diff --git a/README.md b/README.md index c7a7b7e..49836d6 100644 --- a/README.md +++ b/README.md @@ -1,88 +1,89 @@ +[简体中文](README.zh-CN.md) · English + # amber-workbuddy -> ⚠️ **更正(2026-10-02,另一项)**:防御轴的一案 A-d511f9e8 在所有车道上改记 NA(考场判的不是考生交付的文件,判分还要求了题面没写的事)。分母不变,**过案数不变**,每条道的总分都带 `'`。本仓各期成绩表里这一格请按 NA 读,其余内容保留原样,以[更正声明](https://github.com/getaskclaw/amber/blob/main/docs/corrections-2026-10-02-a-d511f9e8.md)为准。 +> ⚠️ **Correction (2026-10-02, second)**: one defense-axis case, A-d511f9e8, is now NA on every lane (the exam room did not grade the file the candidate delivered, and the grader asks for something the task text does not say). The denominator and the **number of passed cases do not change**; every lane's total now carries `'`. In this repo's issue tables, read that cell as NA. Everything else stays as published; the [correction notice](https://github.com/getaskclaw/amber/blob/main/docs/corrections-2026-10-02-a-d511f9e8.en.md) governs. -> ⚠️ **更正(2026-10-02)**:以下考卷在作答时越出考卷、接触了判分材料,不计胜负。deepseek-v4.1-flash @ WorkBuddy 直连道 有 3 张卷(A-a5608487、A-984e80ee、A-24bcf707)改记 NA,成绩 16/24 → **13'/24**。原因是考场隔离缺陷,责任在我们。本页其余内容保留原样,以[更正声明](https://github.com/getaskclaw/amber/blob/main/docs/corrections-2026-10-02.md)为准。 +> ⚠️ **Correction (2026-10-02)**: the papers below were answered by a candidate that stepped outside its paper and touched grading material; they count neither as a pass nor as a fail. deepseek-v4.1-flash @ WorkBuddy direct lane: 3 papers (A-a5608487, A-984e80ee, A-24bcf707) now NA, score 16/24 → **13'/24**. The cause was an isolation defect in our exam setup; the fault is ours. The rest of this page stays as published; where they differ, the [correction notice](https://github.com/getaskclaw/amber/blob/main/docs/corrections-2026-10-02.en.md) governs. -> **W40 起换隔离考场**:从 2026-W40 起本仓的考试在隔离考场里进行,所以 W40 与 W37、W39 的各格跨期不可逐格对比(详见 [2026-W40 期文](results/2026-W40.md))。 +> **From W40 the exam room is isolated**: from 2026-W40 the exams in this repo are taken in an isolated room, so W40 cells are not comparable cell by cell with W37 or W39 (see the [2026-W40 issue](results/2026-W40.en.md)). -用私有题库 **AMBER** 实测 WorkBuddy(CodeBuddy)上的模型,只公开结果,不公开题目。目前有两条通道:ACP 车道(腾讯 CodeBuddy CLI 的 ACP 模式,服务端下发模型目录,W37)和直连车道(`www.workbuddy.ai/v2`,W39、W40)。 -English: [README.en.md](README.en.md) +Running the private **AMBER** benchmark against models on WorkBuddy (CodeBuddy). Results are public; questions are not. There are two channels so far: the ACP lane (the ACP mode of Tencent's CodeBuddy CLI, with a server-pushed model catalog; W37) and the direct lane (`www.workbuddy.ai/v2`; W39 and W40). -> **一句话**:WorkBuddy 直连道上四个模型在 W40 的隔离考场里各做一遍 24 道实战题,过题数是 **deepseek-v4.1-flash 18'/24 · glm-5.3-flash 17'/24 · hy4-preview-f 16'/24 · minimax-m3 16'/24**。四者在写代码、交付、运维、需求这些动手的题上几乎一样,差别在看图案、审查、归因这些判断题。 +> **In one line**: four models on the WorkBuddy direct lane each took the same 24 field tasks in the isolated W40 exam room, and passed **deepseek-v4.1-flash 18'/24 · glm-5.3-flash 17'/24 · hy4-preview-f 16'/24 · minimax-m3 16'/24**. The four are almost the same on the hands-on tasks (coding, delivery, ops, requirements); they differ on the judgment tasks (vision, review, attribution). > -> 分数后的 `'` 表示其中有几题暂不计分(NA),既不算过也不算没过;NA 的原因写在期文里。hy4-preview-f 在榜上有两行:ACP 道(W37)和直连道(W40),是两条不同的道、不同的考场,不并列比较,也不据此说它变强或变弱。 +> A `'` after a score means some cases are NA: not counted as a pass or a fail. The reason for each NA is in the issue. hy4-preview-f has two rows on the board: the ACP lane (W37) and the direct lane (W40). They are different lanes in different exam rooms, so they are not compared, and this page does not say the model got stronger or weaker. -## 成绩一览 +## Scoreboard -| 大类 | 轴 | 考什么 | deepseek-v4.1-flash (WorkBuddy 直连) · [W40](results/2026-W40.md) | glm-5.3-flash (WorkBuddy 直连) · [W40](results/2026-W40.md) | hy4-preview-f (WorkBuddy 直连) · [W40](results/2026-W40.md) | minimax-m3 (WorkBuddy 直连) · [W40](results/2026-W40.md) | hy4-preview-f (WorkBuddy ACP) · [W37](results/2026-W37.md) | +| Group | Axis | What it tests | deepseek-v4.1-flash (WorkBuddy direct) · [W40](results/2026-W40.en.md) | glm-5.3-flash (WorkBuddy direct) · [W40](results/2026-W40.en.md) | hy4-preview-f (WorkBuddy direct) · [W40](results/2026-W40.en.md) | minimax-m3 (WorkBuddy direct) · [W40](results/2026-W40.en.md) | hy4-preview-f (WorkBuddy ACP) · [W37](results/2026-W37.md) | |---|---|---|:-:|:-:|:-:|:-:|:-:| -| 施工面 | 编码 | 照着需求把功能写对 | 5/6 | 5/6 | 5/6 | 5/6 | 5/6 | -| | 交付 | 做完还得交得出东西 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | -| | 运维 | 照规程干脏活 | 6/6 | 6/6 | 6/6 | 5/6 | 6/6 | -| | 需求 | 客户要 A 不要 B | 1/1 | 1/1 | 1/1 | 1/1 | 1/1 | -| | 收敛 | 真干完,不绕圈装忙 | 1/1 | 1/1 | 1/1 | 0/1 | 1/1 | -| 判断面 | UI | 照设计稿做页面 | 0/1 · 1 NA | 0/1 · 1 NA | 0/1 · 1 NA | 0/1 · 1 NA | 1/1 | -| | 视觉 | 给真截图挑毛病 | 1/1 | 0/1 | 0/1 | 1/1 | 0/1 | -| | 防御 | 堵死校验器的漏网口 | 0/2 · 1 NA | 0/2 · 1 NA | 0/2 · 2 NA | 0/2 · 1 NA | 0/2 · 1 NA | -| | 归因 | 毛病对到正确根因 | 0/1 | 0/1 | 0/1 | 0/1 | 0/1 | -| | 审查 | 给别人的交付物挑错 | 1/2 | 1/2 | 0/2 | 1/2 | 1/2 | -| | **合计** | | **18'/24** | **17'/24** | **16'/24** | **16'/24** | **18'/24** | - -每格 = 过了几案/该轴共几案(案 = 一道计分题)。NA = 这一案作废或暂停计分,不算过也不算没过;总分带 `'` 表示其中有 NA。多数轴只有 1–2 案,差一案读数就变,所以别把小差距当结论。各期考试日期不同,数字是当期快照。 +| Building | Coding | Implement the spec correctly | 5/6 | 5/6 | 5/6 | 5/6 | 5/6 | +| | Delivery | Done means handed in | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | +| | Ops | Follow the runbook | 6/6 | 6/6 | 6/6 | 5/6 | 6/6 | +| | Requirements | Ship A when A was asked | 1/1 | 1/1 | 1/1 | 1/1 | 1/1 | +| | Convergence | Finish, don't spin | 1/1 | 1/1 | 1/1 | 0/1 | 1/1 | +| Judging | UI | Build the page to the mock | 0/1 · 1 NA | 0/1 · 1 NA | 0/1 · 1 NA | 0/1 · 1 NA | 1/1 | +| | Vision | Spot defects in screenshots | 1/1 | 0/1 | 0/1 | 1/1 | 0/1 | +| | Defense | Plug