汇总各工作流证据并维护发布边界。", "heartbeat_at": "2026-07-31T05:27:59Z" }, { "id": "sft-trainer", "name": "SFT Trainer", "avatar": "SF", "role": "正式 QLoRA 训练", "status": "completed", "task": "完成 Qwen3-8B 正式 SFT 与 adapter 完整性校验", "progress": 100, "progress_label": "181 / 181 steps", "message": "1 epoch · train loss 0.2582763 · final eval loss 0.2340074 · formal manifest static validation valid · 166.6 MiB", "heartbeat_at": "2026-07-31T05:17:06Z" }, { "id": "adapter-comparator", "name": "Base-vs-Adapter Evaluator", "avatar": "BA", "role": "模型独立对照", "status": "waiting", "task": "在固定同批 10 题上运行本地 model-only base / adapter 对照", "progress_label": "等待源码冻结", "message": "等待安全修复后冻结源码并用新目录重跑；中断目录不计为结果。", "heartbeat_at": "2026-07-31T05:27:59Z" }, { "id": "candidate-evaluator", "name": "Candidate Evaluator", "avatar": "CV", "role": "候选池评测", "status": "completed", "task": "完成 1000 题本地开发 CV 与候选池分析", "progress": 100, "progress_label": "1000 / 1000 tasks", "message": "420 / 1076 outputs · 404 / 1000 task full passes · pass@2 +2 · 913.1s", "heartbeat_at": "2026-07-31T05:17:06Z" }, { "id": "safety-auditor", "name": "Safety Auditor", "avatar": "SA", "role": "发布安全审计", "status": "completed", "task": "审计 sealed gate、retrieval provenance 与 bundle 绑定", "progress": 100, "progress_label": "审计完成", "message": "确认 2 项开放的高严重度发布阻断。", "heartbeat_at": "2026-07-31T05:17:06Z" }, { "id": "model-fetcher", "name": "Model Fetcher", "avatar": "MF", "role": "模型资产同步", "status": "running", "task": "下载第 1 / 6 个候选模型：Qwen3-8B Q4_K_M", "progress": 58.554057, "progress_label": "2,943,971,231 / 5,027,783,488 bytes", "message": "仅使用 7897 下载通道，限速 2 MiB/s。", "heartbeat_at": "2026-07-31T05:27:59Z" }, { "id": "monitor-publisher", "name": "Monitor Publisher", "avatar": "MP", "role": "只读状态发布", "status": "running", "task": "构建并验证多 Agent 公开监控面板", "progress_label": "页面已构建 · 发布验收中", "message": "只发布公开状态，不暴露密钥、token 或本机路径。", "heartbeat_at": "2026-07-31T05:27:59Z" }, { "id": "release-guard", "name": "Release Guard", "avatar": "RG", "role": "发布门禁", "status": "blocked", "task": "核验 sealed gate 与 Kaggle bundle 的发布条件", "progress_label": "等待审计修复", "message": "需关闭 2 项发布阻断并重建 bundle 后才能继续。", "heartbeat_at": "2026-07-31T05:27:59Z" } ], "workstreams": [ { "id": "program-orchestration", "title": "自动化整合", "owner_agent_id": "orchestrator", "status": "running", "description": "协调训练、评测、审计、资产与发布证据", "progress_label": "持续协调", "dependencies": [ "各工作流公开心跳" ], "next_step": "整合审计修复与最终发布状态" }, { "id": "formal-sft", "title": "正式 SFT / QLoRA", "owner_agent_id": "sft-trainer", "status": "completed", "description": "正式 adapter 已生成并通过完整性校验", "progress": 100, "progress_label": "181 / 181 steps", "dependencies": [ "冻结的 Qwen3-8B 快照", "无泄漏 SFT 数据" ], "next_step": "进入 adapter 对照与 sealed gate 链" }, { "id": "base-adapter-comparison", "title": "Base vs Adapter 对照", "owner_agent_id": "adapter-comparator", "status": "waiting", "description": "固定同批 10 题的本地 model-only 独立对照", "progress_label": "等待源码冻结", "dependencies": [ "完整性 valid 的正式 adapter", "固定 validation 