Import Bos direct batch 20261014-A-bos-direct

This commit is contained in:
amos 2026-10-15 00:37:44 +08:00
parent 3525e76b9a
commit f1acc8c938
728 changed files with 33036 additions and 0 deletions

View File

@ -253,3 +253,31 @@ M1(20 题准入、完整尝试、正负例与重判)已有实际证据;M2/
- 共享根目录下 2026-10-12 的旧代码副本(根目录 `AGENTS.md`、`rollout/`、`tasks/` 等)未动,待用户决定。
现在 Codex 可按 [task-factory-brief.md](task-factory-brief.md) 在 36 的工作副本中开工。另:本地连跑 9 次测试,1 次在高负载下报 1 个错误(未能捕获是哪项),其余 8 次全部通过,待再现时定位。
## 2026-10-14 制题批次 20261014-A-bos-direct(完成;2026-10-15 冷迁收尾)
按作业 A 逐题盘点、静态导入、校准、独立评审及首次真实试采。边界:只新增任务与文档,不执行来源 shell/工具脚本、不直接读取或复制凭据、不写来源、不改平台、不推送;controller 仅接收已授权私有文件路径。基线 HEAD 为 `3525e76b9abfba411f8d4b143e2721afc4c53a8f`,fast-forward 已确认最新。
盘点与准入:182 个候选,importable 93(含已有 20)、rglob 84、ROS2 5。73 次新导入尝试中 67 成功、6 被导入器拒绝;67 题可信本地与正式 Docker 校准均 PASS。独立内容评审最终保留 54 题,拒绝 13 题,当前 tasks 为既有 21 题加新增 54 题,共 75 题。每道成功导入题都填写 spec 第 1、4、7 节,其余沿用 Bos 原题;拒绝题包与规格保留在批次 rejected 目录,不进任务池。
- 导入器拒绝 6:公开接口名字检查 1、静态样例解析不支持条件/rglob 循环 3、chmod 权限语句 1、样例数量超过上限 1;保留原拒因,不执行脚本绕过或硬改来源。
- 内容评审拒绝 13:n3_068 的定向 cheat 是语法错误坏例;n2_016 的 gold/valid 精确 PSD 检查存在浮点下溢;另外 n2_014、038、047、049、087、088、095、102、111、115、135 未公开 solve 顶层 list/空批 [] 输出合同。后 11 题由试采暴露后对 65 题逐项专项复核发现;原评审备份与修正记录完整保留。
- BLOCKED 89:84 道 rglob 题需要候选通配、多文件冻结与评分合同;5 道 ROS2 题需要正式环境族及固定镜像准入。本批继续下一题,不扩平台。
最终五族:motor-sensor-control 30、linux-device-abi 10、firmware-image-boot 6、embedded-vision 4、robot-navigation 4;完整 ID 清单及全部拒绝/BLOCKED 原因见冷盘 review.md。n3_094 原单题 real-time-scheduling 族回并到 motor-sensor-control,family 改变产生新 digest `ae01826d377ff5a219796f83593bdc146eb0d4e098a5289ac7b6b9e69fab930a`,新身份重新本地和正式 Docker 校准 PASS;旧回执只作历史事实。可信本地与正式 Docker 各 2195 次样例运行(包括该重校准),每类样例正式各五次符合预期。
闸门 3:54 道合格题均作一次 GLM-5.3 / Pi 0.73.1 真实试采,50 PASS、4 FAIL;全部正常退出、host/native/broker 对账 VERIFIED、cleanup confirmed,冻结候选在新无网评分容器重判 54/54 一致。原始物理尝试共 58 条完整轨迹(52 PASS、6 FAIL),另 4 条来自后被拒绝的公开输出合同缺陷题,保留原 PASS/FAIL、证据与封存事实,在 quality-report 中明确排除,不伪改 INFRA_ERROR 或当作能力信号。58 条原始轨迹全部重判一致,并按 archive 白名单、manifest 逐成员字节数/摘要核对后封存。
有效 FAIL 为 n2_013(巨大整数转 float 溢出,未按公开契约 ValueError)、n2_070(精确方差正值下溢,误过零门限)、n2_071(精确实数行列式门限误用浮点阈值)、n2_084(接收超过公开局部范围的观测)。只读核对候选、公开合同和首个评分失败证据;不声称后续用例通过,不修改候选或向运行 Agent 注入答案。真实 FAIL 不否决任务,54 题 spec 最终状态 ADMITTED;本批未做闸门 4,不能据单次结果作难度分池。
配置与验证:固定 trace5 runtime/actual image `sha256:aa5d00779fe03e9eadbf29c030711a4ecd1e5f894255868b5ca451a425299efc`,GLM-5.3 每次回复上限 131072,sampling=service-default,四路并发,每题 3600 秒防挂死。基线、导入后及最终保留题集单元测试均通过;最终运行 73 项,1 项因 host 无 Node 跳过,其余通过。最终独立只读审阅确认 54 个 digest 与 admission/raw/quality 一致,规格满足模板,13 拒题均不在 tasks,已有 21 题及 rollout/agent/tests 对 HEAD 无改动。 Git whitespace 检查 exit=2:导入原材料/样例的 315 处格式提示(文件尾空行及行尾空白);无 spec 或本节提示,保留冻结原字节,不因格式重写已校准任务。真实日志另存冷盘 git-whitespace-check。
证据目标目录:`/mnt/data/ming01.wang/embedded-rollout/factory/20261014-A-bos-direct/`(inventory/import/gates/calibration/reviews/admission、raw report、quality report、review/summary 与拒绝包);运行证据位于同根 `runs/20261014-A-bos-direct{,-dev,-calibration}/`;58 个封存包位于 `artifacts/20261014-A-bos-direct/`。迁移前 source/回执保持冻结,五个目标均由本批独占新建,不覆盖历史数据。
收尾完成:五个批次目录冷迁时共 19743 个普通文件、1108494513 字节,逐路径文件数/字节、空目录结构及链接均一致,完整窗口 source before/after/final 元数据未变,58 包摘要与字节数全部核对。先核对再删除本批热副本;迁移后五根实测缺失、占用 0 字节(Git 任务包保留),残留容器 0、网络 0、活跃作业 0,disk_low 0、attempt_size 0。/home 可用 154132492288 字节(约 143.55 GiB)。四个目录普通用户直接清理;真实运行目录的权限残留仅用挂载该单个目录的无网 trace5 容器清理,helper exit=0 且已不存在。原始回执与轨迹不改写;最终 review/summary 在 cold 更新,迁移前版本另存,后续派生元数据不计入复制时原始 payload 统计。
迁移历史:第一次顺序复制小文件较慢,为提速按本批专属 PID/进程组正常中止,保留原记录,改为 16 路独占子目录复制。恢复预检发现一个 rsync 遗留 0 字节临时文件,明确记 BLOCKED;核对本批拥有的路径、UID、普通文件属性与源前缀后,完整保存到 migration-quarantine 并记录空文件 SHA,再移除原临时路径恢复复制。首次中止和预检阻塞事实均保留,最终第二次完整迁移已通过;不是模型 INFRA_ERROR,未删历史数据或为腾空间清理。准备版辅助脚本也在验证 cold 留档摘要后清理。
本地提交范围为 54 道新任务目录及本节;只提交 main、不推送。下一步按来源或平台所需合同解除已列拒绝/BLOCKED 后,以新身份重新准入;闸门 4 仍为另批范围。
限制:Salty CLI/连接不可用,上传及下载摘要核对单列 BLOCKED,本地封存不冒充已上传。既有 Bos wrapper 的 checker/candidate 同 UID 临时限制按 spec 的 shortcut 记录,未声称任意恶意候选隔离已解决;本批不改平台。原回执绝对路径与字节不改写,冷迁后入口尚无路径重定位合同,直接重跑需恢复原路径或审定合同升级。后续先由来源/平台对应升级解除拒绝或 BLOCKED,再以新身份重校准/试采;难度分池另批处理。

View File

@ -0,0 +1,40 @@
#include <termios.h>
#include <unistd.h>
#include <fcntl.h>
#include <poll.h>
#include <signal.h>
#include <cerrno>
#include <cstdlib>
#include <cstdio>
#include <cstring>
#include <climits>
#include <string>
static volatile sig_atomic_t interrupted=0;
static void signal_handler(int){interrupted=1;}
static bool write_all(int fd,const char* p,size_t n){
while(n){ssize_t k=write(fd,p,n);if(k<0){if(errno==EINTR){if(interrupted)return false;continue;}return false;}if(!k)return false;p+=k;n-=k;}return true;
}
int main(int argc,char** argv){
if(argc>3){fprintf(stderr,"usage: serial_TXRX [device [rx_count]]\n");return 2;}
unsigned long long limit=ULLONG_MAX;
if(argc==3){std::string s=argv[2];if(s.empty()||s.find_first_not_of("0123456789")!=std::string::npos){fprintf(stderr,"invalid count\n");return 2;}errno=0;char* end=nullptr;limit=strtoull(argv[2],&end,10);if(errno||!limit||*end){fprintf(stderr,"invalid count\n");return 2;}}
int fd=open(argc>1?argv[1]:"/dev/ttyS0",O_RDWR|O_NOCTTY);
if(fd<0){perror("open serial");return 1;}
termios old{},cfg{};if(tcgetattr(fd,&old)){perror("tcgetattr");close(fd);return 1;}cfg=old;cfmakeraw(&cfg);cfsetispeed(&cfg,B115200);cfsetospeed(&cfg,B115200);
cfg.c_cflag=(cfg.c_cflag&~(CSIZE|PARENB|CSTOPB|CRTSCTS))|CS8|CLOCAL|CREAD;cfg.c_iflag&=~(IXON|IXOFF|IXANY);cfg.c_cc[VMIN]=1;cfg.c_cc[VTIME]=0;
if(tcsetattr(fd,TCSANOW,&cfg)){perror("tcsetattr");close(fd);return 1;}
struct sigaction sa{};sa.sa_handler=signal_handler;sigemptyset(&sa.sa_mask);sigaction(SIGINT,&sa,nullptr);
bool input=true;unsigned long long received=0;int status=0;
while(!interrupted&&received<limit){
pollfd fds[2]={{input?STDIN_FILENO:-1,POLLIN,0},{fd,POLLIN,0}};
int result=poll(fds,2,-1);if(result<0){if(errno==EINTR)continue;status=1;break;}
char buf[4096];
if(input&&(fds[0].revents&(POLLIN|POLLHUP))){ssize_t n=read(0,buf,sizeof(buf));if(n==0)break;else if(n>0){if(!write_all(fd,buf,n)){if(!interrupted)status=1;break;}}else if(errno!=EINTR){status=1;break;}}
if(fds[1].revents&(POLLIN|POLLHUP|POLLERR)){
size_t capacity=sizeof(buf);if(limit-received<capacity)capacity=limit-received;
ssize_t n=read(fd,buf,capacity);if(n>0){if(!write_all(1,buf,n)){if(!interrupted)status=1;break;}received+=n;}else if(n==0){status=received==limit?0:1;break;}else if(errno!=EINTR){status=1;break;}
}
}
if(tcsetattr(fd,TCSANOW,&old)){perror("restore");status=1;}close(fd);return interrupted?0:status;
}

View File

@ -0,0 +1,40 @@
#include <termios.h>
#include <unistd.h>
#include <fcntl.h>
#include <poll.h>
#include <signal.h>
#include <cerrno>
#include <cstdlib>
#include <cstdio>
#include <cstring>
#include <climits>
#include <string>
static volatile sig_atomic_t interrupted=0;
static void signal_handler(int){interrupted=1;}
static bool write_all(int fd,const char* p,size_t n){
while(n){ssize_t k=write(fd,p,n);if(k<0){if(errno==EINTR){if(interrupted)return false;continue;}return false;}if(!k)return false;p+=k;n-=k;}return true;
}
int main(int argc,char** argv){
if(argc>3){fprintf(stderr,"usage: serial_TXRX [device [rx_count]]\n");return 2;}
unsigned long long limit=ULLONG_MAX;
if(argc==3){std::string s=argv[2];if(s.empty()||s.find_first_not_of("0123456789")!=std::string::npos){fprintf(stderr,"invalid count\n");return 2;}errno=0;char* end=nullptr;limit=strtoull(argv[2],&end,10);if(errno||!limit||*end){fprintf(stderr,"invalid count\n");return 2;}}
int fd=open(argc>1?argv[1]:"/dev/ttyS0",O_RDWR|O_NOCTTY);
if(fd<0){perror("open serial");return 1;}
termios old{},cfg{};if(tcgetattr(fd,&old)){perror("tcgetattr");close(fd);return 1;}cfg=old;cfmakeraw(&cfg);cfsetispeed(&cfg,B115200);cfsetospeed(&cfg,B115200);
cfg.c_cflag=(cfg.c_cflag&~(CSIZE|PARENB|CSTOPB|CRTSCTS))|CS8|CLOCAL|CREAD;cfg.c_iflag&=~(IXON|IXOFF|IXANY);cfg.c_cc[VMIN]=1;cfg.c_cc[VTIME]=0;
if(tcsetattr(fd,TCSANOW,&cfg)){perror("tcsetattr");close(fd);return 1;}
struct sigaction sa{};sa.sa_handler=signal_handler;sigemptyset(&sa.sa_mask);sigaction(SIGINT,&sa,nullptr);
bool input=true;unsigned long long received=0;int status=0;
while(!interrupted&&received<limit){
pollfd fds[2]={{input?STDIN_FILENO:-1,POLLIN,0},{fd,POLLIN,0}};
int result=poll(fds,2,-1);if(result<0){if(errno==EINTR)continue;status=1;break;}
char buf[4096];
if(input&&(fds[0].revents&(POLLIN|POLLHUP))){ssize_t n=read(0,buf,sizeof(buf));if(n==0)input=false;else if(n>0){if(!write_all(1,buf,n)){if(!interrupted)status=1;break;}}else if(errno!=EINTR){status=1;break;}}
if(fds[1].revents&(POLLIN|POLLHUP|POLLERR)){
size_t capacity=sizeof(buf);if(limit-received<capacity)capacity=limit-received;
ssize_t n=read(fd,buf,capacity);if(n>0){if(!write_all(1,buf,n)){if(!interrupted)status=1;break;}received+=n;}else if(n==0){status=received==limit?0:1;break;}else if(errno!=EINTR){status=1;break;}
}
}
if(tcsetattr(fd,TCSANOW,&old)){perror("restore");status=1;}close(fd);return interrupted?0:status;
}

View File

@ -0,0 +1,16 @@
TARGETS=['serial_TXRX.cpp']
import pathlib, shutil, subprocess, sys, tempfile
own=pathlib.Path(__file__).resolve().parent
try:
with tempfile.TemporaryDirectory(dir='/tmp',prefix='author1-grade-') as d:
tmp=pathlib.Path(d)
for name in TARGETS:
p=pathlib.Path(sys.argv[1])/name
if not p.is_file() or p.is_symlink():raise ValueError('missing target '+name)
shutil.copyfile(p,tmp/name)
shutil.copyfile(own/'tests.py',tmp/'_tests.py')
p=subprocess.run([sys.executable,'-I','-B',str(tmp/'_tests.py')],cwd=tmp,text=True,capture_output=True,timeout=25)
if p.returncode or p.stdout.strip()!='VERIFIED_ALL_CASES':raise ValueError((p.stdout+p.stderr)[-900:])
print('PASS')
except Exception as exc:
print('FAIL:',str(exc).replace(str(tmp),'<scratch>') if 'tmp' in globals() else str(exc));sys.exit(1)