every hole in the validator | 0/2 · 1 NA | 0/2 · 1 NA | 0/2 · 2 NA | 0/2 · 1 NA | 0/2 · 1 NA | +| | Attribution | Pin defects to their root cause | 0/1 | 0/1 | 0/1 | 0/1 | 0/1 | +| | Review | Inspect someone else's work | 1/2 | 1/2 | 0/2 | 1/2 | 1/2 | +| | **Total** | | **18'/24** | **17'/24** | **16'/24** | **16'/24** | **18'/24** | + +Each cell = cases passed / cases on that axis (a case is one scored task). NA = the case was voided or put on hold; it counts as neither a pass nor a fail, and a total carrying `'` contains at least one NA. Most axes hold only 1–2 cases, so one case moves the reading: do not over-read small gaps. Sittings are from different weeks; every number is a snapshot. -W37 是 ACP 道的 hy4-preview-f,W40 是直连道、隔离考场:两列不可逐格对比。 +The W37 column is the ACP-lane hy4-preview-f; W40 is the direct lane in the isolated room: the columns are not comparable cell by cell. -## 这是什么 +## What this is -- 「道」= 同一个模型名在不同家的卖场/接口;「案」= 一道题,「卷」= 一场考试记录(一案多卷 = 一道题的几个变体场次)。 +- A 'lane' is one vendor's shop/API for a model name; a 'case' is one task, a 'run' is one sitting (a multi-variant case has several runs). -- 每期一篇 `results/YYYY-Www.md`:同题、同 harness(跑考试并记分的程序),对目标模型跑全库;同车道跨模型并排。 -- 一期固定报告:题集规模与哈希、每案找茬分(d2 分,我们的打分,算法不公开)与通过/失败、终端终态(程序跑完时的退出状态)、token 用量(若车道上报——本道不上报,以墙钟代替)与时延、环境指纹、按证据纪律写的定性裁决。 -- 题目、oracle(判分器)、transcript(答题全过程记录)、中间产物**永不公开**(见下「发布纪律」)。 -- 姐妹仓:[amber-gpt](https://github.com/getaskclaw/amber-gpt)(GPT 系周测)、[amber-crof](https://github.com/getaskclaw/amber-crof)(CrofAI 周测)、[amber-ollama](https://github.com/getaskclaw/amber-ollama)(Ollama Cloud 周测)、[amber-devin](https://github.com/getaskclaw/amber-devin)(Devin 周测)、[amber-opencode](https://github.com/getaskclaw/amber-opencode)(OpenCode Go 道)、[amber-commandcode](https://github.com/getaskclaw/amber-commandcode)(CommandCode 道)、[amber-deepseek](https://github.com/getaskclaw/amber-deepseek)(DeepSeek 官方道)、[amber-doubao](https://github.com/getaskclaw/amber-doubao)、[amber-goldenpotato](https://github.com/getaskclaw/amber-goldenpotato)、[amber-kimi](https://github.com/getaskclaw/amber-kimi)、[amber-stepfun](https://github.com/getaskclaw/amber-stepfun)。同一个 deepseek-v4.1 家族在官方/转发道的成绩见对应仓库;本仓的对照轴是 **WorkBuddy 各车道上的模型**(ACP 道与直连道分开看),跨仓引用一律带日期与档位声明。 -- AMBER 是 agentic 实战题库(施工/运维/审查/视觉/需求漂移——题中要求中途变化),规范与制题工具见 [getaskclaw/amber](https://github.com/getaskclaw/amber);考题本体私有。 +- One `results/YYYY-Www.md` per edition: same questions, same harness (the program that runs the exam and scores it), full library per model; models on the same lane side by side. +- Each edition pins: library size and hashes, per-case defect-hunt score and pass/fail, terminal states (how the run process exited), token usage where the lane reports it (this lane does not — wall clock stands in), latency, environment fingerprint, and qualitative verdicts written under evidence discipline. +- Questions, oracles, transcripts (full answer logs)s