BRANCH EVIDENCE · UPDATED 01 NOV 2026 / 最新分支证据
Original Chinese packages from ry-rcua-tasks. English descriptions explain those versions. Screenshot replays reconstruct elapsed time; they are not continuous screen recordings. Samples are not aggregate model success rates.
来自 ry-rcua-tasks 的原始中文任务。英文仅为说明。回放由截图按时序重建，不是连续录屏；单局样本不能作为模型整体成功率。
Four implemented Chinese task packages have been imported from ry-rcua-tasks@cb92943. The three livestream tasks require separately supplied source video; their source, setup instructions and branch model evidence are retained. Their runs do not extend the fixed 17-task leaderboard.
已导入4个中文任务实现。3个直播任务需要本地提供原始视频；源码、安装说明及分支模型证据均已保留。其运行结果不计入固定17题榜单。
Three tiers, 90 frozen episodes, a 54-section venue map, blockers, competitor inventory and a fee-inclusive budget trap. The current package passes 126 contract/harness tests. These are individual samples at the easy t1 tier, not a three-tier evaluation.
三档难度、90个冻结episode、54分区场馆图、流程阻断、竞争库存与含费预算陷阱。当前包通过126项契约与harness测试。下方为t1宽松档单局样本，不是三档完整评测。
This package is retained for the fixed 17-task replay cohort; the source branch has superseded it with newer tracks. The stream advances through inference. Respond to real requests and ignore unrelated chatter. Model pilot evidence for five episodes is retained. Runtime-seeded episodes are not yet a frozen comparable benchmark; prompt variants and model protocols must be read with each result.
此包为固定17题回放保留的历史版本，上游分支已由新赛道替代。直播流在推理期间持续推进；及时响应真正的请求，对无关闲聊保持克制。保留5局模型试跑证据，运行时种子生成的episode尚未构成冻结可比基准；请结合每项结果的提示版本和模型协议阅读。
All three packages have an engine, trusted journal-replay grader and contract tests. They are actively being deepened. A matched GPT / Haiku single-episode pilot for each task is now available on the model leaderboard. Difficulty tiers, frozen episodes and evaluation harnesses remain incomplete. Recorded hand-driven scores are functional evidence, not difficulty or model baselines. Meeting is captions-only, without audio.
三者均有引擎、可信日志重放评分器与契约测试，仍在深化中。现已收录每题各一次GPT与Haiku对照试跑，见模型榜单。难度分档、冻结episode与评测harness待完善。手动运行分数仅为功能证据，不代表难度或模型基线；会议任务仅字幕，无音频。
Native Blender rocket: tutorial, rubric and four model pilots / Blender火箭教程、评分与四次试跑 ↗
All model evidence belongs to the exact version, language, budget and harness it records. Real OSWorld VM setup/evaluate remains unverified. / 所有模型证据仅对应各自版本、语言、预算和harness，真实OSWorld VM接入尚未验证。