批次" ], "next_step": "安全修复完成后使用新目录重跑完整对照" }, { "id": "candidate-cv1000", "title": "候选池 CV1000", "owner_agent_id": "candidate-evaluator", "status": "completed", "description": "本地开发 CV 与 pass@2 / oracle 分析已完成", "progress": 100, "progress_label": "1000 / 1000 tasks", "dependencies": [ "候选池 v2", "本地开发 split" ], "next_step": "保留 420 / 1076 为可复现本地基线" }, { "id": "release-safety-audit", "title": "发布安全审计", "owner_agent_id": "safety-auditor", "status": "completed", "description": "sealed retrieval 与 gate / bundle 证据链审计完成", "progress": 100, "progress_label": "2 findings open", "dependencies": [ "sealed gate 配置", "runtime source 清单" ], "next_step": "修复两项阻断并重新审计" }, { "id": "candidate-model-fetch", "title": "候选模型资产", "owner_agent_id": "model-fetcher", "status": "running", "description": "顺序同步候选模型并保留可续传状态", "progress": 58.554057, "progress_label": "第 1 / 6 个 · 58.55%", "dependencies": [ "7897 下载通道", "2 MiB/s 限速" ], "next_step": "完成 Qwen3-8B Q4_K_M 后继续剩余候选" }, { "id": "public-monitor", "title": "公开监控面板", "owner_agent_id": "monitor-publisher", "status": "running", "description": "只读展示 Agent、任务、心跳、阻断与事件", "progress_label": "页面已构建 · 发布验收中", "dependencies": [ "公开 JSON 快照", "版本化 schema" ], "next_step": "完成静态 smoke 并交付发布" }, { "id": "release-readiness", "title": "Kaggle 发布门禁", "owner_agent_id": "release-guard", "status": "blocked", "description": "正式发布条件尚未闭合", "progress_label": "阻断中", "dependencies": [ "审计修复", "bundle 重建" ], "blocked_by": [ "sealed-retrieval-provenance", "gate-bundle-source-binding" ], "next_step": "关闭 findings、重建 bundle 并复验 source hash" } ], "findings": [ { "id": "sealed-retrieval-provenance", "severity": "critical", "status": "open", "title": "Sealed retrieval provenance 未闭合", "description": "Sealed gate 尚未完整冻结 retrieval 资产来源与同题排除证据，存在 self-match 证据链风险。", "action": "绑定可信训练资产 provenance，补强同题排除，并将配置纳入 gate 证据。", "affects": [ "sealed gate", "release" ] }, { "id": "gate-bundle-source-binding", "severity": "high", "status": "open", "title": "Gate 与 bundle source 未完整交叉绑定", "description": "Gate 与 bundle 尚未对全部 runtime source 做 hash 逐项交叉验证，且现有 bundle 已过期。", "action": "源码冻结后重建 bundle，并强制 gate 与 bundle runtime source hash 一致。", "affects": [ "bundle", "release" ] } ], "events": [ { "id": "base-adapter-comparison-deferred", "type": "warning", "title": "Base vs Adapter 对照已暂缓", "description": "为避免依赖源码版本不一致，早期运行已停止；中断目录不计为结果。" }, { "id": "base-adapter-comparison-started", "type": "info", "title": "Base vs Adapter 对照已启动", "description": "固定同批 10 题的本地 model-only 对照曾进入 base 阶段，随后因源码冻结要求停止。" }, { "id": "monitor-delegated", "type": "delegation", "title": "监控面板已委派", "description": "Monitor Publisher 开始构建公开只读状态页面。" }, { "id": "safety-audit-completed", "type": "audit", "title": "发布安全审计完成", "description": "审计确认 2 项开放的高严重度发布阻断。" }, { "id": "cv1000-completed", "type": "success", "title": "候选池 CV1000 完成", "description": "1000 题运行完成；pass@2 为 420 / 1076，oracle pool 为 421 / 1076。" }, { "id": "formal-sft-completed", "type": "success", "title": "正式 SFT 完成", "description": "181 / 181 steps 完成，formal adapter 完整性状态为 valid。" } ] }