View File

@ -0,0 +1,40 @@
#include <termios.h>
#include <unistd.h>
#include <fcntl.h>
#include <poll.h>
#include <signal.h>
#include <cerrno>
#include <cstdlib>
#include <cstdio>
#include <cstring>
#include <climits>
#include <string>
static volatile sig_atomic_t interrupted=0;
static void signal_handler(int){interrupted=1;}
static bool write_all(int fd,const char* p,size_t n){
while(n){ssize_t k=write(fd,p,n);if(k<0){if(errno==EINTR){if(interrupted)return false;continue;}return false;}if(!k)return false;p+=k;n-=k;}return true;
}
int main(int argc,char** argv){
if(argc>3){fprintf(stderr,"usage: serial_TXRX [device [rx_count]]\n");return 2;}
unsigned long long limit=ULLONG_MAX;
if(argc==3){std::string s=argv[2];if(s.empty()||s.find_first_not_of("0123456789")!=std::string::npos){fprintf(stderr,"invalid count\n");return 2;}errno=0;char* end=nullptr;limit=strtoull(argv[2],&end,10);if(errno||!limit||*end){fprintf(stderr,"invalid count\n");return 2;}}
int fd=open(argc>1?argv[1]:"/dev/ttyS0",O_RDWR|O_NOCTTY);
if(fd<0){perror("open serial");return 1;}
termios old{},cfg{};if(tcgetattr(fd,&old)){perror("tcgetattr");close(fd);return 1;}cfg=old;cfmakeraw(&cfg);cfsetispeed(&cfg,B115200);cfsetospeed(&cfg,B115200);
cfg.c_cflag=(cfg.c_cflag&~(CSIZE|PARENB|CSTOPB|CRTSCTS))|CS8|CLOCAL|CREAD;cfg.c_iflag&=~(IXON|IXOFF|IXANY);cfg.c_cc[VMIN]=1;cfg.c_cc[VTIME]=0;
if(tcsetattr(fd,TCSANOW,&cfg)){perror("tcsetattr");close(fd);return 1;}
struct sigaction sa{};sa.sa_handler=signal_handler;sigemptyset(&sa.sa_mask);sigaction(SIGINT,&sa,nullptr);
bool input=true;unsigned long long received=0;int status=0;
while(!interrupted&&received<limit){
pollfd fds[2]={{input?STDIN_FILENO:-1,POLLIN,0},{fd,POLLIN,0}};
int result=poll(fds,2,-1);if(result<0){if(errno==EINTR)continue;status=1;break;}
char buf[4096];
if(input&&(fds[0].revents&(POLLIN|POLLHUP))){ssize_t n=read(0,buf,sizeof(buf));if(n==0)input=false;else if(n>0){if(!write_all(fd,buf,n)){if(!interrupted)status=1;break;}}else if(errno!=EINTR){status=1;break;}}
if(fds[1].revents&(POLLIN|POLLHUP|POLLERR)){
size_t capacity=sizeof(buf);if(limit-received<capacity)capacity=limit-received;
ssize_t n=read(fd,buf,capacity);if(n>0){if(!write_all(1,buf,n)){if(!interrupted)status=1;break;}received+=n;}else if(n==0){status=received==limit?0:1;break;}else if(errno!=EINTR){status=1;break;}
}
}
if(tcsetattr(fd,TCSANOW,&old)){perror("restore");status=1;}close(fd);return interrupted?0:status;
}

View File

@ -0,0 +1 @@
// Implement README.md.

View File

@ -0,0 +1,3 @@
当前目录是待实现的空工程。请按 README.md 完成C++双向串口服务与配置恢复,并自行验证。
最终只提交并评分以下文件:`serial_TXRX.cpp`。

View File

@ -0,0 +1,95 @@
# 任务规格:S500_a1_03167860845e
本规格按[任务制作规范](../../docs/task-authoring.md)记录路线A导入审阅;不得提供给Agent。
## 1. 身份与来源
| 字段 | 填写 |
| --- | --- |
| task-id / 任务版本 | S500_a1_03167860845e / 1 |
| 作者 / 业务评审人 / 独立评审人 | 本批Codex(导入);业务评审未单独指定 / Codex review_1 |
| 审核状态 | ADMITTED |
| 原项目与冻结版本 | Bos-qiao task_factory_500_20260914;本批本地只读复制快照 `factory/20261014-A-bos-direct/bos-envs/S500_a1_03167860845e/`;原task.json SHA-256 `728c0e1ec3afaa51bc7b80d5b14a420269932ee4f48e03f365bbaa5319ce8549` |
| 来源路线 | A 导入成熟任务;原创新建软件任务(bos.origin未声明;依据公开README的新建空工程声明,不标为历史现场) |
| 来源 ID | `S500_a1_03167860845e`;source_qid=`q_d37292e903167860845e` |
| 来源与授权 | 原来源 `/mnt/data/ming01.wang/Bos-qiao/envs/S500_a1_03167860845e`;依用户2026-10-12接入/试跑授权与本次作业A授权,只保留导入软件题材料与控制侧证据 |
| 隐私处理 | 无真实用户会话原文;不读/复制私有凭据;公开材料仅任务软件契约和部署路径 |
| 替身测试 | 软件:在空工程实现独立TX/RX方向、stdin EOF后保留RX、精确rx_count退出和SIGINT时termios恢复;原公开材料明确为新建代码需求。 用合成输入、受控回调/PTY或本机只读软件接口可复现;不依赖实体外设 |
| split 分组键 | `bos-qiao/task_factory_500_20260914/envs/S500_a1_03167860845e`(同源变体使用同组,不把重复采样算新题) |
| 来源组 / 根因族 | Bos-qiao/q_d37292e903167860845e / `serial-bidirectional-forwarding-and-termios-restore` |
| 与已有任务的关系 | 独立问题;在本批与既有题的公开需求/根因对照中未发现仅参数或问法替换的重复。 |
| 验证范围 | 下列验收矩阵中的确定软件行为;family=`linux-device-abi`;最终仅冻结 `serial_TXRX.cpp` |
| 未验证范围 | 未单独注入每种短write/EINTR;未断言CLOCAL/CREAD与CRTSCTS各位,公开要求仍存在。PTY软件时序测试使用秒级超时;不证明实体UART、电气、波特率或性能。 |
实际问题说明:在空工程实现独立TX/RX方向、stdin EOF后保留RX、精确rx_count退出和SIGINT时termios恢复;原公开材料明确为新建代码需求。
## 2. 初态与环境
沿用 Bos 原题。
## 3. 公开题面与接口
沿用 Bos 原题。
## 4. 验收矩阵与合法解
| 需求 | 输入/场景 | 可观察结果与允许误差 | 检查方法 | 公开/受保护 | 超时 |
| --- | --- | --- | --- | --- | --- |
| 双向二进制转发 | 种子501;TX长度1/37/90000,RX长度分别114/150/90113,RX与TX载荷不同 | PTY端收到TX原字节;stdout恰为RX,退出0 | 独立生成TX/RX,字节完全相等,10秒communicate | 规则公开;具体输入/期望受保护 | 评分总120秒;内部25/30秒 |
| raw配置与退出恢复 | 服务运行时读取PTY termios;rx_count正常结束 | 115200/8N1;无ICANON/ECHO/OPOST/IXON/IXOFF/ICRNL/INLCR;退出后原termios全相等(忽略PENDIN) | termios快照断言 | 规则公开;具体输入/期望受保护 | 评分总120秒;内部25/30秒 |
| 非法CLI | 缺失设备;count=0/-1/1x;多余参数 | 非0退出且stderr非空 | 3秒subprocess断言 | 规则公开;具体输入/期望受保护 | 评分总120秒;内部25/30秒 |
| 持续模式安全退出 | 仅device参数,stdin发送ready后SIGINT | PTY得到ready;退出0、stdout空、原配置恢复 | 受控SIGINT+3秒退出 | 规则公开;具体输入/期望受保护 | 评分总120秒;内部25/30秒 |
- 合法替代解:既有valid_alternative以733字节缓冲区替代gold的4096字节;select/poll或受控线程均可,只要满足独立双向、short read/write与恢复语义。
- 不允许的捷径:cheat在stdin EOF直接结束,被EOF后的RX输入抓住;cheat_stdin_echo把TX写stdout,被PTY端TX和stdout独立RX双重比较抓住。
- 独立复算与冻结:check.py将声明目标复制到干净/tmp再运行/编译;正式期望来自公开约定的独立字节/数学/受控事件计算,不运行gold求答案,不复用Agent中间产物。判据附件为 `check.py`, `tests.py`。
- 判断范围与覆盖限制:未单独注入每种短write/EINTR;未断言CLOCAL/CREAD与CRTSCTS各位,公开要求仍存在。PTY软件时序测试使用秒级超时;不证明实体UART、电气、波特率或性能。
- grader输出:统一verify.py返回一行JSON;PASS/FAIL checks=1、退出0/1,环境/基础设施错误为UNVERIFIED/INFRA_ERROR、退出2;内部行为断言全部完成才PASS。
- 隔离复核:沿用固定runtime的Agent/评分分容器、candidate冻结只读与UID65534探针;Docker校准回执才证明正式探针及清理。verify.py内已有同UID原checker与候选共进程的shortcut,不能据此宣称已具备独立UID的恶意checker防篡改。
<!-- shortcut: 现有Bos评分包装器的check.py与候选同UID且共享checker进程,正式UID探针不等于完整恶意候选防篡改;升级条件为独立checker身份/进程边界合同准入。 -->
- 历史答案可见性:Agent仅得到prompt与template,spec/grader/gold不放入公开工作区;历史归档不由本题挂载。spec不进入task digest。
## 5. 四类样例与校准
沿用 Bos 原题。
## 6. 真实试采与证据
沿用 Bos 原题。
## 7. 评审与准入
- [x] 独立内容评审完成;来源、替身测试、公开行为、判据、合法解、负例及局限见第 1、4 节和下方采样前记录。
- [x] 最终 task/verifier/image 身份的正式 Docker 校准,每类样例各五次符合预期,整体 PASS、cleanup confirmed。
- [x] GLM-5.3 / Pi 0.73.1 首次真实试采正常退出,host/native/broker 对账和工具反馈因果闭合 VERIFIED,清理 confirmed。
- [x] 冻结候选在新无网评分容器重判一致,按 rollout.archive 显式证据合同重新核对后本地封存。
- [x] 重复来源、根因、软件边界和现有隔离局限已登记;闸门 4 尚未执行。
最终审核状态:**ADMITTED**。任务失败与轨迹完整分开记录,首次任务 verdict 为 **PASS**,不据单次结果作难度分池。人工干预:只读诊断与控制侧评审;未向运行 Agent 注入判据/答案或修改候选。
最终身份:family `linux-device-abi`;task digest `140c1239781160127f9ec4338d71802f0df714f017ff3437627b888d3ab2acc3`;verifier digest `2ce99a20ed07a4b16716568ee1efbd08f35e09b1300468ab443896171bde1cfe`;固定 runtime / actual image ID `sha256:aa5d00779fe03e9eadbf29c030711a4ecd1e5f894255868b5ca451a425299efc`;模型 `glm-5.3`、每次回复 max_tokens=131072、sampling=service-default,单题 Agent 防挂死期限 3600 秒。
最终正式校准:`/mnt/data/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-calibration/calibration-13b464f05a9a44edadd4df99d269c7e2`;本地原执行路径 `/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-calibration/calibration-13b464f05a9a44edadd4df99d269c7e2`。首次 controller:`/mnt/data/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct/pi-b7ec8afb5a8d4d6dbdb7c47033a36775`;原路径 `/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct/pi-b7ec8afb5a8d4d6dbdb7c47033a36775`。证据 `VERIFIED`,Agent exit=0 / termination=exit,cleanup confirmed,独立重判一致。完整轨迹与独立评分见本批冷数据 `factory/20261014-A-bos-direct/report.json`、`admission.jsonl`。
本地封存包:`/mnt/data/ming01.wang/embedded-rollout/artifacts/20261014-A-bos-direct/pi-b7ec8afb5a8d4d6dbdb7c47033a36775.tar.gz`;SHA-256 `721118a4cd4a3f560704051c60ab59afdb0c6c5ab9732f4b44d2dac6795e3f4a`,2206407 字节。Salty 上传/下载摘要核对 **BLOCKED**:执行主机没有可用已认证 CLI 或连接;没有声明已上传。此阻塞不替代已完成的本地封存事实。冷迁按文件数与总字节核对,见批次 `migration-verification.json`;回执原字节和原绝对路径保持,当前入口不能直接对迁移后的路径重跑,需显式恢复原路径或升级路径合同。
### 采样前独立内容评审记录
下方保留采样前状态和当时的证据。若包含旧归族校准,只用于历史事实;当前准入由上述最终身份、校准和完整试采授权。
- [x] 2026-10-14,Codex review_1已逐项静态核对task/prompt/template、check及附件、gold/valid/全部cheat;未执行来源脚本、未改来源或平台。
- [x] 公开README与隐藏行为逐项对照;矩阵内受检正确行为均有公开依据,初态桩无gold泄露,目标自包含,软件替身范围明确。
- [x] root cause与来源组记录,已对照当前任务公开主题确认解题路径不同;合法替代解按行为而非内部补丁评审。
- [x] 固定镜像正式五次/样例校准整体PASS且cleanup confirmed(仅根据已见实际回执)。
- [ ] 闸门3真实试采的工具轨迹、独立评分、VERIFIED对账与清理:本评审尚未核对;由批次controller补齐,不能据本规格宣称ADMITTED。
- [ ] 闸门4≥5次同配置难度分池:本批作业A不执行。
样例合同:gold=PASS、template=FAIL、noop=FAIL、cheat=FAIL、cheat_stdin_echo=FAIL、valid_alternative=PASS;各类需正式重复5次,不能以静态审阅代替。
校准证据:
- 可信本地开发:`/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-dev/calibration-463a4c5089f949a991ca700cb400b899`;verdict=PASS,cleanup=confirmed,trials=30;task_digest=`140c1239781160127f9ec4338d71802f0df714f017ff3437627b888d3ab2acc3`;实际image_id=`None`。
- 固定镜像Docker:`/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-calibration/calibration-13b464f05a9a44edadd4df99d269c7e2`;verdict=PASS,cleanup=confirmed,trials=30;task_digest=`140c1239781160127f9ec4338d71802f0df714f017ff3437627b888d3ab2acc3`;实际image_id=`sha256:aa5d00779fe03e9eadbf29c030711a4ecd1e5f894255868b5ca451a425299efc`。
独立结论:APPROVED_FOR_CALIBRATION/PILOT(题目内容),人工生命周期状态为CALIBRATED,非ADMITTED。任务导入摘要 `140c1239781160127f9ec4338d71802f0df714f017ff3437627b888d3ab2acc3`;正式镜像固定为 `ccr-29eug8s3-pub.cnc.bj.baidubce.com/algo/bench-runtime@sha256:aa5d00779fe03e9eadbf29c030711a4ecd1e5f894255868b5ca451a425299efc`。限制:未单独注入每种短write/EINTR;未断言CLOCAL/CREAD与CRTSCTS各位,公开要求仍存在。PTY软件时序测试使用秒级超时;不证明实体UART、电气、波特率或性能。 后续以最终正式校准与真实试采证据更新本节。

View File

@ -0,0 +1,36 @@
{
"id": "S500_a1_03167860845e",
"version": "1",
"family": "linux-device-abi",
"source": "bos-qiao/task_factory_500_20260914/envs/S500_a1_03167860845e",
"template": "template",
"verifier": "verify.py",
"candidate_files": [
"serial_TXRX.cpp"
],
"timeout_seconds": 120,
"fixtures": {
"gold": "PASS",
"template": "FAIL",
"noop": "FAIL",
"cheat": "FAIL",
"cheat_stdin_echo": "FAIL",
"valid_alternative": "PASS"
},
"grader_files": [
"check.py",
"tests.py"
],
"bos": {
"source_qid": "q_d37292e903167860845e",
"origin": null,
"image": "rlenv/base@sha256:a00f7d220b2de8a5e3520d2d3eb3038dd80e5f10612ee9e53537e2f671abdef8",
"budget_s": 600,
"fixture_scripts": {
"gold": "gold.sh",
"cheat": "cheat_early_eof.sh",
"cheat_stdin_echo": "cheat_stdin_echo.sh",
"valid_alternative": "valid_alternative.sh"
}
}
}

View File

@ -0,0 +1,8 @@
# C++双向串口服务与配置恢复
这是依据原始新建代码需求创建的空标准工程,不是历史故障现场恢复。
实现 serial_TXRX.cpp,使用 g++ -std=c++17 -O2 serial_TXRX.cpp -o serial_TXRX 编译。程序是双向字节转发串口服务,stdin字节发往串口,串口字节写stdout,两方向数据互不替代;不是串口自回显。
CLI:serial_TXRX [device [rx_count]];device缺省/dev/ttyS0。rx_count可省略(持续服务直到SIGINT),若提供则为十进制正整数,累计从串口转发恰好rx_count字节后正常退出0。参数数量错误或非法count非0退出;设备不存在也非0退出,stderr给出可读错误。不输出调试信息到stdout。
串口公开固定115200、8数据位、无校验、1停止位、无软件/硬件流控,启用CLOCAL/CREAD;raw模式,禁止ICANON/ECHO/OPOST和输入字节转换。用termios保存原配置,退出时恢复。stdin EOF仅关闭stdin方向,仍等待串口数据;中断SIGINT正常退出0且恢复配置。必须支持二进制0字节、CR/LF、0xFF、大批量和短read/write,不安装库。
自测可使用Python标准库PTY:服务将stdin独立载荷TX发往PTY,测试方读取TX并向PTY注入不同RX,stdout必须仅为RX。只证明软件串口配置与转发,不证明实体UART、电气或实际波特率性能。
只依赖标准库;不访问硬件或网络。可在当前目录编写自己的自测;额外文件不影响判分。

View File

@ -0,0 +1 @@
// Implement README.md.