and intermediate artifacts are **never published** (see "Publishing rules"). +- Sister repos: [amber-gpt](https://github.com/getaskclaw/amber-gpt), [amber-crof](https://github.com/getaskclaw/amber-crof), [amber-ollama](https://github.com/getaskclaw/amber-ollama), [amber-devin](https://github.com/getaskclaw/amber-devin), [amber-opencode](https://github.com/getaskclaw/amber-opencode), [amber-commandcode](https://github.com/getaskclaw/amber-commandcode), [amber-deepseek](https://github.com/getaskclaw/amber-deepseek), [amber-doubao](https://github.com/getaskclaw/amber-doubao), [amber-goldenpotato](https://github.com/getaskclaw/amber-goldenpotato), [amber-kimi](https://github.com/getaskclaw/amber-kimi), [amber-stepfun](https://github.com/getaskclaw/amber-stepfun). Scores for the same deepseek-v4.1 family on official/relay lanes live in those repos; this repo's comparison axis is **across models on the WorkBuddy lanes** (the ACP lane and the direct lane are read separately). Cross-repo citations always carry date and band. +- AMBER is an agentic field benchmark (build / ops / review / vision / requirement-drift (the requirements change mid-task)); spec and authoring tools at [getaskclaw/amber](https://github.com/getaskclaw/amber); the questions themselves are private. -## 一分钟看懂 W40 +## W40 in one minute -![W40 分歧图:24 案里四个直连道模型结果不同的 5 案](results/assets/2026-W40-diff.zh.png?v=20261005) +![W40 divergence: the 5 of 24 cases where the four direct-lane models differ](results/assets/2026-W40-diff.en.png?v=20261005) -同一条直连道、同一天(2026-10-04,UTC)、high 档、同一个隔离考场、同 24 案同哈希:**deepseek-v4.1-flash 18'/24、glm-5.3-flash 17'/24、hy4-preview-f 16'/24、minimax-m3 16'/24**。看图案(A-ea80d793)首考时考场的图片请求被接口拒绝,修好考场后只重考了这一格;品牌题 A-d9b79b46 和防御案 A-d511f9e8 在四条道上都挂起(NA)。逐案矩阵、考试条件与各 NA 的说明见 [2026-W40 期文](results/2026-W40.md)。另 19 案四者结果相同:14 案全过,3 案全没过(A-87c472cb、A-a317e74b、A-cdc3d11a),2 案全 NA(A-d511f9e8、A-d9b79b46)。完整 24 案矩阵见期文。 +Same direct lane, same day (2026-10-04, UTC), high band, same isolated exam room, same 24 cases and hashes: **deepseek-v4.1-flash 18'/24, glm-5.3-flash 17'/24, hy4-preview-f 16'/24, minimax-m3 16'/24**. On the vision case (A-ea80d793) the room's image request was rejected in the first sitting; after the room fix only that cell was re-taken. The brand case A-d9b79b46 and the defense case A-d511f9e8 are on hold (NA) on all four lanes. The per-case matrix, the exam conditions and the explanation of each NA are in the [2026-W40 issue](results/2026-W40.en.md). The other 19 cases have the same result for all four: 14 all pass, 3 all fail (A-87c472cb, A-a317e74b, A-cdc3d11a), 2 all NA (A-d511f9e8, A-d9b79b46). The full 24-case matrix is in the issue. -## 一分钟看懂 W37 +## W37 in one minute -同一 ACP 车道、同 high 档、同 23 案同哈希:**hy4-preview-f 17/23**(A-be92627f 外的核验面仍挂,但施工 5/6 + OPS 6/6 + UI 12/12)、**deepseek-v4.1-flash 15/23**(核验案 A-be92627f 9/9 = 该案史上首个过案、核验面第二席)。逐案矩阵与车道账本见 [2026-W37 期文](results/2026-W37.md)。 +Same ACP lane, same high band, same 23 cases and hashes: **hy4-preview-f 17/23** (build 5/6 + OPS 6/6 sweep + ui-build 12/12), **deepseek-v4.1-flash 15/23** (A-be92627f 9/9 — first-ever pass on that case; second verify-face pass overall). Per-case matrix and lane ledger in the [2026-W37 issue](results/2026-W37.md). -上面的 W37 数字是当期口径(23 案)。上方「成绩一览」按现行库把 W37 这条 ACP 道的 hy4-preview-f 记为 18'/24:在 17/23 之外加上后来补测通过的收敛案,防御案 A-d511f9e8 随全车道挂起记 NA。W37 与 W40 的通道和考场都不同,不逐格对比。 +The W37 numbers above use the issue's own count (23 cases). The "Scoreboard" above counts this ACP-lane hy4-preview-f row as 18'/24 on the current library: the 17/23 plus the convergence case that was added later and passed, with defense case A-d511f9e8 as NA through the all-lane hold. W37 and W40 differ in channel and exam room, so they are not compared cell by cell. -## 发布纪律(红线) +## Publishing rules (red lines) -1. 只发:分数与聚合、token 用量(若车道上报)、速度、定性裁决。 -2. 永不发:题目内容、oracle/判分器、transcript、考生工作区、任何能复原题面的中间产物。 -3. 每期必钉:模型 ID、effort 档(思考力度档位)、日期(UTC)、harness 版本、每案内容哈希(bundle_sha,每题内容的哈希指纹)。哈希用于对照 [amber](https://github.com/getaskclaw/amber) 的公开哈希清单,自证题集未变。 -4. 案号与题目结构属私有面:公开结果里案例只用稳定别名(A-xxxxxxxx,哈希派生)+ bundle 哈希作句柄;内部案号、变体名、题目描述永不出现。 -5. 基调:这是社区实测,不是对厂商的攻击。数据说话,措辞克制。 +1. Publish only: scores and aggregates, token usage (where the lane reports it), speed, qualitative verdicts. +2. Never publish: question content, oracles/graders, transcripts, candidate workspaces, any intermediate that could reconstruct a question. +3. Every edition pins: model ID, effort band (the thinking-effort setting), date (UTC), harness version, per-case content hash (bundle_sha (per-case content-hash fingerprint)) — checkable against the public hash index in [amber](https://github.com/getaskclaw/amber). +4. Case IDs and question structure are private: published results use only stable aliases (A-xxxxxxxx, hash-derived) plus bundle hashes as handles; internal case IDs, variant names, and question descriptions never appear. +5. Tone: this is a community measurement, not an attack on any vendor. Data speaks; wording stays restrained. -## 一个方法论前提 +## A methodological caveat -同名模型、同 provider,两次跑也可能不同分——推理参数、负载、服务端版本都在漂;转发/聚合车道还多一层装帧差。所以这里的一切结论都带日期与档位,且定期重测。单日数字是快照,不是定律。 +Same model name, same provider, two runs can still differ — sampling parameters, load, and server-side versions all drift; relay/aggregator lanes add a framing layer on top. Every conclusion here carries a date and a band, and lanes get re-measured regularly. A single day's number is a snapshot, not a law. -## 结果索引 +## Results index -| 期 | 考生 | 成绩(W37 为 23 案 / 公共子集 21;W40 为 24 案) | 一句话 | +| Edition | Candidate | Score (W37: 23 cases / 21-case public subset; W40: 24 cases) | One-liner | |---|---|---|---| -| [2026-W40](results/2026-W40.md) | **deepseek-v4.1-flash**(WorkBuddy 直连) | **18'/24** | 四者最高、没有超时;看图案重考过;归因 A-a317e74b 14/15 只差 1 项 | -| [2026-W40](results/2026-W40.md) | glm-5.3-flash(WorkBuddy 直连) | **17'/24** | 审查 A-47eea242 找茬分 4 过;这个通道不接受图片输入,看图案没过 | -| [2026-W40](results/2026-W40.md) | hy4-preview-f(WorkBuddy 直连) | **16'/24** | 3 个 NA(防御 2、UI 1);看到图但看图案没过;与 W37 的 ACP 道是两条道,不并列比较 | -| [2026-W40](results/2026-W40.md) | minimax-m3(WorkBuddy 直连) | **16'/24** | 负案最多(6);收敛案没过;品牌题记 NA(挂起;2026-10-05 由「负」改记 NA);缺的 9 案当天补考 | -| [2026-W37](results/2026-W37.md) | **hy4-preview-f**(x0.00 免费档) | **17/23**(15/21) | 入档即第二梯队;OPS 6/6 + UI 12/12 满分;核验 0/3,审查/视觉仍挂 | -| [2026-W37](results/2026-W37.md) | deepseek-v4.1-flash(x0.00 免费档) | **15/23**(13/21) | A-be92627f 9/9 = 该案史上首个过案(核验面第二席);同脑比官方 GA 道低一案,结构互换 | -| [2026-W38 更正特刊](results/2026-W38-correction.md) | W38 全库复核:本仓改判 0 格 · 挂起 1 格 | W37 hy4 视觉 1 格挂起 | +| [2026-W40](results/2026-W40.en.md) | **deepseek-v4.1-flash** (WorkBuddy direct) | **18'/24** | Highest of the four, no timeouts; passed the vision re-take; attribution A-a317e74b 14/15, one check short | +| [2026-W40](results/2026-W40.en.md) | glm-5.3-flash (WorkBuddy direct) | **17'/24** | Passed review A-47eea242 (fault-finding 4); this channel does not accept image input, vision case not passed | +| [2026-W40](results/2026-W40.en.md) | hy4-preview-f (WorkBuddy direct) | **16'/24** | 3 NA (2 defense, 1 UI); saw the picture but did not pass the vision case; a different lane from the W37 ACP row, not compared | +| [2026-W40](results/2026-W40.en.md) | minimax-m3 (WorkBuddy direct) | **16'/24** | Most losses (6); did not pass the convergence case; brand case NA (on hold; changed from a loss to NA on 2026-10-05); 9 missing cases made up the same day | +| [2026-W37](results/2026-W37.md) | **hy4-preview-f** (x0.00 free tier) | **17/23** (15/21) | Second-tier entry on debut; OPS 6/6 sweep + ui-build 12/12; verify 0/3, review/vision still fail | +| [2026-W37](results/2026-W37.md) | deepseek-v4.1-flash (x0.00 free tier) | **15/23** (13/21) | A-be92627f 9/9 = first-ever pass on that case (second verify-face pass overall); one case below its official-GA sibling with swapped structure | +| [2026-W38 correction notice](results/2026-W38-correction.en.md) | W38 full-library review: 0 cells reversed · 1 held here | 1 W37 hy4 vision cell held | -## 免责 +## Disclaimer -与腾讯、CodeBuddy/WorkBuddy 无任何隶属/赞助关系。分数是特定周、特定档位的快照,不构成采购建议。 +No affiliation with or sponsorship by Tencent or CodeBuddy/WorkBuddy. Scores are snapshots of a specific week and band, not purchasing advice. diff --git a/README.zh-CN.md b/README.zh-CN.md new file mode 100644 index 0000000..06f3a25 --- /dev/null +++ b/README.zh-CN.md @@ -0,0 +1,89 @@ +[English](README.md) · 简体中文 + +# amber-workbuddy + +> ⚠️ **更正(2026-10-02,另一项)**:防御轴的一案 A-d511f9e8 在所有车道上改记 NA(考场判的不是考生交付的文件,判分还要求了题面没写的事)。分母不变,**过案数不变**,每条道的总分都带 `'`。本仓各期成绩表里这一格请按 NA 读,其余内容保留原样,以[更正声明](https://github.com/getaskclaw/amber/blob/main/docs/corrections-2026-10-02-a-d511f9e8.md)为准。 + +> ⚠️ **更正(2026-10-02)**:以下考卷在作答时越出考卷、接触了判分材料,不计胜负。deepseek-v4.1-flash @ WorkBuddy 直连道 有 3 张卷(A-a5608487、A-984e80ee、A-24bcf707)改记 NA,成绩 16/24 → **13'/24**。原因是考场隔离缺陷,责任在我们。本页其余内容保留原样,以[更正声明](https://github.com/getaskclaw/amber/blob/main/docs/corrections-2026-10-02.md)为准。 + +> **W40 起换隔离考场**:从 2026-W40 起本仓的考试在隔离考场里进行,所以 W40 与 W37、W39 的各格跨期不可逐格对比(详见 [2026-W40 期文](results/2026-W40.md))。 + +用私有题库 **AMBER** 实测 WorkBuddy(CodeBuddy)上的模型,只公开结果,不公开题目。目前有两条通道:ACP 车道(腾讯 CodeBuddy CLI 的 ACP 模式,服务端下发模型目录,W37)和直连车道(`www.workbuddy.ai/v2`,W39、W40)。 + +> **一句话**:WorkBuddy 直连道上四个模型在 W40 的隔离考场里各做一遍 24 道实战题,过题数是 **deepseek-v4.1-flash 18'/24 · glm-5.3-flash 17'/24 · hy4-preview-f 16'/24 · minimax-m3 16'/24**。四者在写代码、交付、运维、需求这些动手的题上几乎一样,差别在看图案、审查、归因这些判断题。 +> +> 分数后的 `'` 表示其中有几题暂不计分(NA),既不算过也不算没过;NA 的原因写在期文里。hy4-preview-f 在榜上有两行:ACP 道(W37)和直连道(W40),是两条不同的道、不同的考场,不并列比较,也不据此说它变强或变弱。 + +## 成绩一览 + + + +| 大类 | 轴 | 考什么 | deepseek-v4.1-flash (WorkBuddy 直连) · [W40](results/2026-W40.md) | glm-5.3-flash (WorkBuddy 直连) · [W40](results/2026-W40.md) | hy4-preview-f (WorkBuddy 直连) · [W40](results/2026-W40.md) | minimax-m3 (WorkBuddy 直连) · [W40](results/2026-W40.md) | hy4-preview-f (WorkBuddy ACP) · [W37](results/2026-W37.md) | +|---|---|---|:-:|:-:|:-:|:-:|:-:| +| 施工面 | 编码 | 照着需求把功能写对 | 5/6 | 5/6 | 5/6 | 5/6 | 5/6 | +| | 交付 | 做完还得交得出东西 | 3/3 | 3/3 | 3/3 | 3/3 | 3/3 | +| | 运维 | 照规程干脏活 | 6/6 | 6/6 | 6/6 | 5/6 | 6/6 | +| | 需求 | 客户要 A 不要 B | 1/1 | 1/1 | 1/1 | 1/1 | 1/1 | +| | 收敛 | 真干完,不绕圈装忙 | 1/1 | 1/1 | 1/1 | 0/1 | 1/1 | +| 判断面 | UI | 照设计稿做页面 | 0/1 · 1 NA | 0/1 · 1 NA | 0/1 · 1 NA | 0/1 · 1 NA | 1/1 | +| | 视觉 | 给真截图挑毛病 | 1/1 | 0/1 | 0/1 | 1/1 | 0/1 | +| | 防御 | 堵死校验器的漏网口 | 0/2 · 1 NA | 0/2 · 1 NA | 0/2 · 2 NA | 0/2 · 1 NA | 0/2 · 1 NA | +| | 归因 | 毛病对到正确根因 | 0/1 | 0/1 | 0/1 | 0/1 | 0/1 | +| | 审查 | 给别人的交付物挑错 | 1/2 | 1/2 | 0/2 | 1/2 | 1/2 | +| | **合计** | | **18'/24** | **17'/24** | **16'/24** | **16'/24** | **18'/24** | + +每格 = 过了几案/该轴共几案(案 = 一道计分题)。NA = 这一案作废或暂停计分,不算过也不算没过;总分带 `'` 表示其中有 NA。多数轴只有 1–2 案,差一案读数就变,所以别把小差距当结论。各期考试日期不同,数字是当期快照。 + + + +W37 是 ACP 道的 hy4-preview-f,W40 是直连道、隔离考场:两列不可逐格对比。 + +## 这是什么 + +- 「道」= 同一个模型名在不同家的卖场/接口;「案」= 一道题,「卷」= 一场考试记录(一案多卷 = 一道题的几个变体场次)。 + +- 每期一篇 `results/YYYY-Www.md`:同题、同 harness(跑考试并记分的程序),对目标模型跑全库;同车道跨模型并排。 +- 一期固定报告:题集规模与哈希、每案找茬分(d2 分,我们的打分,算法不公开)与通过/失败、终端终态(程序跑完时的退出状态)、token 用量(若车道上报——本道不上报,以墙钟代替)与时延、环境指纹、按证据纪律写的定性裁决。 +- 题目、oracle(判分器)、transcript(答题全过程记录)、中间产物**永不公开**(见下「发布纪律」)。 +- 姐妹仓:[amber-gpt](https://github.com/getaskclaw/amber-gpt)(GPT 系周测)、[amber-crof](https://github.com/getaskclaw/amber-crof)(CrofAI 周测)、[amber-ollama](https://github.com/getaskclaw/amber-ollama)(Ollama Cloud 周测)、[amber-devin](https://github.com/getaskclaw/amber-devin)(Devin 周测)、[amber-opencode](https://github.com/getaskclaw/amber-opencode)(OpenCode Go 道)、[amber-commandcode](https://github.com/getaskclaw/amber-commandcode)(CommandCode 道)、[amber-deepseek](https://github.com/getaskclaw/amber-deepseek)(DeepSeek 官方道)、[amber-doubao](https://github.com/getaskclaw/amber-doubao)、[amber-goldenpotato](https://github.com/getaskclaw/amber-goldenpotato)、[amber-kimi](https://github.com/getaskclaw/amber-kimi)、[amber-stepfun](https://github.com/getaskclaw/amber-stepfun)。