View File

@ -0,0 +1,36 @@
import sys, pathlib, random, math, json, struct
sys.path.insert(0,str(pathlib.Path(__file__).resolve().parent))
import subprocess,os,pty,termios,threading,select,signal
p=subprocess.run(['g++','-std=c++17','-O2','serial_TXRX.cpp','-o','serial_TXRX'],capture_output=True,text=True,timeout=15);assert p.returncode==0,p.stderr
exe=str(pathlib.Path('serial_TXRX').resolve());r=random.Random(501)
def attrs(fd):
a=termios.tcgetattr(fd);a[3]&=~getattr(termios,'PENDIN',0);return a
for size in [1,37,90000]:
master,slave=pty.openpty();name=os.ttyname(slave);before=attrs(slave)
tx=r.randbytes(size);rx=r.randbytes(size+113)
proc=subprocess.Popen([exe,name,str(len(rx))],stdin=subprocess.PIPE,stdout=subprocess.PIPE,stderr=subprocess.PIPE);result=[]
thread=threading.Thread(target=lambda:result.append(proc.communicate(tx,timeout=10)));thread.start()
got=b''
while len(got)<len(tx):
assert select.select([master],[],[],4)[0],'TX stalled'
got+=os.read(master,65536)
assert got==tx,'TX bytes changed'
cfg=termios.tcgetattr(slave)
assert not cfg[3]&(termios.ECHO|termios.ICANON) and not cfg[1]&termios.OPOST
assert not cfg[0]&(termios.IXON|termios.IXOFF|termios.ICRNL|termios.INLCR)
assert cfg[4]==termios.B115200 and cfg[5]==termios.B115200
assert cfg[2]&termios.CSIZE==termios.CS8 and not cfg[2]&(termios.PARENB|termios.CSTOPB)
pos=0
while pos<len(rx):pos+=os.write(master,rx[pos:pos+1003])
thread.join(10);assert not thread.is_alive() and proc.returncode==0,result
assert result[0][0]==rx,'RX bytes changed or replaced by TX'
assert attrs(slave)==before,'termios not restored'
os.close(master);os.close(slave)
for args in [['/no/such/device'],['/no/such/device','0'],['/no/such/device','-1'],['/no/such/device','1x'],['x','2','3']]:
p=subprocess.run([exe]+args,capture_output=True,timeout=3);assert p.returncode!=0 and p.stderr
master,slave=pty.openpty();before=attrs(slave);proc=subprocess.Popen([exe,os.ttyname(slave)],stdin=subprocess.PIPE,stdout=subprocess.PIPE,stderr=subprocess.PIPE)
proc.stdin.write(b'ready');proc.stdin.flush();assert select.select([master],[],[],3)[0];assert os.read(master,5)==b'ready'
proc.send_signal(signal.SIGINT);out,err=proc.communicate(timeout=3);assert proc.returncode==0 and not out and attrs(slave)==before
os.close(master);os.close(slave)
print('VERIFIED_ALL_CASES')

View File

@ -0,0 +1,40 @@
#include <termios.h>
#include <unistd.h>
#include <fcntl.h>
#include <poll.h>
#include <signal.h>
#include <cerrno>
#include <cstdlib>
#include <cstdio>
#include <cstring>
#include <climits>
#include <string>
static volatile sig_atomic_t interrupted=0;
static void signal_handler(int){interrupted=1;}
static bool write_all(int fd,const char* p,size_t n){
while(n){ssize_t k=write(fd,p,n);if(k<0){if(errno==EINTR){if(interrupted)return false;continue;}return false;}if(!k)return false;p+=k;n-=k;}return true;
}
int main(int argc,char** argv){
if(argc>3){fprintf(stderr,"usage: serial_TXRX [device [rx_count]]\n");return 2;}
unsigned long long limit=ULLONG_MAX;
if(argc==3){std::string s=argv[2];if(s.empty()||s.find_first_not_of("0123456789")!=std::string::npos){fprintf(stderr,"invalid count\n");return 2;}errno=0;char* end=nullptr;limit=strtoull(argv[2],&end,10);if(errno||!limit||*end){fprintf(stderr,"invalid count\n");return 2;}}
int fd=open(argc>1?argv[1]:"/dev/ttyS0",O_RDWR|O_NOCTTY);
if(fd<0){perror("open serial");return 1;}
termios old{},cfg{};if(tcgetattr(fd,&old)){perror("tcgetattr");close(fd);return 1;}cfg=old;cfmakeraw(&cfg);cfsetispeed(&cfg,B115200);cfsetospeed(&cfg,B115200);
cfg.c_cflag=(cfg.c_cflag&~(CSIZE|PARENB|CSTOPB|CRTSCTS))|CS8|CLOCAL|CREAD;cfg.c_iflag&=~(IXON|IXOFF|IXANY);cfg.c_cc[VMIN]=1;cfg.c_cc[VTIME]=0;
if(tcsetattr(fd,TCSANOW,&cfg)){perror("tcsetattr");close(fd);return 1;}
struct sigaction sa{};sa.sa_handler=signal_handler;sigemptyset(&sa.sa_mask);sigaction(SIGINT,&sa,nullptr);
bool input=true;unsigned long long received=0;int status=0;
while(!interrupted&&received<limit){
pollfd fds[2]={{input?STDIN_FILENO:-1,POLLIN,0},{fd,POLLIN,0}};
int result=poll(fds,2,-1);if(result<0){if(errno==EINTR)continue;status=1;break;}
char buf[733];
if(input&&(fds[0].revents&(POLLIN|POLLHUP))){ssize_t n=read(0,buf,sizeof(buf));if(n==0)input=false;else if(n>0){if(!write_all(fd,buf,n)){if(!interrupted)status=1;break;}}else if(errno!=EINTR){status=1;break;}}
if(fds[1].revents&(POLLIN|POLLHUP|POLLERR)){
size_t capacity=sizeof(buf);if(limit-received<capacity)capacity=limit-received;
ssize_t n=read(fd,buf,capacity);if(n>0){if(!write_all(1,buf,n)){if(!interrupted)status=1;break;}received+=n;}else if(n==0){status=received==limit?0:1;break;}else if(errno!=EINTR){status=1;break;}
}
}
if(tcsetattr(fd,TCSANOW,&old)){perror("restore");status=1;}close(fd);return interrupted?0:status;
}

View File

@ -0,0 +1,93 @@
"""Run an imported Bos-qiao check.py unchanged, with the candidate user's privileges."""
import json
import os
from pathlib import Path
import shutil
import subprocess
import sys
import tempfile
GRADER = Path(__file__).resolve().parent
# Runs the original check unchanged; failing child processes also copy their stderr tail out for private triage.
CHECK = (
"import runpy, subprocess, sys\n"
"real = subprocess.run\n"
"def run(*args, **kwargs):\n"
" result = real(*args, **kwargs)\n"
" if result.returncode and result.stderr:\n"
" tail = result.stderr if isinstance(result.stderr, str) else result.stderr.decode('utf-8', 'replace')\n"
" sys.__stderr__.write(tail[-2000:] + '\\n')\n"
" return result\n"
"subprocess.run = run\n"
"sys.argv = sys.argv[1:]\n"
"runpy.run_path(sys.argv[0], run_name='__main__')\n")
class UnsupportedEnvironment(RuntimeError):
pass
def candidate_user():
os.setgroups([])
os.setgid(65534)
os.setuid(65534)
def isolation():
if os.environ.get("ROLLOUT_GRADE_ISOLATION") != "uid65534":
return None
if sys.platform != "linux" or os.geteuid() != 0:
raise UnsupportedEnvironment("isolated grader requires a Linux root supervisor")
probe = (
"from pathlib import Path\n"
"try: Path('/grader/verify.py').read_bytes()\n"
"except PermissionError: raise SystemExit(0)\n"
"raise SystemExit(1)\n"
)
try:
result = subprocess.run([sys.executable, "-I", "-c", probe],
capture_output=True, timeout=10, preexec_fn=candidate_user)
except subprocess.SubprocessError as error:
raise UnsupportedEnvironment("candidate identity could not be isolated") from error
if result.returncode != 0:
raise UnsupportedEnvironment("candidate can read the private grader or isolation probe failed")
return candidate_user
def verify(candidate):
child_user = isolation()
if shutil.which("gcc") is None:
raise UnsupportedEnvironment("C compiler unavailable")
with tempfile.TemporaryDirectory(prefix="bos-grade-") as temporary:
root = Path(temporary)
work, grader = root / "work", root / "grader"
shutil.copytree(candidate, work, symlinks=True)
shutil.copytree(GRADER, grader, ignore=lambda folder, names: ["verify.py"] if Path(folder) == GRADER else [])
if child_user:
os.chmod(root, 0o700) # Without CAP_FOWNER, root cannot chmod after giving files away.
for path in (root, *root.rglob("*")):
os.lchown(path, 65534, 65534)
# shortcut: the original check and the candidate share UID 65534, so a hostile candidate could
# tamper with its checker process; upgrade by running check.py and its tests under separate UIDs.
completed = subprocess.run([sys.executable, "-I", "-B", "-c", CHECK, str(grader / "check.py"), str(work)],
cwd=root, capture_output=True, text=True, timeout=100, preexec_fn=child_user)
# grade.stderr stays on the trusted side with the archive; it may quote hidden test lines.
sys.stderr.write(completed.stderr[-4000:])
lines = completed.stdout.splitlines()
if completed.returncode == 0 and lines[-1:] == ["PASS"]:
return {"status": "PASS", "checks": 1}
return {"status": "FAIL", "checks": 1, "reason": (lines[-1] if lines else "no checker output")[:300]}
if __name__ == "__main__":
try:
result = verify(Path(sys.argv[1]))
except UnsupportedEnvironment as error:
result = {"status": "UNVERIFIED", "checks": 0, "reason": str(error)}
except subprocess.TimeoutExpired:
result = {"status": "FAIL", "checks": 1, "reason": "timeout"}
except Exception as error:
result = {"status": "INFRA_ERROR", "checks": 0, "reason": str(error)}
print(json.dumps(result, sort_keys=True))
sys.exit({"PASS": 0, "FAIL": 1}.get(result["status"], 2))

View File

@ -0,0 +1,47 @@
import struct,json,math,socket
MAGIC=bytes.fromhex('aa55bb77')
def encode(type_id,payload):
if type(type_id) is not int or not 0<=type_id<=255 or not isinstance(payload,bytes) or len(payload)>4096:raise ValueError('frame')
return MAGIC+struct.pack('<BI',type_id,len(payload))+payload
class Decoder:
def __init__(self):self.buffer=b''
def feed(self,chunk):
self.buffer=chunk;out=[]
while len(self.buffer)>=4:
i=self.buffer.find(MAGIC)
if i<0:self.buffer=self.buffer[-3:];break
self.buffer=self.buffer[i:]
if len(self.buffer)<9:break
t=self.buffer[4];n=int.from_bytes(self.buffer[5:9],'little')
if n>4096:self.buffer=self.buffer[1:];continue
if len(self.buffer)<9+n:break
out.append((t,self.buffer[9:9+n]));self.buffer=self.buffer[9+n:]
return out
def dispatch(type_id,payload,on_line_follow,on_velocity,on_resend):
if type_id==126:
if payload:return False
on_resend();return True
if type_id not in (99,100):return False
try:d=json.loads(payload.decode('utf8'))
except (ValueError,UnicodeError):return False
if not isinstance(d,dict):return False
if type_id==99:
if d.get('cmd')!='line_follow' or type(d.get('enable')) is not bool:return False
on_line_follow(d['enable'])
else:
x=d.get('linear_x');z=d.get('angular_z')
if any(type(v) not in (int,float) or not math.isfinite(v) for v in (x,z)):return False
on_velocity(float(x),float(z))
return True
def serve_once(on_line_follow,on_velocity,on_resend,ready=None):
with socket.socket(socket.AF_INET,socket.SOCK_STREAM) as listener:
listener.setsockopt(socket.SOL_SOCKET,socket.SO_REUSEADDR,1);listener.bind(('127.0.0.1',9999));listener.listen(1)
if ready:ready()
conn,_=listener.accept()
with conn:
decoder=Decoder()
while True:
chunk=conn.recv(1024)
if not chunk:break
for t,p in decoder.feed(chunk):dispatch(t,p,on_line_follow,on_velocity,on_resend)

View File

@ -0,0 +1,47 @@
import struct,json,math,socket
MAGIC=bytes.fromhex('aa55bb77')
def encode(type_id,payload):
if type(type_id) is not int or not 0<=type_id<=255 or not isinstance(payload,bytes) or len(payload)>4096:raise ValueError('frame')
return MAGIC+struct.pack('<BI',type_id,len(payload))+payload
class Decoder:
def __init__(self):self.buffer=b''
def feed(self,chunk):
self.buffer+=chunk;out=[]
while len(self.buffer)>=4:
i=self.buffer.find(MAGIC)
if i<0:self.buffer=self.buffer[-3:];break
self.buffer=self.buffer[i:]
if len(self.buffer)<9:break
t=self.buffer[4];n=int.from_bytes(self.buffer[5:9],'little')
if n>4096:self.buffer=self.buffer[1:];continue
if len(self.buffer)<9+n:break
out.append((t,self.buffer[9:9+n]));self.buffer=self.buffer[9+n:]
return out
def dispatch(type_id,payload,on_line_follow,on_velocity,on_resend):
if type_id==126:
if payload:return False
on_resend();return True
if type_id not in (99,100):return False
try:d=json.loads(payload.decode('utf8'))
except (ValueError,UnicodeError):return False
if not isinstance(d,dict):return False
if type_id==99:
if d.get('cmd')!='line_follow' or 'enable' not in d:return False
on_line_follow(bool(d['enable']))
else:
x=d.get('linear_x');z=d.get('angular_z')
if any(type(v) not in (int,float) or not math.isfinite(v) for v in (x,z)):return False
on_velocity(float(x),float(z))
return True
def serve_once(on_line_follow,on_velocity,on_resend,ready=None):
with socket.socket(socket.AF_INET,socket.SOCK_STREAM) as listener:
listener.setsockopt(socket.SOL_SOCKET,socket.SO_REUSEADDR,1);listener.bind(('127.0.0.1',9999));listener.listen(1)
if ready:ready()
conn,_=listener.accept()
with conn:
decoder=Decoder()
while True:
chunk=conn.recv(1024)
if not chunk:break
for t,p in decoder.feed(chunk):dispatch(t,p,on_line_follow,on_velocity,on_resend)

View File

@ -0,0 +1,16 @@
TARGETS=['tcp_bridge.py']
import pathlib, shutil, subprocess, sys, tempfile
own=pathlib.Path(__file__).resolve().parent
try:
with tempfile.TemporaryDirectory(dir='/tmp',prefix='author1-grade-') as d:
tmp=pathlib.Path(d)
for name in TARGETS:
p=pathlib.Path(sys.argv[1])/name
if not p.is_file() or p.is_symlink():raise ValueError('missing target '+name)
shutil.copyfile(p,tmp/name)
shutil.copyfile(own/'tests.py',tmp/'_tests.py')
p=subprocess.run([sys.executable,'-I','-B',str(tmp/'_tests.py')],cwd=tmp,text=True,capture_output=True,timeout=25)
if p.returncode or p.stdout.strip()!='VERIFIED_ALL_CASES':raise ValueError((p.stdout+p.stderr)[-900:])
print('PASS')
except Exception as exc:
print('FAIL:',str(exc).replace(str(tmp),'<scratch>') if 'tmp' in globals() else str(exc));sys.exit(1)

View File

@ -0,0 +1,47 @@
import struct,json,math,socket
MAGIC=bytes.fromhex('aa55bb77')
def encode(type_id,payload):
if type(type_id) is not int or not 0<=type_id<=255 or not isinstance(payload,bytes) or len(payload)>4096:raise ValueError('frame')
return MAGIC+struct.pack('<BI',type_id,len(payload))+payload
class Decoder:
def __init__(self):self.buffer=b''
def feed(self,chunk):
self.buffer+=chunk;out=[]
while len(self.buffer)>=4:
i=self.buffer.find(MAGIC)
if i<0:self.buffer=self.buffer[-3:];break
self.buffer=self.buffer[i:]
if len(self.buffer)<9:break
t=self.buffer[4];n=int.from_bytes(self.buffer[5:9],'little')
if n>4096:self.buffer=self.buffer[1:];continue
if len(self.buffer)<9+n:break
out.append((t,self.buffer[9:9+n]));self.buffer=self.buffer[9+n:]
return out
def dispatch(type_id,payload,on_line_follow,on_velocity,on_resend):
if type_id==126:
if payload:return False
on_resend();return True
if type_id not in (99,100):return False
try:d=json.loads(payload.decode('utf8'))
except (ValueError,UnicodeError):return False
if not isinstance(d,dict):return False
if type_id==99:
if d.get('cmd')!='line_follow' or type(d.get('enable')) is not bool:return False
on_line_follow(d['enable'])
else:
x=d.get('linear_x');z=d.get('angular_z')
if any(type(v) not in (int,float) or not math.isfinite(v) for v in (x,z)):return False
on_velocity(float(x),float(z))
return True
def serve_once(on_line_follow,on_velocity,on_resend,ready=None):
with socket.socket(socket.AF_INET,socket.SOCK_STREAM) as listener:
listener.setsockopt(socket.SOL_SOCKET,socket.SO_REUSEADDR,1);listener.bind(('127.0.0.1',9999));listener.listen(1)
if ready:ready()
conn,_=listener.accept()
with conn:
decoder=Decoder()
while True:
chunk=conn.recv(1024)
if not chunk:break
for t,p in decoder.feed(chunk):dispatch(t,p,on_line_follow,on_velocity,on_resend)

View File

@ -0,0 +1,2 @@
"""Implement README.md."""
raise NotImplementedError

View File

@ -0,0 +1,3 @@
当前目录是待实现的空工程。请按 README.md 完成TCP协议流拆帧与控制回调服务,并自行验证。
最终只提交并评分以下文件:`tcp_bridge.py`。

View File

@ -0,0 +1,92 @@
# 任务规格:S500_a1_2718bf3679a5
本规格仅供控制侧作者与评审,不进入 Agent。规范见 [任务制作规范](../../docs/task-authoring.md)。路线 A 仅填写第 1、4、7 节。
## 1. 身份与来源
| 字段 | 填写 |
| --- | --- |
| task-id / 任务版本 | S500_a1_2718bf3679a5 / 1 |
| 作者 / 业务评审人 / 独立评审人 | Bos 原作者;本批导入 Codex / 沿用 Bos 原题 / Codex review_0 |
| 审核状态 | ADMITTED |
| 原项目与冻结版本 | Bos-qiao task_factory_500_20260914;本批本地只读复制快照 `factory/20261014-A-bos-direct/bos-envs/S500_a1_2718bf3679a5/`;原 task.json SHA-256 `06abeee7b3c9e23f4a2b75547ea138ce4a96d8cefd222f754527acb3a1c39fcf`;未获得独立上游 commit |
| 来源路线 | A 导入成熟任务;历史需求种子的新建工程;bos.origin=null,README明确不是历史故障恢复 |
| 来源 ID | S500_a1_2718bf3679a5;source_qid=q_4efbda322718bf3679a5 |
| 来源与授权 | bos-qiao/task_factory_500_20260914/envs/S500_a1_2718bf3679a5;用户 2026-10-12 授权本项目资源与试跑、2026-10-14 明确授权 Bos direct 导入。仅保留授权任务资产,不执行来源样例脚本 |
| 隐私处理 | 沿用 Bos 已公开软件材料;本批未引入用户会话原文,未读取凭据;历史来源标记不扩张为新回放证据 |
| 替身测试 | software:AA55BB77流编码/增量解码、严格控制回调、一次loopback TCP会话可以用合成输入/受控桩复现,不依赖真实设备 |
| split 分组键 | bos:q_4efbda322718bf3679a5 |
| 来源组 / 根因族 | Bos direct / tcp-stream-resynchronization-dispatch |
| 与已有任务的关系 | 与既有任务/本批目录题面标题和行为比对,未发现仅换参数/问法的同根因重复;同大族仍须保持同源split键 |
| 验证范围 | AA55BB77流编码/增量解码、严格控制回调、一次loopback TCP会话 |
| 未验证范围 | ROS/BLE/JPEG/地图、实体设备、生产网络时序 |
实际问题说明:实现TCP字节流拆帧及软件控制回调;核心是跨recv边界保留魔数、坏长度重同步和严格JSON字段分发。
## 2. 初态与环境
沿用 Bos 原题。
## 3. 公开题面与接口
沿用 Bos 原题。
## 4. 验收矩阵与合法解
| 需求 | 输入/场景 | 可观察结果与允许误差 | 检查方法 | 公开/受保护 | 超时 |
| --- | --- | --- | --- | --- | --- |
| 编码边界 | type=0..255,payload长度0/1/3/100/4096 | 头9字节:AA55BB77+单字节type+uint32小端长度;载荷逐字节相等 | 240帧独立组装对比 | 行为公开于 README;输入组合受保护 | 评分总120秒;探针25/30秒以内 |
| 分片/粘包/实例隔离 | noise前缀、随机1..299字节块、两个独立Decoder、空feed | 返回全部原(type,payload)有序列表;半个魔数保留,空feed无新帧 | 随机流与跨实例分片断言 | 行为公开于 README;输入组合受保护 | 评分总120秒;探针25/30秒以内 |
| 坏长度重同步 | 长度0xffffffff坏头后接合法帧 | 丢1字节重同步后仍返回后续合法帧 | 公开策略的坏头组合 | 行为公开于 README;输入组合受保护 | 评分总120秒;探针25/30秒以内 |
| 编码拒绝 | True/-1/256 type;字符串payload;4097字节 | ValueError | 逐项异常断言 | 行为公开于 README;输入组合受保护 | 评分总120秒;探针25/30秒以内 |
| 严格分发与回调异常 | 99 line_follow bool;100有限数;126空;未知/坏UTF8/NaN/bool数 | 成功True且对应单次回调;非法False且无回调;回调RuntimeError原样传播 | 受控事件列表及异常断言 | 行为公开于 README;输入组合受保护 | 评分总120秒;探针25/30秒以内 |
| TCP生命周期 | listen后ready;单客户端每7字节发送4种帧×30后半关闭 | 依序120个回调,无ACK要求,EOF退出None并关闭socket | loopback真实连接与线程join | 行为公开于 README;输入组合受保护 | 评分总120秒;探针25/30秒以内 |
- 独立判据:tests.py用struct/字节拼接独立产生240个帧并按随机分片发送,逐项比较返回帧和受控回调事件;真实loopback仅验证公开服务生命周期。
- 合法替代解:valid_alternative改用struct.unpack读取长度;也允许bytearray、状态机等不同实现。判据不查源码风格。
- 不允许的捷径:cheat丢弃跨feed碎片由随机分片/实例隔离拦截;cheat_truthy_enable把1转bool由非法enable无回调拦截;noop无完整API被执行检查拒绝。
- 判断范围:必须通过实际行为;编译/导入成功不等于通过。ROS/BLE/JPEG/地图、实体设备、生产网络时序不在验收结论内。
- grader 输出:统一 verify.py 包装为一行 JSON;PASS/FAIL checks>0,退出 0/1;UNVERIFIED/INFRA_ERROR 退出 2。内部 `check.py` 必须退出0且末行PASS才成功。
- 隔离复核:沿用 bos_verify.py 的受保护grader/UID65534执行探针及全新 `/tmp` 副本;正式镜像使用 trace5 digest。正式校准已核实固定镜像身份与cleanup confirmed;不可读 /grader 探针沿用wrapper,临时checker的同UID局限见shortcut;可信本地结果不证明隔离。
- shortcut: 原Bos check及候选共享UID65534,敌对候选可读取或干扰其checker进程;本批沿用已有边界,不声称全面防恶意隔离,升级条件是check与候选分UID及受控结果通道。
- 历史答案可见性:Agent只获公开prompt/template;gold、正式测试、本规格、factory/runs和历史候选均不挂进工作容器。
- 覆盖局限:固定端口9999及3秒ready/join仅覆盖隔离loopback环境;数值极大JSON、全部非法字段组合及网络异常分支未穷举。
## 5. 四类样例与校准
沿用 Bos 原题。
## 6. 真实试采与证据
沿用 Bos 原题。
## 7. 评审与准入
- [x] 独立内容评审完成;来源、替身测试、公开行为、判据、合法解、负例及局限见第 1、4 节和下方采样前记录。
- [x] 最终 task/verifier/image 身份的正式 Docker 校准,每类样例各五次符合预期,整体 PASS、cleanup confirmed。
- [x] GLM-5.3 / Pi 0.73.1 首次真实试采正常退出,host/native/broker 对账和工具反馈因果闭合 VERIFIED,清理 confirmed。
- [x] 冻结候选在新无网评分容器重判一致,按 rollout.archive 显式证据合同重新核对后本地封存。
- [x] 重复来源、根因、软件边界和现有隔离局限已登记;闸门 4 尚未执行。
最终审核状态:**ADMITTED**。任务失败与轨迹完整分开记录,首次任务 verdict 为 **PASS**,不据单次结果作难度分池。人工干预:只读诊断与控制侧评审;未向运行 Agent 注入判据/答案或修改候选。
最终身份:family `motor-sensor-control`;task digest `bf229f45c24e27a78d337fdbce4110c9f3877cc2902babc37befeb6991919494`;verifier digest `2ce99a20ed07a4b16716568ee1efbd08f35e09b1300468ab443896171bde1cfe`;固定 runtime / actual image ID `sha256:aa5d00779fe03e9eadbf29c030711a4ecd1e5f894255868b5ca451a425299efc`;模型 `glm-5.3`、每次回复 max_tokens=131072、sampling=service-default,单题 Agent 防挂死期限 3600 秒。
最终正式校准:`/mnt/data/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-calibration/calibration-e7c67701beaf43df863fd1bcd5b5eb80`;本地原执行路径 `/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-calibration/calibration-e7c67701beaf43df863fd1bcd5b5eb80`。首次 controller:`/mnt/data/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct/pi-4f252f5b8f3042dfb672917430c18dc5`;原路径 `/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct/pi-4f252f5b8f3042dfb672917430c18dc5`。证据 `VERIFIED`,Agent exit=0 / termination=exit,cleanup confirmed,独立重判一致。完整轨迹与独立评分见本批冷数据 `factory/20261014-A-bos-direct/report.json`、`admission.jsonl`。
本地封存包:`/mnt/data/ming01.wang/embedded-rollout/artifacts/20261014-A-bos-direct/pi-4f252f5b8f3042dfb672917430c18dc5.tar.gz`;SHA-256 `8985222c2c03c31079a1f0241ddf63094bc7db76b37621de8332ce43917486ec`,920342 字节。Salty 上传/下载摘要核对 **BLOCKED**:执行主机没有可用已认证 CLI 或连接;没有声明已上传。此阻塞不替代已完成的本地封存事实。冷迁按文件数与总字节核对,见批次 `migration-verification.json`;回执原字节和原绝对路径保持,当前入口不能直接对迁移后的路径重跑,需显式恢复原路径或升级路径合同。
### 采样前独立内容评审记录
下方保留采样前状态和当时的证据。若包含旧归族校准,只用于历史事实;当前准入由上述最终身份、校准和完整试采授权。
- [x] 独立只读核对本题 prompt、template、check/grader_files、gold、全部cheat与合法替代样例;来源、软件替身、公开约束及独立复算已核对。
- [x] 本批固定 trace5 Docker 全部声明样例各五次符合预期,整体PASS且cleanup confirmed,真实回执见下。
- [ ] task/verifier/image与真实试采一致,原生证据VERIFIED、独立评分和清理尚待闸门3。
- [x] 来源组、根因及未验证范围已记录。review_0未发现与当前任务列表同根因的仅参数变体;跨未导入候选全面去重尚不在本次已审集合。
- [x] 独立内容评审完成;本评审只读,不创建容器/网络或执行来源脚本。
审核记录:2026-10-14,内容评审 APPROVED,审核状态 CALIBRATED;作者本批 Codex,独立评审 Codex review_0。证据为本题冻结文件与 `factory/20261014-A-bos-direct/review-0.jsonl`;当前未发现规范 §3/§5/§6 的实质拒绝原因。正式校准已核实,真实试采和难度分池(每题>=5次)未完成,因此不得标记 ADMITTED。固定端口9999及3秒ready/join仅覆盖隔离loopback环境;数值极大JSON、全部非法字段组合及网络异常分支未穷举。
本批校准可见证据:可信本地 `/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-dev/calibration-8d6c24eb00cf419dabfb7954dbd8bf67` verdict=PASS(仅开发证据);正式 Docker `/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-calibration/calibration-e7c67701beaf43df863fd1bcd5b5eb80` / receipt.json:30次,6类×5,每次符合声明预期,cleanup confirmed;task digest `bf229f45c24e27a78d337fdbce4110c9f3877cc2902babc37befeb6991919494`,verifier digest `2ce99a20ed07a4b16716568ee1efbd08f35e09b1300468ab443896171bde1cfe`,实际image ID `sha256:aa5d00779fe03e9eadbf29c030711a4ecd1e5f894255868b5ca451a425299efc`;镜像引用采用本批trace5 digest。闸门3真实模型尚未完成,因此人工状态仍CALIBRATED,未ADMITTED。

View File

@ -0,0 +1,36 @@
{
"id": "S500_a1_2718bf3679a5",
"version": "1",
"family": "motor-sensor-control",
"source": "bos-qiao/task_factory_500_20260914/envs/S500_a1_2718bf3679a5",
"template": "template",
"verifier": "verify.py",
"candidate_files": [
"tcp_bridge.py"
],
"timeout_seconds": 120,
"fixtures": {
"gold": "PASS",
"template": "FAIL",
"noop": "FAIL",
"cheat": "FAIL",
"cheat_truthy_enable": "FAIL",
"valid_alternative": "PASS"
},
"grader_files": [
"check.py",
"tests.py"
],
"bos": {
"source_qid": "q_4efbda322718bf3679a5",
"origin": null,
"image": "python@sha256:78387bc3881b8273120a12ebe6c1ab22b018ccc2c9adf565ae1ac9b536e184ea",
"budget_s": 600,
"fixture_scripts": {
"gold": "gold.sh",
"cheat": "cheat_lost_fragments.sh",
"cheat_truthy_enable": "cheat_truthy_enable.sh",
"valid_alternative": "valid_alternative.sh"
}
}
}