同一个 deepseek-v4.1 家族在官方/转发道的成绩见对应仓库;本仓的对照轴是 **WorkBuddy 各车道上的模型**(ACP 道与直连道分开看),跨仓引用一律带日期与档位声明。 +- AMBER 是 agentic 实战题库(施工/运维/审查/视觉/需求漂移——题中要求中途变化),规范与制题工具见 [getaskclaw/amber](https://github.com/getaskclaw/amber);考题本体私有。 + +## 一分钟看懂 W40 + +![W40 分歧图:24 案里四个直连道模型结果不同的 5 案](results/assets/2026-W40-diff.zh.png?v=20261005) + +同一条直连道、同一天(2026-10-04,UTC)、high 档、同一个隔离考场、同 24 案同哈希:**deepseek-v4.1-flash 18'/24、glm-5.3-flash 17'/24、hy4-preview-f 16'/24、minimax-m3 16'/24**。看图案(A-ea80d793)首考时考场的图片请求被接口拒绝,修好考场后只重考了这一格;品牌题 A-d9b79b46 和防御案 A-d511f9e8 在四条道上都挂起(NA)。逐案矩阵、考试条件与各 NA 的说明见 [2026-W40 期文](results/2026-W40.md)。另 19 案四者结果相同:14 案全过,3 案全没过(A-87c472cb、A-a317e74b、A-cdc3d11a),2 案全 NA(A-d511f9e8、A-d9b79b46)。完整 24 案矩阵见期文。 + +## 一分钟看懂 W37 + +同一 ACP 车道、同 high 档、同 23 案同哈希:**hy4-preview-f 17/23**(A-be92627f 外的核验面仍挂,但施工 5/6 + OPS 6/6 + UI 12/12)、**deepseek-v4.1-flash 15/23**(核验案 A-be92627f 9/9 = 该案史上首个过案、核验面第二席)。逐案矩阵与车道账本见 [2026-W37 期文](results/2026-W37.md)。 + +上面的 W37 数字是当期口径(23 案)。上方「成绩一览」按现行库把 W37 这条 ACP 道的 hy4-preview-f 记为 18'/24:在 17/23 之外加上后来补测通过的收敛案,防御案 A-d511f9e8 随全车道挂起记 NA。W37 与 W40 的通道和考场都不同,不逐格对比。 + +## 发布纪律(红线) + +1. 只发:分数与聚合、token 用量(若车道上报)、速度、定性裁决。 +2. 永不发:题目内容、oracle/判分器、transcript、考生工作区、任何能复原题面的中间产物。 +3. 每期必钉:模型 ID、effort 档(思考力度档位)、日期(UTC)、harness 版本、每案内容哈希(bundle_sha,每题内容的哈希指纹)。哈希用于对照 [amber](https://github.com/getaskclaw/amber) 的公开哈希清单,自证题集未变。 +4. 案号与题目结构属私有面:公开结果里案例只用稳定别名(A-xxxxxxxx,哈希派生)+ bundle 哈希作句柄;内部案号、变体名、题目描述永不出现。 +5. 基调:这是社区实测,不是对厂商的攻击。数据说话,措辞克制。 + +## 一个方法论前提 + +同名模型、同 provider,两次跑也可能不同分——推理参数、负载、服务端版本都在漂;转发/聚合车道还多一层装帧差。所以这里的一切结论都带日期与档位,且定期重测。单日数字是快照,不是定律。 + +## 结果索引 + +| 期 | 考生 | 成绩(W37 为 23 案 / 公共子集 21;W40 为 24 案) | 一句话 | +|---|---|---|---| +| [2026-W40](results/2026-W40.md) | **deepseek-v4.1-flash**(WorkBuddy 直连) | **18'/24** | 四者最高、没有超时;看图案重考过;归因 A-a317e74b 14/15 只差 1 项 | +| [2026-W40](results/2026-W40.md) | glm-5.3-flash(WorkBuddy 直连) | **17'/24** | 审查 A-47eea242 找茬分 4 过;这个通道不接受图片输入,看图案没过 | +| [2026-W40](results/2026-W40.md) | hy4-preview-f(WorkBuddy 直连) | **16'/24** | 3 个 NA(防御 2、UI 1);看到图但看图案没过;与 W37 的 ACP 道是两条道,不并列比较 | +| [2026-W40](results/2026-W40.md) | minimax-m3(WorkBuddy 直连) | **16'/24** | 负案最多(6);收敛案没过;品牌题记 NA(挂起;2026-10-05 由「负」改记 NA);缺的 9 案当天补考 | +| [2026-W37](results/2026-W37.md) | **hy4-preview-f**(x0.00 免费档) | **17/23**(15/21) | 入档即第二梯队;OPS 6/6 + UI 12/12 满分;核验 0/3,审查/视觉仍挂 | +| [2026-W37](results/2026-W37.md) | deepseek-v4.1-flash(x0.00 免费档) | **15/23**(13/21) | A-be92627f 9/9 = 该案史上首个过案(核验面第二席);同脑比官方 GA 道低一案,结构互换 | +| [2026-W38 更正特刊](results/2026-W38-correction.md) | W38 全库复核:本仓改判 0 格 · 挂起 1 格 | W37 hy4 视觉 1 格挂起 | + +## 免责 + +与腾讯、CodeBuddy/WorkBuddy 无任何隶属/赞助关系。分数是特定周、特定档位的快照,不构成采购建议。