View File

@ -0,0 +1,10 @@
# TCP协议流拆帧与控制回调服务
这是依据原始新建代码需求创建的空标准工程,不是历史故障现场恢复。
实现 tcp_bridge.py,纯Python标准库,不依赖rclpy。协议:魔数 AA55BB77 + type_id单字节 + payload长度uint32小端 + 原始payload,固定头9字节。
encode(type_id,payload:bytes)->bytes:type_id必须是0..255的int(拒绝bool),payload必须bytes且长度<=4096;非法ValueError。
Decoder.feed(chunk:bytes)->list[(int,bytes)]:支持任意分片/粘包,忽略魔数前噪声、保留跨chunk魔数前缀;载荷含魔数不影响拆帧。长度>4096属于坏头,丢弃一个字节后重新同步;未知type仍返回其原始payload;空载荷合法。不同实例独立。
dispatch(type_id,payload,on_line_follow,on_velocity,on_resend)->bool:0x63解析UTF-8 JSON对象,要求cmd=='line_follow'且enable严格bool;调用on_line_follow(enable)。0x64要求linear_x/angular_z为有限数字(拒绝bool),调用on_velocity(float(x),float(z))。0x7E只接受空payload,调用on_resend()。成功返回True;未知type、字段缺失/非法、坏JSON/UTF-8、非空重发返回False且无回调;额外JSON字段允许。回调异常原样传播。
serve_once(on_line_follow,on_velocity,on_resend,ready=None):TCP监听127.0.0.1:9999,listen后调用ready()(如提供);接受单个连接,循环recv直到EOF,用Decoder和dispatch驱动回调,关闭连接及监听socket并返回None。无需输出或发送ACK。本地软件测试从localhost:9999按协议发送分片,受控回调替代ROS publishers,不涉及BLE、JPEG生产、地图生成或实体设备。这是原seed明确允许的纯Python TCP拆帧/分发范围。
原文例子将JSON载荷标为42字节,与实际文本长度不一致;长度必须根据UTF-8实际bytes计算,不能写死42。名义其它帧类型01/02/03/04/05可由Decoder承载但dispatch忽略;服务器支持多个帧直到客户端半关闭写端。
只依赖标准库;不访问硬件或网络。可在当前目录编写自己的自测;额外文件不影响判分。

View File

@ -0,0 +1,2 @@
"""Implement README.md."""
raise NotImplementedError

View File

@ -0,0 +1,40 @@
import sys, pathlib, random, math, json, struct
sys.path.insert(0,str(pathlib.Path(__file__).resolve().parent))
import tcp_bridge as b,socket,threading
r=random.Random(455);frames=[];exp=[]
for i in range(240):
p=r.randbytes(r.choice([0,1,3,100,4096]));t=r.randrange(256);f=bytes.fromhex('aa55bb77')+bytes([t])+len(p).to_bytes(4,'little')+p
assert b.encode(t,p)==f;frames.append(f);exp.append((t,p))
d=b.Decoder();stream=b'noise'+b''.join(frames);got=[];i=0
while i<len(stream):
n=r.randrange(1,300);got.extend(d.feed(stream[i:i+n]));i+=n
assert got==exp and d.feed(b'')==[]
a=b.Decoder();c=b.Decoder();assert a.feed(frames[1][:3])==[];assert c.feed(frames[2])==[exp[2]];assert a.feed(frames[1][3:])==[exp[1]]
bad=bytes.fromhex('aa55bb77')+bytes([100])+bytes.fromhex('ffffffff')
assert b.Decoder().feed(bad+frames[0])==[exp[0]]
for t,p in [(True,b''),(-1,b''),(256,b''),(1,'x'),(1,b'x'*4097)]:
try:b.encode(t,p)
except ValueError:pass
else:raise AssertionError('invalid encode')
events=[]
cb=(lambda x:events.append(('line',x)),lambda x,z:events.append(('vel',x,z)),lambda:events.append(('resend',)))
valid=[(99,b'{"cmd":"line_follow","enable":true}',('line',True)),(99,b'{"cmd":"line_follow","enable":false,"extra":1}',('line',False)),(100,b'{"linear_x":0.08,"angular_z":-0.5}',('vel',.08,-.5)),(126,b'',('resend',))]
for t,p,e in valid:assert b.dispatch(t,p,*cb) is True and events[-1]==e
for t,p in [(100,b'{"linear_x":true,"angular_z":0}'),(99,b'{"cmd":"line_follow","enable":1}'),(99,b'[]'),(100,b'{}'),(100,b'{"linear_x":NaN,"angular_z":0}'),(100,b'\xff'),(126,b'x'),(1,b'{}')]:
old=list(events);assert b.dispatch(t,p,*cb) is False and events==old
try:b.dispatch(126,b'',lambda _:None,lambda *_:None,lambda:(_ for _ in ()).throw(RuntimeError('callback')))
except RuntimeError:pass
else:raise AssertionError('callback exception')
# Real isolated loopback TCP: ready event is listener lifecycle, no timing scoring.
events=[];ready=threading.Event();errors=[]
def server():
try:assert b.serve_once(*cb,ready=ready.set) is None
except BaseException as ex:errors.append(ex);ready.set()
th=threading.Thread(target=server,daemon=True);th.start();assert ready.wait(3) and not errors
with socket.create_connection(('127.0.0.1',9999),timeout=3) as client:
stream=b''.join(b.encode(t,p) for t,p,_ in valid)*30
for i in range(0,len(stream),7):client.sendall(stream[i:i+7])
client.shutdown(socket.SHUT_WR)
th.join(3);assert not th.is_alive() and not errors and events==[x[2] for x in valid]*30
print('VERIFIED_ALL_CASES')

View File

@ -0,0 +1,47 @@
import struct,json,math,socket
MAGIC=bytes.fromhex('aa55bb77')
def encode(type_id,payload):
if type(type_id) is not int or not 0<=type_id<=255 or not isinstance(payload,bytes) or len(payload)>4096:raise ValueError('frame')
return MAGIC+struct.pack('<BI',type_id,len(payload))+payload
class Decoder:
def __init__(self):self.buffer=b''
def feed(self,chunk):
self.buffer+=chunk;out=[]
while len(self.buffer)>=4:
i=self.buffer.find(MAGIC)
if i<0:self.buffer=self.buffer[-3:];break
self.buffer=self.buffer[i:]
if len(self.buffer)<9:break
t=self.buffer[4];n=struct.unpack('<I',self.buffer[5:9])[0]
if n>4096:self.buffer=self.buffer[1:];continue
if len(self.buffer)<9+n:break
out.append((t,self.buffer[9:9+n]));self.buffer=self.buffer[9+n:]
return out
def dispatch(type_id,payload,on_line_follow,on_velocity,on_resend):
if type_id==126:
if payload:return False
on_resend();return True
if type_id not in (99,100):return False
try:d=json.loads(payload.decode('utf8'))
except (ValueError,UnicodeError):return False
if not isinstance(d,dict):return False
if type_id==99:
if d.get('cmd')!='line_follow' or type(d.get('enable')) is not bool:return False
on_line_follow(d['enable'])
else:
x=d.get('linear_x');z=d.get('angular_z')
if any(type(v) not in (int,float) or not math.isfinite(v) for v in (x,z)):return False
on_velocity(float(x),float(z))
return True
def serve_once(on_line_follow,on_velocity,on_resend,ready=None):
with socket.socket(socket.AF_INET,socket.SOCK_STREAM) as listener:
listener.setsockopt(socket.SOL_SOCKET,socket.SO_REUSEADDR,1);listener.bind(('127.0.0.1',9999));listener.listen(1)
if ready:ready()
conn,_=listener.accept()
with conn:
decoder=Decoder()
while True:
chunk=conn.recv(1024)
if not chunk:break
for t,p in decoder.feed(chunk):dispatch(t,p,on_line_follow,on_velocity,on_resend)

View File

@ -0,0 +1,93 @@
"""Run an imported Bos-qiao check.py unchanged, with the candidate user's privileges."""
import json
import os
from pathlib import Path
import shutil
import subprocess
import sys
import tempfile
GRADER = Path(__file__).resolve().parent
# Runs the original check unchanged; failing child processes also copy their stderr tail out for private triage.
CHECK = (
"import runpy, subprocess, sys\n"
"real = subprocess.run\n"
"def run(*args, **kwargs):\n"
" result = real(*args, **kwargs)\n"
" if result.returncode and result.stderr:\n"
" tail = result.stderr if isinstance(result.stderr, str) else result.stderr.decode('utf-8', 'replace')\n"
" sys.__stderr__.write(tail[-2000:] + '\\n')\n"
" return result\n"
"subprocess.run = run\n"
"sys.argv = sys.argv[1:]\n"
"runpy.run_path(sys.argv[0], run_name='__main__')\n")
class UnsupportedEnvironment(RuntimeError):
pass
def candidate_user():
os.setgroups([])
os.setgid(65534)
os.setuid(65534)
def isolation():
if os.environ.get("ROLLOUT_GRADE_ISOLATION") != "uid65534":
return None
if sys.platform != "linux" or os.geteuid() != 0:
raise UnsupportedEnvironment("isolated grader requires a Linux root supervisor")
probe = (
"from pathlib import Path\n"
"try: Path('/grader/verify.py').read_bytes()\n"
"except PermissionError: raise SystemExit(0)\n"
"raise SystemExit(1)\n"
)
try:
result = subprocess.run([sys.executable, "-I", "-c", probe],
capture_output=True, timeout=10, preexec_fn=candidate_user)
except subprocess.SubprocessError as error:
raise UnsupportedEnvironment("candidate identity could not be isolated") from error
if result.returncode != 0:
raise UnsupportedEnvironment("candidate can read the private grader or isolation probe failed")
return candidate_user
def verify(candidate):
child_user = isolation()
if shutil.which("gcc") is None:
raise UnsupportedEnvironment("C compiler unavailable")
with tempfile.TemporaryDirectory(prefix="bos-grade-") as temporary:
root = Path(temporary)
work, grader = root / "work", root / "grader"
shutil.copytree(candidate, work, symlinks=True)
shutil.copytree(GRADER, grader, ignore=lambda folder, names: ["verify.py"] if Path(folder) == GRADER else [])
if child_user:
os.chmod(root, 0o700) # Without CAP_FOWNER, root cannot chmod after giving files away.
for path in (root, *root.rglob("*")):
os.lchown(path, 65534, 65534)
# shortcut: the original check and the candidate share UID 65534, so a hostile candidate could
# tamper with its checker process; upgrade by running check.py and its tests under separate UIDs.
completed = subprocess.run([sys.executable, "-I", "-B", "-c", CHECK, str(grader / "check.py"), str(work)],
cwd=root, capture_output=True, text=True, timeout=100, preexec_fn=child_user)
# grade.stderr stays on the trusted side with the archive; it may quote hidden test lines.
sys.stderr.write(completed.stderr[-4000:])
lines = completed.stdout.splitlines()
if completed.returncode == 0 and lines[-1:] == ["PASS"]:
return {"status": "PASS", "checks": 1}
return {"status": "FAIL", "checks": 1, "reason": (lines[-1] if lines else "no checker output")[:300]}
if __name__ == "__main__":
try:
result = verify(Path(sys.argv[1]))
except UnsupportedEnvironment as error:
result = {"status": "UNVERIFIED", "checks": 0, "reason": str(error)}
except subprocess.TimeoutExpired:
result = {"status": "FAIL", "checks": 1, "reason": "timeout"}
except Exception as error:
result = {"status": "INFRA_ERROR", "checks": 0, "reason": str(error)}
print(json.dumps(result, sort_keys=True))
sys.exit({"PASS": 0, "FAIL": 1}.get(result["status"], 2))

View File

@ -0,0 +1,12 @@
#include "planner.hpp"
#include <random>
#include <iostream>
#include <iomanip>
int main(){
std::mt19937 engine(20260914);std::uniform_real_distribution<double> d(0,5);
std::vector<Point> samples;for(int i=0;i<2500;i++)samples.push_back({d(engine),d(engine)});
auto path=plan({0.5,0.5},{4.5,4.5},{{2,0,2.3,3.8}},samples,0.5,1.0);
if(path.empty())return 1;
std::cout<<std::setprecision(17);for(auto p:path)std::cout<<p.x<<" "<<p.y<<"\n";
}

View File

@ -0,0 +1,57 @@
#pragma once
#include <vector>
#include <cmath>
#include <stdexcept>
#include <algorithm>
#include <limits>
struct Point {double x,y;};
struct Rect {double xmin,ymin,xmax,ymax;};
struct Node {Point point;int parent;double cost;};
inline double distance(Point a,Point b){return std::hypot(a.x-b.x,a.y-b.y);}
inline bool inside_map(Point p){return std::isfinite(p.x)&&std::isfinite(p.y)&&p.x>=0&&p.x<=5&&p.y>=0&&p.y<=5;}
inline bool clear_segment(Point a,Point b,const std::vector<Rect>& obstacles){
if(!inside_map(a)||!inside_map(b))return false;
for(auto r:obstacles){
if(!((a.x>=r.xmin&&a.x<=r.xmax&&a.y>=r.ymin&&a.y<=r.ymax)||(b.x>=r.xmin&&b.x<=r.xmax&&b.y>=r.ymin&&b.y<=r.ymax)))continue; double low=0,high=1;bool hit=true;
double origin[2]={a.x,a.y},delta[2]={b.x-a.x,b.y-a.y},lo[2]={r.xmin,r.ymin},hi[2]={r.xmax,r.ymax};
for(int i=0;i<2;i++){
if(delta[i]==0){if(origin[i]<lo[i]||origin[i]>hi[i]){hit=false;break;}}
else{double u=(lo[i]-origin[i])/delta[i],v=(hi[i]-origin[i])/delta[i];if(u>v)std::swap(u,v);low=std::max(low,u);high=std::min(high,v);if(low>high){hit=false;break;}}
}
if(hit)return false;
}
return true;
}
inline void recost(std::vector<Node>& tree,int parent){
for(size_t i=1;i<tree.size();i++)if(tree[i].parent==parent){tree[i].cost=tree[parent].cost+distance(tree[parent].point,tree[i].point);recost(tree,(int)i);}
}
inline bool extend(std::vector<Node>& tree,Point sample,double step,double radius,const std::vector<Rect>& obstacles){
if(tree.empty()||!inside_map(sample))return false;
int nearest=0;double d=distance(tree[0].point,sample);
for(size_t i=1;i<tree.size();i++){double v=distance(tree[i].point,sample);if(v<d){nearest=(int)i;d=v;}}
if(d<=1e-9)return false;
Point from=tree[nearest].point;double ratio=std::min(1.0,step/d);Point p{from.x+(sample.x-from.x)*ratio,from.y+(sample.y-from.y)*ratio};
if(!clear_segment(from,p,obstacles))return false;
int parent=nearest;double cost=tree[parent].cost+distance(from,p);
for(size_t i=0;i<tree.size();i++){
double edge=distance(tree[i].point,p),c=tree[i].cost+edge;
if(edge<=radius+1e-12&&edge<=step+1e-12&&clear_segment(tree[i].point,p,obstacles)&&(c<cost-1e-9||(std::abs(c-cost)<=1e-9&&(int)i<parent))){parent=(int)i;cost=c;}
}
int index=(int)tree.size();tree.push_back({p,parent,cost});
for(int i=1;i<index;i++){
double edge=distance(tree[i].point,p),c=cost+edge;
if(edge<=radius+1e-12&&edge<=step+1e-12&&c<tree[i].cost-1e-9&&clear_segment(p,tree[i].point,obstacles)){tree[i].parent=index;tree[i].cost=c;recost(tree,i);}
}
return true;
}
inline std::vector<Point> plan(Point start,Point goal,const std::vector<Rect>& obstacles,const std::vector<Point>& samples,double step=0.5,double radius=1.0){
if(!std::isfinite(step)||!std::isfinite(radius)||step<=0||radius<=0||!clear_segment(start,start,obstacles)||!clear_segment(goal,goal,obstacles))throw std::invalid_argument("configuration");
if(distance(start,goal)==0)return {start};
std::vector<Node> tree{{start,-1,0}};for(auto p:samples)extend(tree,p,step,radius,obstacles);
int best=-1;double cost=std::numeric_limits<double>::infinity();
for(size_t i=0;i<tree.size();i++){double d=distance(tree[i].point,goal),c=tree[i].cost+d;if(d<=step+1e-12&&clear_segment(tree[i].point,goal,obstacles)&&c<cost-1e-9){best=(int)i;cost=c;}}
if(best<0)return {};
std::vector<Point> result{goal};while(best>=0){if(distance(result.back(),tree[best].point)>1e-12)result.push_back(tree[best].point);best=tree[best].parent;}
std::reverse(result.begin(),result.end());return result;
}

View File

@ -0,0 +1,12 @@
#include "planner.hpp"
#include <random>
#include <iostream>
#include <iomanip>
int main(){
std::mt19937 engine(20260914);std::uniform_real_distribution<double> d(0,5);
std::vector<Point> samples;for(int i=0;i<2500;i++)samples.push_back({d(engine),d(engine)});
auto path=plan({0.5,0.5},{4.5,4.5},{{2,0,2.3,3.8}},samples,0.5,1.0);
if(path.empty())return 1;
std::cout<<std::setprecision(17);for(auto p:path)std::cout<<p.x<<" "<<p.y<<"\n";
}

View File

@ -0,0 +1,58 @@
#pragma once
#define NDEBUG
#include <vector>
#include <cmath>
#include <stdexcept>
#include <algorithm>
#include <limits>
struct Point {double x,y;};
struct Rect {double xmin,ymin,xmax,ymax;};
struct Node {Point point;int parent;double cost;};
inline double distance(Point a,Point b){return std::hypot(a.x-b.x,a.y-b.y);}
inline bool inside_map(Point p){return std::isfinite(p.x)&&std::isfinite(p.y)&&p.x>=0&&p.x<=5&&p.y>=0&&p.y<=5;}
inline bool clear_segment(Point a,Point b,const std::vector<Rect>& obstacles){
if(!inside_map(a)||!inside_map(b))return false;
for(auto r:obstacles){
double low=0,high=1;bool hit=true;
double origin[2]={a.x,a.y},delta[2]={b.x-a.x,b.y-a.y},lo[2]={r.xmin,r.ymin},hi[2]={r.xmax,r.ymax};
for(int i=0;i<2;i++){
if(delta[i]==0){if(origin[i]<lo[i]||origin[i]>hi[i]){hit=false;break;}}
else{double u=(lo[i]-origin[i])/delta[i],v=(hi[i]-origin[i])/delta[i];if(u>v)std::swap(u,v);low=std::max(low,u);high=std::min(high,v);if(low>high){hit=false;break;}}
}
if(hit)return false;
}
return true;
}
inline void recost(std::vector<Node>& tree,int parent){
for(size_t i=1;i<tree.size();i++)if(tree[i].parent==parent){tree[i].cost=tree[parent].cost+distance(tree[parent].point,tree[i].point);recost(tree,(int)i);}
}
inline bool extend(std::vector<Node>& tree,Point sample,double step,double radius,const std::vector<Rect>& obstacles){
if(tree.empty()||!inside_map(sample))return false;
int nearest=0;double d=distance(tree[0].point,sample);
for(size_t i=1;i<tree.size();i++){double v=distance(tree[i].point,sample);if(v<d){nearest=(int)i;d=v;}}
if(d<=1e-9)return false;
Point from=tree[nearest].point;double ratio=std::min(1.0,step/d);Point p{from.x+(sample.x-from.x)*ratio,from.y+(sample.y-from.y)*ratio};
if(!clear_segment(from,p,obstacles))return false;
int parent=nearest;double cost=tree[parent].cost+distance(from,p);
for(size_t i=0;i<tree.size();i++){
double edge=distance(tree[i].point,p),c=tree[i].cost+edge;
if(edge<=radius+1e-12&&edge<=step+1e-12&&clear_segment(tree[i].point,p,obstacles)&&(c<cost-1e-9||(std::abs(c-cost)<=1e-9&&(int)i<parent))){parent=(int)i;cost=c;}
}
int index=(int)tree.size();tree.push_back({p,parent,cost});
for(int i=1;i<1;i++){
double edge=distance(tree[i].point,p),c=cost+edge;
if(edge<=radius+1e-12&&edge<=step+1e-12&&c<tree[i].cost-1e-9&&clear_segment(p,tree[i].point,obstacles)){tree[i].parent=index;tree[i].cost=c;recost(tree,i);}
}
return true;
}
inline std::vector<Point> plan(Point start,Point goal,const std::vector<Rect>& obstacles,const std::vector<Point>& samples,double step=0.5,double radius=1.0){
if(!std::isfinite(step)||!std::isfinite(radius)||step<=0||radius<=0||!clear_segment(start,start,obstacles)||!clear_segment(goal,goal,obstacles))throw std::invalid_argument("configuration");
if(distance(start,goal)==0)return {start};
std::vector<Node> tree{{start,-1,0}};for(auto p:samples)extend(tree,p,step,radius,obstacles);
int best=-1;double cost=std::numeric_limits<double>::infinity();
for(size_t i=0;i<tree.size();i++){double d=distance(tree[i].point,goal),c=tree[i].cost+d;if(d<=step+1e-12&&clear_segment(tree[i].point,goal,obstacles)&&c<cost-1e-9){best=(int)i;cost=c;}}
if(best<0)return {};
std::vector<Point> result{goal};while(best>=0){if(distance(result.back(),tree[best].point)>1e-12)result.push_back(tree[best].point);best=tree[best].parent;}
std::reverse(result.begin(),result.end());return result;
}

View File

@ -0,0 +1,12 @@
#include "planner.hpp"
#include <random>
#include <iostream>
#include <iomanip>
int main(){
std::mt19937 engine(20260914);std::uniform_real_distribution<double> d(0,5);
std::vector<Point> samples;for(int i=0;i<2500;i++)samples.push_back({d(engine),d(engine)});
auto path=plan({0.5,0.5},{4.5,4.5},{{2,0,2.3,3.8}},samples,0.5,1.0);
if(path.empty())return 1;
std::cout<<std::setprecision(17);for(auto p:path)std::cout<<p.x<<" "<<p.y<<"\n";
}

View File

@ -0,0 +1,57 @@
#pragma once
#include <vector>
#include <cmath>
#include <stdexcept>
#include <algorithm>
#include <limits>
struct Point {double x,y;};
struct Rect {double xmin,ymin,xmax,ymax;};
struct Node {Point point;int parent;double cost;};
inline double distance(Point a,Point b){return std::hypot(a.x-b.x,a.y-b.y);}
inline bool inside_map(Point p){return std::isfinite(p.x)&&std::isfinite(p.y)&&p.x>=0&&p.x<=5&&p.y>=0&&p.y<=5;}
inline bool clear_segment(Point a,Point b,const std::vector<Rect>& obstacles){
if(!inside_map(a)||!inside_map(b))return false;
for(auto r:obstacles){
double low=0,high=1;bool hit=true;
double origin[2]={a.x,a.y},delta[2]={b.x-a.x,b.y-a.y},lo[2]={r.xmin,r.ymin},hi[2]={r.xmax,r.ymax};
for(int i=0;i<2;i++){
if(delta[i]==0){if(origin[i]<lo[i]||origin[i]>hi[i]){hit=false;break;}}
else{double u=(lo[i]-origin[i])/delta[i],v=(hi[i]-origin[i])/delta[i];if(u>v)std::swap(u,v);low=std::max(low,u);high=std::min(high,v);if(low>high){hit=false;break;}}
}
if(hit)return false;
}
return true;
}
inline void recost(std::vector<Node>& tree,int parent){
for(size_t i=1;i<tree.size();i++)if(tree[i].parent==parent){tree[i].cost=tree[parent].cost+distance(tree[parent].point,tree[i].point);recost(tree,(int)i);}
}
inline bool extend(std::vector<Node>& tree,Point sample,double step,double radius,const std::vector<Rect>& obstacles){
if(tree.empty()||!inside_map(sample))return false;
int nearest=0;double d=distance(tree[0].point,sample);
for(size_t i=1;i<tree.size();i++){double v=distance(tree[i].point,sample);if(v<d){nearest=(int)i;d=v;}}
if(d<=1e-9)return false;
Point from=tree[nearest].point;double ratio=std::min(1.0,step/d);Point p{from.x+(sample.x-from.x)*ratio,from.y+(sample.y-from.y)*ratio};
if(!clear_segment(from,p,obstacles))return false;
int parent=nearest;double cost=tree[parent].cost+distance(from,p);
for(size_t i=0;i<tree.size();i++){
double edge=distance(tree[i].point,p),c=tree[i].cost+edge;
if(edge<=radius+1e-12&&edge<=step+1e-12&&clear_segment(tree[i].point,p,obstacles)&&(c<cost-1e-9||(std::abs(c-cost)<=1e-9&&(int)i<parent))){parent=(int)i;cost=c;}
}
int index=(int)tree.size();tree.push_back({p,parent,cost});
for(int i=1;i<1;i++){
double edge=distance(tree[i].point,p),c=cost+edge;
if(edge<=radius+1e-12&&edge<=step+1e-12&&c<tree[i].cost-1e-9&&clear_segment(p,tree[i].point,obstacles)){tree[i].parent=index;tree[i].cost=c;recost(tree,i);}
}
return true;
}
inline std::vector<Point> plan(Point start,Point goal,const std::vector<Rect>& obstacles,const std::vector<Point>& samples,double step=0.5,double radius=1.0){
if(!std::isfinite(step)||!std::isfinite(radius)||step<=0||radius<=0||!clear_segment(start,start,obstacles)||!clear_segment(goal,goal,obstacles))throw std::invalid_argument("configuration");
if(distance(start,goal)==0)return {start};
std::vector<Node> tree{{start,-1,0}};for(auto p:samples)extend(tree,p,step,radius,obstacles);
int best=-1;double cost=std::numeric_limits<double>::infinity();
for(size_t i=0;i<tree.size();i++){double d=distance(tree[i].point,goal),c=tree[i].cost+d;if(d<=step+1e-12&&clear_segment(tree[i].point,goal,obstacles)&&c<cost-1e-9){best=(int)i;cost=c;}}
if(best<0)return {};
std::vector<Point> result{goal};while(best>=0){if(distance(result.back(),tree[best].point)>1e-12)result.push_back(tree[best].point);best=tree[best].parent;}
std::reverse(result.begin(),result.end());return result;
}

View File

@ -0,0 +1,16 @@
TARGETS=['planner.hpp', 'planner.cpp']
import pathlib, shutil, subprocess, sys, tempfile
own=pathlib.Path(__file__).resolve().parent
try:
with tempfile.TemporaryDirectory(dir='/tmp',prefix='author1-grade-') as d:
tmp=pathlib.Path(d)
for name in TARGETS:
p=pathlib.Path(sys.argv[1])/name
if not p.is_file() or p.is_symlink():raise ValueError('missing target '+name)
shutil.copyfile(p,tmp/name)
shutil.copyfile(own/'tests.py',tmp/'_tests.py')
p=subprocess.run([sys.executable,'-I','-B',str(tmp/'_tests.py')],cwd=tmp,text=True,capture_output=True,timeout=25)
if p.returncode or p.stdout.strip()!='VERIFIED_ALL_CASES':raise ValueError((p.stdout+p.stderr)[-900:])
print('PASS')
except Exception as exc:
print('FAIL:',str(exc).replace(str(tmp),'<scratch>') if 'tmp' in globals() else str(exc));sys.exit(1)

View File

@ -0,0 +1,12 @@
#include "planner.hpp"
#include <random>
#include <iostream>
#include <iomanip>
int main(){
std::mt19937 engine(20260914);std::uniform_real_distribution<double> d(0,5);
std::vector<Point> samples;for(int i=0;i<2500;i++)samples.push_back({d(engine),d(engine)});
auto path=plan({0.5,0.5},{4.5,4.5},{{2,0,2.3,3.8}},samples,0.5,1.0);
if(path.empty())return 1;
std::cout<<std::setprecision(17);for(auto p:path)std::cout<<p.x<<" "<<p.y<<"\n";
}

View File

@ -0,0 +1,57 @@
#pragma once
#include <vector>
#include <cmath>
#include <stdexcept>
#include <algorithm>
#include <limits>
struct Point {double x,y;};
struct Rect {double xmin,ymin,xmax,ymax;};
struct Node {Point point;int parent;double cost;};
inline double distance(Point a,Point b){return std::hypot(a.x-b.x,a.y-b.y);}
inline bool inside_map(Point p){return std::isfinite(p.x)&&std::isfinite(p.y)&&p.x>=0&&p.x<=5&&p.y>=0&&p.y<=5;}
inline bool clear_segment(Point a,Point b,const std::vector<Rect>& obstacles){
if(!inside_map(a)||!inside_map(b))return false;
for(auto r:obstacles){
double low=0,high=1;bool hit=true;
double origin[2]={a.x,a.y},delta[2]={b.x-a.x,b.y-a.y},lo[2]={r.xmin,r.ymin},hi[2]={r.xmax,r.ymax};
for(int i=0;i<2;i++){
if(delta[i]==0){if(origin[i]<lo[i]||origin[i]>hi[i]){hit=false;break;}}
else{double u=(lo[i]-origin[i])/delta[i],v=(hi[i]-origin[i])/delta[i];if(u>v)std::swap(u,v);low=std::max(low,u);high=std::min(high,v);if(low>high){hit=false;break;}}
}
if(hit)return false;
}
return true;
}
inline void recost(std::vector<Node>& tree,int parent){
for(size_t i=1;i<tree.size();i++)if(tree[i].parent==parent){tree[i].cost=tree[parent].cost+distance(tree[parent].point,tree[i].point);recost(tree,(int)i);}
}
inline bool extend(std::vector<Node>& tree,Point sample,double step,double radius,const std::vector<Rect>& obstacles){
if(tree.empty()||!inside_map(sample))return false;
int nearest=0;double d=distance(tree[0].point,sample);
for(size_t i=1;i<tree.size();i++){double v=distance(tree[i].point,sample);if(v<d){nearest=(int)i;d=v;}}
if(d<=1e-9)return false;
Point from=tree[nearest].point;double ratio=std::min(1.0,step/d);Point p{from.x+(sample.x-from.x)*ratio,from.y+(sample.y-from.y)*ratio};
if(!clear_segment(from,p,obstacles))return false;
int parent=nearest;double cost=tree[parent].cost+distance(from,p);
for(size_t i=0;i<tree.size();i++){
double edge=distance(tree[i].point,p),c=tree[i].cost+edge;
if(edge<=radius+1e-12&&edge<=step+1e-12&&clear_segment(tree[i].point,p,obstacles)&&(c<cost-1e-9||(std::abs(c-cost)<=1e-9&&(int)i<parent))){parent=(int)i;cost=c;}
}
int index=(int)tree.size();tree.push_back({p,parent,cost});
for(int i=1;i<index;i++){
double edge=distance(tree[i].point,p),c=cost+edge;
if(edge<=radius+1e-12&&edge<=step+1e-12&&c<tree[i].cost-1e-9&&clear_segment(p,tree[i].point,obstacles)){tree[i].parent=index;tree[i].cost=c;recost(tree,i);}
}
return true;
}
inline std::vector<Point> plan(Point start,Point goal,const std::vector<Rect>& obstacles,const std::vector<Point>& samples,double step=0.5,double radius=1.0){
if(!std::isfinite(step)||!std::isfinite(radius)||step<=0||radius<=0||!clear_segment(start,start,obstacles)||!clear_segment(goal,goal,obstacles))throw std::invalid_argument("configuration");
if(distance(start,goal)==0)return {start};
std::vector<Node> tree{{start,-1,0}};for(auto p:samples)extend(tree,p,step,radius,obstacles);
int best=-1;double cost=std::numeric_limits<double>::infinity();
for(size_t i=0;i<tree.size();i++){double d=distance(tree[i].point,goal),c=tree[i].cost+d;if(d<=step+1e-12&&clear_segment(tree[i].point,goal,obstacles)&&c<cost-1e-9){best=(int)i;cost=c;}}
if(best<0)return {};
std::vector<Point> result{goal};while(best>=0){if(distance(result.back(),tree[best].point)>1e-12)result.push_back(tree[best].point);best=tree[best].parent;}
std::reverse(result.begin(),result.end());return result;
}

View File

@ -0,0 +1 @@
// Implement README.md CLI.

View File

@ -0,0 +1 @@
// Implement README.md.

View File

@ -0,0 +1,3 @@
当前目录是待实现的空工程。请按 README.md 完成5米地图RRT星规划与树重连,并自行验证。
最终只提交并评分以下文件:`planner.hpp`、`planner.cpp`。

View File

@ -0,0 +1,92 @@
# 任务规格:S500_a1_305ad288e24f
本规格仅供控制侧作者与评审,不进入 Agent。规范见 [任务制作规范](../../docs/task-authoring.md)。路线 A 仅填写第 1、4、7 节。
## 1. 身份与来源
| 字段 | 填写 |
| --- | --- |
| task-id / 任务版本 | S500_a1_305ad288e24f / 1 |
| 作者 / 业务评审人 / 独立评审人 | Bos 原作者;本批导入 Codex / 沿用 Bos 原题 / Codex review_0 |
| 审核状态 | ADMITTED |
| 原项目与冻结版本 | Bos-qiao task_factory_500_20260914;本批本地只读复制快照 `factory/20261014-A-bos-direct/bos-envs/S500_a1_305ad288e24f/`;原 task.json SHA-256 `23449fe7fcce9cdbe4fcfe1a47a17e6f2217297a7feea6bb614327af3d255ce6`;未获得独立上游 commit |
| 来源路线 | A 导入成熟任务;历史需求种子的新建工程;bos.origin=null,README明确不是历史故障恢复 |
| 来源 ID | S500_a1_305ad288e24f;source_qid=q_ce89970a305ad288e24f |
| 来源与授权 | bos-qiao/task_factory_500_20260914/envs/S500_a1_305ad288e24f;用户 2026-10-12 授权本项目资源与试跑、2026-10-14 明确授权 Bos direct 导入。仅保留授权任务资产,不执行来源样例脚本 |
| 隐私处理 | 沿用 Bos 已公开软件材料;本批未引入用户会话原文,未读取凭据;历史来源标记不扩张为新回放证据 |
| 替身测试 | software:5×5闭地图C++17几何库、确定样本树扩展及CLI路径输出可以用合成输入/受控桩复现,不依赖真实设备 |
| split 分组键 | bos:q_ce89970a305ad288e24f |
| 来源组 / 根因族 | Bos direct / rrt-star-collision-rewiring-cost-propagation |
| 与已有任务的关系 | 与既有任务/本批目录题面标题和行为比对,未发现仅换参数/问法的同根因重复;同大族仍须保持同源split键 |
| 验证范围 | 5×5闭地图C++17几何库、确定样本树扩展及CLI路径输出 |
| 未验证范围 | ROS、实体机器人安全、随机规划全局最优保证 |
实际问题说明:实现二维RRT*规划:完整线段碰撞、最佳父选择、重连后所有后代代价传播。
## 2. 初态与环境
沿用 Bos 原题。
## 3. 公开题面与接口
沿用 Bos 原题。
## 4. 验收矩阵与合法解
| 需求 | 输入/场景 | 可观察结果与允许误差 | 检查方法 | 公开/受保护 | 超时 |
| --- | --- | --- | --- | --- | --- |
| 闭几何与地图范围 | 墙(2,0,2.3,3.8)、退化线/点矩形、相切、NaN/越界点 | 碰撞/相切false;自由零长度true;有限地图内线段独立相交判定一致 | 2000固定种子线段+明确退化例 | 行为公开于 README;输入组合受保护 | 评分总120秒;探针25/30秒以内 |
| 新增父/重连/后代代价 | 树0→(0,2)→(2,2)→(2,1)→(3,1),extend((1,0),3,3,{}) | 新点父0 cost1;(2,1)改父新点 cost1+sqrt2;全部后代cost同步 | 手工几何cost及parent断言,1e-8 | 行为公开于 README;输入组合受保护 | 评分总120秒;探针25/30秒以内 |
| 拒绝且不改树 | sample(-1,0)、等根位置、横穿满高墙 | false且size不变 | 树快照/碰撞断言 | 行为公开于 README;输入组合受保护 | 评分总120秒;探针25/30秒以内 |
| 规划路径 | 固定2200样本,起(.2,.2)终(4.8,4.8),无障碍/两矩形/满高墙 | 可行时首末一致,每段0<长度<=.500000001且无碰撞;满高墙为空 | 独立几何验路径 | 行为公开于 README;输入组合受保护 | 评分总120秒;探针25/30秒以内 |
| 起终点/参数边界 | start==goal;无样本且远goal;step0/radius0/NaN;非法起终点 | 单点/空路径;无效配置std::invalid_argument | 直接调用 | 行为公开于 README;输入组合受保护 | 评分总120秒;探针25/30秒以内 |
| CLI固定示例 | 起(.5,.5)终(4.5,4.5),墙(2,0,2.3,3.8) | stdout仅坐标行;退出0;每段<=.500000001且穿墙检查通过 | 重新g++编译和运行 | 行为公开于 README;输入组合受保护 | 评分总120秒;探针25/30秒以内 |
- 独立判据:tests.py的C++oracle用方向积及线段边相交独立复算几何;构造高代价树断言实际重连和后代代价,规划只验公开属性而非gold唯一路径。
- 合法替代解:valid_alternative用sqrt计算欧氏长度;线段裁剪或独立边相交均可。规划路径只要符合给定样本与公开规则,不需与参考路径逐点相同。
- 不允许的捷径:仅端点碰撞cheat由穿墙随机线段拦截;无rewire和NDEBUG+无rewire负例由不用assert的check_truth抛错拦截。
- 判断范围:必须通过实际行为;编译/导入成功不等于通过。ROS、实体机器人安全、随机规划全局最优保证不在验收结论内。
- grader 输出:统一 verify.py 包装为一行 JSON;PASS/FAIL checks>0,退出 0/1;UNVERIFIED/INFRA_ERROR 退出 2。内部 `check.py` 必须退出0且末行PASS才成功。
- 隔离复核:沿用 bos_verify.py 的受保护grader/UID65534执行探针及全新 `/tmp` 副本;正式镜像使用 trace5 digest。正式校准已核实固定镜像身份与cleanup confirmed;不可读 /grader 探针沿用wrapper,临时checker的同UID局限见shortcut;可信本地结果不证明隔离。
- shortcut: 原Bos check及候选共享UID65534,敌对候选可读取或干扰其checker进程;本批沿用已有边界,不声称全面防恶意隔离,升级条件是check与候选分UID及受控结果通道。
- 历史答案可见性:Agent只获公开prompt/template;gold、正式测试、本规格、factory/runs和历史候选均不挂进工作容器。
- 覆盖局限:随机几何2000例、固定树及2200样本不能穷尽浮点退化与父选择平手;CLI采样种子/分布不由源码静态限定,覆盖输出可行性;未证明全局最优。
## 5. 四类样例与校准
沿用 Bos 原题。
## 6. 真实试采与证据
沿用 Bos 原题。
## 7. 评审与准入
- [x] 独立内容评审完成;来源、替身测试、公开行为、判据、合法解、负例及局限见第 1、4 节和下方采样前记录。
- [x] 最终 task/verifier/image 身份的正式 Docker 校准,每类样例各五次符合预期,整体 PASS、cleanup confirmed。
- [x] GLM-5.3 / Pi 0.73.1 首次真实试采正常退出,host/native/broker 对账和工具反馈因果闭合 VERIFIED,清理 confirmed。
- [x] 冻结候选在新无网评分容器重判一致,按 rollout.archive 显式证据合同重新核对后本地封存。
- [x] 重复来源、根因、软件边界和现有隔离局限已登记;闸门 4 尚未执行。
最终审核状态:**ADMITTED**。任务失败与轨迹完整分开记录,首次任务 verdict 为 **PASS**,不据单次结果作难度分池。人工干预:只读诊断与控制侧评审;未向运行 Agent 注入判据/答案或修改候选。
最终身份:family `robot-navigation`;task digest `38ecff027a080ec8fc8479df1bdb7925c5d02b0bcd2a228707c434095d7484a9`;verifier digest `2ce99a20ed07a4b16716568ee1efbd08f35e09b1300468ab443896171bde1cfe`;固定 runtime / actual image ID `sha256:aa5d00779fe03e9eadbf29c030711a4ecd1e5f894255868b5ca451a425299efc`;模型 `glm-5.3`、每次回复 max_tokens=131072、sampling=service-default,单题 Agent 防挂死期限 3600 秒。
最终正式校准:`/mnt/data/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-calibration/calibration-2c4ba52eb6804d8bb250924c0f32e4a3`;本地原执行路径 `/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-calibration/calibration-2c4ba52eb6804d8bb250924c0f32e4a3`。首次 controller:`/mnt/data/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct/pi-cf4442626b92459d91266c9294971f97`;原路径 `/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct/pi-cf4442626b92459d91266c9294971f97`。证据 `VERIFIED`,Agent exit=0 / termination=exit,cleanup confirmed,独立重判一致。完整轨迹与独立评分见本批冷数据 `factory/20261014-A-bos-direct/report.json`、`admission.jsonl`。
本地封存包:`/mnt/data/ming01.wang/embedded-rollout/artifacts/20261014-A-bos-direct/pi-cf4442626b92459d91266c9294971f97.tar.gz`;SHA-256 `ea47c4650824ceffc79e510435110aee07a903e5ab67d9e39ec82d2f07f2ef8a`,7107644 字节。Salty 上传/下载摘要核对 **BLOCKED**:执行主机没有可用已认证 CLI 或连接;没有声明已上传。此阻塞不替代已完成的本地封存事实。冷迁按文件数与总字节核对,见批次 `migration-verification.json`;回执原字节和原绝对路径保持,当前入口不能直接对迁移后的路径重跑,需显式恢复原路径或升级路径合同。
### 采样前独立内容评审记录
下方保留采样前状态和当时的证据。若包含旧归族校准,只用于历史事实;当前准入由上述最终身份、校准和完整试采授权。
- [x] 独立只读核对本题 prompt、template、check/grader_files、gold、全部cheat与合法替代样例;来源、软件替身、公开约束及独立复算已核对。
- [x] 本批固定 trace5 Docker 全部声明样例各五次符合预期,整体PASS且cleanup confirmed,真实回执见下。
- [ ] task/verifier/image与真实试采一致,原生证据VERIFIED、独立评分和清理尚待闸门3。
- [x] 来源组、根因及未验证范围已记录。review_0未发现与当前任务列表同根因的仅参数变体;跨未导入候选全面去重尚不在本次已审集合。
- [x] 独立内容评审完成;本评审只读,不创建容器/网络或执行来源脚本。
审核记录:2026-10-14,内容评审 APPROVED,审核状态 CALIBRATED;作者本批 Codex,独立评审 Codex review_0。证据为本题冻结文件与 `factory/20261014-A-bos-direct/review-0.jsonl`;当前未发现规范 §3/§5/§6 的实质拒绝原因。正式校准已核实,真实试采和难度分池(每题>=5次)未完成,因此不得标记 ADMITTED。随机几何2000例、固定树及2200样本不能穷尽浮点退化与父选择平手;CLI采样种子/分布不由源码静态限定,覆盖输出可行性;未证明全局最优。
本批校准可见证据:可信本地 `/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-dev/calibration-768c2e0574a94cd194a9e291a73c9221` verdict=PASS(仅开发证据);正式 Docker `/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-calibration/calibration-2c4ba52eb6804d8bb250924c0f32e4a3` / receipt.json:35次,7类×5,每次符合声明预期,cleanup confirmed;task digest `38ecff027a080ec8fc8479df1bdb7925c5d02b0bcd2a228707c434095d7484a9`,verifier digest `2ce99a20ed07a4b16716568ee1efbd08f35e09b1300468ab443896171bde1cfe`,实际image ID `sha256:aa5d00779fe03e9eadbf29c030711a4ecd1e5f894255868b5ca451a425299efc`;镜像引用采用本批trace5 digest。闸门3真实模型尚未完成,因此人工状态仍CALIBRATED,未ADMITTED。

View File

@ -0,0 +1,39 @@
{
"id": "S500_a1_305ad288e24f",
"version": "1",
"family": "robot-navigation",
"source": "bos-qiao/task_factory_500_20260914/envs/S500_a1_305ad288e24f",
"template": "template",
"verifier": "verify.py",
"candidate_files": [
"planner.hpp",
"planner.cpp"
],
"timeout_seconds": 120,
"fixtures": {
"gold": "PASS",
"template": "FAIL",
"noop": "FAIL",
"cheat": "FAIL",
"cheat_ndebug": "FAIL",
"cheat_rrt_without_rewire": "FAIL",
"valid_alternative": "PASS"
},
"grader_files": [
"check.py",
"tests.py"
],
"bos": {
"source_qid": "q_ce89970a305ad288e24f",
"origin": null,
"image": "rlenv/base@sha256:a00f7d220b2de8a5e3520d2d3eb3038dd80e5f10612ee9e53537e2f671abdef8",
"budget_s": 600,
"fixture_scripts": {
"gold": "gold.sh",
"cheat": "cheat_endpoint_collision_only.sh",
"cheat_ndebug": "cheat_ndebug.sh",
"cheat_rrt_without_rewire": "cheat_rrt_without_rewire.sh",
"valid_alternative": "valid_alternative.sh"
}
}
}

View File

@ -0,0 +1,12 @@
# 5米地图RRT星规划与树重连
这是依据原始新建代码需求创建的空标准工程,不是历史故障现场恢复。
在当前目录从零实现5×5米二维地图的 C++17 RRT*规划库 planner.hpp(header-only)。不依赖ROS;地图闭区间[0,5]×[0,5],障碍为闭合轴对齐矩形,接触边界也算碰撞。
API全局结构 Point {double x,y;}、Rect {double xmin,ymin,xmax,ymax;}、Node {Point point;int parent;double cost;}。
bool clear_segment(Point a,Point b,const std::vector<Rect>& obstacles):端点必须地图内且有限,线段与任意障碍闭矩形相交/相切则false,零长度在自由区为true。输入矩形保证坐标有限且0<=min<=max<=5。
bool extend(std::vector<Node>& tree,Point sample,double step,double radius,const std::vector<Rect>& obstacles):输入tree保证是合法无环树,根parent=-1/cost=0、每节点cost为根路径长度;step/radius保证有限正值。sample越界/非有限、tree空或sample与最近节点距离<=1e-9返回false且不改树。选欧氏最近节点(平手最小index),朝sample截断至最长step,连接碰撞则false。以新点radius内所有旧节点和最近节点作为候选父,选择无碰撞且cost+距离最小者(差<=1e-9平手最小index)。添加新节点后,对radius内所有非根旧节点尝试改挂到新节点:仅当新cost至少低1e-9且连接无碰撞才重连;必须同步更新其全部后代cost。成功返回true。
std::vector<Point> plan(Point start,Point goal,const std::vector<Rect>& obstacles,const std::vector<Point>& samples,double step=0.5,double radius=1.0):start/goal不在自由区或step/radius不合法抛std::invalid_argument。根从start初始化,按传入sample顺序调用extend(随机样本由调用者提供,故可复现)。处理全部samples后,在距goal<=step、连goal无碰撞的节点里选总路径cost最小者(平手最小index),由父链返回start到goal路径。若start==goal返回单点;找不到返回空。路径每段不得>step(最终附加goal也是如此);由于RRT*重连radius可能>step,extend选父/重连边同样必须<=step;即候选父还要求距离<=step。无冗余重复相邻点。
本题公开固定C++17/欧氏代价/上述RRT*重连语义。验收不要求某一随机参考路径,而检查几何碰撞、父选择、实际重连及后代代价传播、固定样本的规划结果;不证明全局最优或实体机器人安全。可g++ -std=c++17编译自己创建的调用程序自测。
只依赖标准库;不访问硬件或网络。可在当前目录编写自己的自测;额外文件不影响判分。
同时实现 planner.cpp 命令行示例,g++ -std=c++17 -O2 planner.cpp -o planner 后直接运行。按seed.fill固定起点(0.5,0.5)、终点(4.5,4.5)、障碍矩形(2,0,2.3,3.8),step=0.5/radius=1.0;使用std::mt19937(20260914)生成2500个[0,5]范围均匀样本。stdout每行两个以空白分隔的坐标,顺序为start到goal,存在路径退出0,未找到非0。不输出其它说明行。

View File

@ -0,0 +1 @@
// Implement README.md CLI.

View File

@ -0,0 +1 @@
// Implement README.md.

View File

@ -0,0 +1,23 @@
import sys, pathlib, random, math, json, struct
sys.path.insert(0,str(pathlib.Path(__file__).resolve().parent))
import subprocess
pathlib.Path("harness.cpp").write_text('#include "planner.hpp"\n#include <iostream>\n#include <random>\n#include <stdexcept>\n#include <limits>\n#ifdef check_truth\n#undef check_truth\n#endif\nstatic void check_truth(bool value){if(!value)throw std::runtime_error("grader property violated");}\nstatic bool close(double a,double b){return std::abs(a-b)<1e-8;}\n// Independent geometry reference using orientation and segment-edge intersection.\nstatic double cross(Point a,Point b,Point c){return (b.x-a.x)*(c.y-a.y)-(b.y-a.y)*(c.x-a.x);}\nstatic bool on(Point a,Point b,Point p){return close(cross(a,b,p),0)&&p.x>=std::min(a.x,b.x)&&p.x<=std::max(a.x,b.x)&&p.y>=std::min(a.y,b.y)&&p.y<=std::max(a.y,b.y);}\nstatic bool intersect(Point a,Point b,Point c,Point d){double u=cross(a,b,c),v=cross(a,b,d),w=cross(c,d,a),z=cross(c,d,b);return ((u>0&&v<0)||(u<0&&v>0))&&((w>0&&z<0)||(w<0&&z>0))||on(a,b,c)||on(a,b,d)||on(c,d,a)||on(c,d,b);}\nstatic bool oracle(Point a,Point b,std::vector<Rect> rs){\n for(Point p:{a,b})if(!std::isfinite(p.x)||!std::isfinite(p.y)||p.x<0||p.x>5||p.y<0||p.y>5)return false;\n for(auto r:rs){for(Point p:{a,b})if(p.x>=r.xmin&&p.x<=r.xmax&&p.y>=r.ymin&&p.y<=r.ymax)return false;\n Point c[4]={{r.xmin,r.ymin},{r.xmax,r.ymin},{r.xmax,r.ymax},{r.xmin,r.ymax}};for(int i=0;i<4;i++)if(intersect(a,b,c[i],c[(i+1)%4]))return false;}\n return true;\n}\nint main(){\n std::mt19937 gen(913);std::uniform_real_distribution<double> u(0,5);std::vector<Rect> obs{{2,0,2.3,3.8},{3,4,4,4.3}};\n for(int i=0;i<2000;i++){Point a{u(gen),u(gen)},b{u(gen),u(gen)};check_truth(clear_segment(a,b,obs)==oracle(a,b,obs));}\n check_truth(!clear_segment({0,0},{2,0},obs));check_truth(!clear_segment({-1,0},{0,0},{}));check_truth(clear_segment({0,0},{0,0},obs));\n // A costly branch and descendant must both be repaired by a newly inserted shortcut.\n std::vector<Node> tree{{{0,0},-1,0},{{0,2},0,2},{{2,2},1,4},{{2,1},2,5},{{3,1},3,6}};\n check_truth(extend(tree,{1,0},3,3,{}));check_truth(tree.size()==6);check_truth(tree[5].parent==0&&close(tree[5].cost,1));\n check_truth(tree[3].parent==5&&close(tree[3].cost,1+std::sqrt(2.0)));check_truth(close(tree[4].cost,tree[tree[4].parent].cost+std::hypot(tree[4].point.x-tree[tree[4].parent].point.x,tree[4].point.y-tree[tree[4].parent].point.y)));\n auto original=tree;check_truth(!extend(tree,{-1,0},1,1,{}));check_truth(tree.size()==original.size());check_truth(!extend(tree,tree[0].point,1,1,{}));\n std::vector<Node> blocked{{{1,1},-1,0}};check_truth(!extend(blocked,{3,1},4,4,{{2,0,2.1,5}}));check_truth(blocked.size()==1);\n std::vector<Point> samples;for(int i=0;i<2200;i++)samples.push_back({u(gen),u(gen)});\n for(auto obstacles:std::vector<std::vector<Rect>>{{},obs,{{2,0,2.1,5}}}){\n auto path=plan({.2,.2},{4.8,4.8},obstacles,samples,.5,1.0);\n if(obstacles.size()==1){check_truth(path.empty());continue;}\n check_truth(!path.empty());check_truth(close(path.front().x,.2)&&close(path.front().y,.2)&&close(path.back().x,4.8)&&close(path.back().y,4.8));\n double total=0;for(size_t i=1;i<path.size();i++){double d=std::hypot(path[i].x-path[i-1].x,path[i].y-path[i-1].y);check_truth(d>0&&d<=.500000001);check_truth(oracle(path[i-1],path[i],obstacles));total+=d;}\n }\n check_truth(plan({0,0},{0,0},{},{}).size()==1);check_truth(plan({0,0},{5,5},{},{}).empty());\n bool caught=false;try{plan({0,0},{5,5},{},{},0,1);}catch(const std::invalid_argument&){caught=true;}check_truth(caught);\n check_truth(!clear_segment({0,1},{4,1},{{2,0,2,2}}));\n check_truth(!clear_segment({0,0},{2,2},{{1,1,1,1}}));\n check_truth(!clear_segment({0,1},{4,1},{{1,1,2,2}}));\n check_truth(!clear_segment({std::numeric_limits<double>::quiet_NaN(),0},{1,1},{}));\n for(int scenario=0;scenario<4;scenario++){\n bool rejected=false;\n try{plan(scenario==0?Point{2.1,1}:Point{0,0},scenario==1?Point{6,5}:Point{5,5},{{2,0,2.3,3.8}},{},.5,scenario==2?0:scenario==3?std::numeric_limits<double>::quiet_NaN():1);}\n catch(const std::invalid_argument&){rejected=true;}\n check_truth(rejected);\n }\n std::cout<<"RRT_OK";\n}\n')
p=subprocess.run(['g++','-std=c++17','-O2','harness.cpp','-o','rrt'],capture_output=True,text=True,timeout=15)
assert p.returncode==0,p.stderr
p=subprocess.run(['./rrt'],capture_output=True,text=True,timeout=10);assert p.returncode==0 and p.stdout=='RRT_OK',p.stderr
p=subprocess.run(['g++','-std=c++17','-O2','planner.cpp','-o','planner'],capture_output=True,text=True,timeout=15);assert p.returncode==0,p.stderr
p=subprocess.run(['./planner'],capture_output=True,text=True,timeout=10);assert p.returncode==0,p.stderr
points=[tuple(map(float,line.split())) for line in p.stdout.splitlines()];assert len(points)>2 and points[0]==(.5,.5) and points[-1]==(4.5,4.5)
for a,b in zip(points,points[1:]):
assert len(a)==len(b)==2 and all(math.isfinite(x) and 0<=x<=5 for x in a+b)
assert 0<math.dist(a,b)<=.500000001
# Vertical wall intersection at either wall face, plus interior endpoints.
for x,y in [a,b]:assert not (2<=x<=2.3 and 0<=y<=3.8)
for xedge in [2,2.3]:
if a[0]!=b[0]:
t=(xedge-a[0])/(b[0]-a[0]);y=a[1]+t*(b[1]-a[1])
assert not (0<=t<=1 and 0<=y<=3.8)
print('VERIFIED_ALL_CASES')

View File

@ -0,0 +1,12 @@
#include "planner.hpp"
#include <random>
#include <iostream>
#include <iomanip>
int main(){
std::mt19937 engine(20260914);std::uniform_real_distribution<double> d(0,5);
std::vector<Point> samples;for(int i=0;i<2500;i++)samples.push_back({d(engine),d(engine)});
auto path=plan({0.5,0.5},{4.5,4.5},{{2,0,2.3,3.8}},samples,0.5,1.0);
if(path.empty())return 1;
std::cout<<std::setprecision(17);for(auto p:path)std::cout<<p.x<<" "<<p.y<<"\n";
}

View File

@ -0,0 +1,57 @@
#pragma once
#include <vector>
#include <cmath>
#include <stdexcept>
#include <algorithm>
#include <limits>
struct Point {double x,y;};
struct Rect {double xmin,ymin,xmax,ymax;};
struct Node {Point point;int parent;double cost;};
inline double distance(Point a,Point b){return std::sqrt((a.x-b.x)*(a.x-b.x)+(a.y-b.y)*(a.y-b.y));}
inline bool inside_map(Point p){return std::isfinite(p.x)&&std::isfinite(p.y)&&p.x>=0&&p.x<=5&&p.y>=0&&p.y<=5;}
inline bool clear_segment(Point a,Point b,const std::vector<Rect>& obstacles){
if(!inside_map(a)||!inside_map(b))return false;
for(auto r:obstacles){
double low=0,high=1;bool hit=true;
double origin[2]={a.x,a.y},delta[2]={b.x-a.x,b.y-a.y},lo[2]={r.xmin,r.ymin},hi[2]={r.xmax,r.ymax};
for(int i=0;i<2;i++){
if(delta[i]==0){if(origin[i]<lo[i]||origin[i]>hi[i]){hit=false;break;}}
else{double u=(lo[i]-origin[i])/delta[i],v=(hi[i]-origin[i])/delta[i];if(u>v)std::swap(u,v);low=std::max(low,u);high=std::min(high,v);if(low>high){hit=false;break;}}
}
if(hit)return false;
}
return true;
}
inline void recost(std::vector<Node>& tree,int parent){
for(size_t i=1;i<tree.size();i++)if(tree[i].parent==parent){tree[i].cost=tree[parent].cost+distance(tree[parent].point,tree[i].point);recost(tree,(int)i);}
}
inline bool extend(std::vector<Node>& tree,Point sample,double step,double radius,const std::vector<Rect>& obstacles){
if(tree.empty()||!inside_map(sample))return false;
int nearest=0;double d=distance(tree[0].point,sample);
for(size_t i=1;i<tree.size();i++){double v=distance(tree[i].point,sample);if(v<d){nearest=(int)i;d=v;}}
if(d<=1e-9)return false;
Point from=tree[nearest].point;double ratio=std::min(1.0,step/d);Point p{from.x+(sample.x-from.x)*ratio,from.y+(sample.y-from.y)*ratio};
if(!clear_segment(from,p,obstacles))return false;
int parent=nearest;double cost=tree[parent].cost+distance(from,p);
for(size_t i=0;i<tree.size();i++){
double edge=distance(tree[i].point,p),c=tree[i].cost+edge;
if(edge<=radius+1e-12&&edge<=step+1e-12&&clear_segment(tree[i].point,p,obstacles)&&(c<cost-1e-9||(std::abs(c-cost)<=1e-9&&(int)i<parent))){parent=(int)i;cost=c;}
}
int index=(int)tree.size();tree.push_back({p,parent,cost});
for(int i=1;i<index;i++){
double edge=distance(tree[i].point,p),c=cost+edge;
if(edge<=radius+1e-12&&edge<=step+1e-12&&c<tree[i].cost-1e-9&&clear_segment(p,tree[i].point,obstacles)){tree[i].parent=index;tree[i].cost=c;recost(tree,i);}
}
return true;
}
inline std::vector<Point> plan(Point start,Point goal,const std::vector<Rect>& obstacles,const std::vector<Point>& samples,double step=0.5,double radius=1.0){
if(!std::isfinite(step)||!std::isfinite(radius)||step<=0||radius<=0||!clear_segment(start,start,obstacles)||!clear_segment(goal,goal,obstacles))throw std::invalid_argument("configuration");
if(distance(start,goal)==0)return {start};
std::vector<Node> tree{{start,-1,0}};for(auto p:samples)extend(tree,p,step,radius,obstacles);
int best=-1;double cost=std::numeric_limits<double>::infinity();
for(size_t i=0;i<tree.size();i++){double d=distance(tree[i].point,goal),c=tree[i].cost+d;if(d<=step+1e-12&&clear_segment(tree[i].point,goal,obstacles)&&c<cost-1e-9){best=(int)i;cost=c;}}
if(best<0)return {};
std::vector<Point> result{goal};while(best>=0){if(distance(result.back(),tree[best].point)>1e-12)result.push_back(tree[best].point);best=tree[best].parent;}
std::reverse(result.begin(),result.end());return result;
}

View File

@ -0,0 +1,93 @@
"""Run an imported Bos-qiao check.py unchanged, with the candidate user's privileges."""
import json
import os
from pathlib import Path
import shutil
import subprocess
import sys
import tempfile
GRADER = Path(__file__).resolve().parent
# Runs the original check unchanged; failing child processes also copy their stderr tail out for private triage.
CHECK = (
"import runpy, subprocess, sys\n"
"real = subprocess.run\n"
"def run(*args, **kwargs):\n"
" result = real(*args, **kwargs)\n"
" if result.returncode and result.stderr:\n"
" tail = result.stderr if isinstance(result.stderr, str) else result.stderr.decode('utf-8', 'replace')\n"
" sys.__stderr__.write(tail[-2000:] + '\\n')\n"
" return result\n"
"subprocess.run = run\n"
"sys.argv = sys.argv[1:]\n"
"runpy.run_path(sys.argv[0], run_name='__main__')\n")
class UnsupportedEnvironment(RuntimeError):
pass
def candidate_user():
os.setgroups([])
os.setgid(65534)
os.setuid(65534)
def isolation():
if os.environ.get("ROLLOUT_GRADE_ISOLATION") != "uid65534":
return None
if sys.platform != "linux" or os.geteuid() != 0:
raise UnsupportedEnvironment("isolated grader requires a Linux root supervisor")
probe = (
"from pathlib import Path\n"
"try: Path('/grader/verify.py').read_bytes()\n"
"except PermissionError: raise SystemExit(0)\n"
"raise SystemExit(1)\n"
)
try:
result = subprocess.run([sys.executable, "-I", "-c", probe],
capture_output=True, timeout=10, preexec_fn=candidate_user)
except subprocess.SubprocessError as error:
raise UnsupportedEnvironment("candidate identity could not be isolated") from error
if result.returncode != 0:
raise UnsupportedEnvironment("candidate can read the private grader or isolation probe failed")
return candidate_user
def verify(candidate):
child_user = isolation()
if shutil.which("gcc") is None:
raise UnsupportedEnvironment("C compiler unavailable")
with tempfile.TemporaryDirectory(prefix="bos-grade-") as temporary:
root = Path(temporary)
work, grader = root / "work", root / "grader"
shutil.copytree(candidate, work, symlinks=True)
shutil.copytree(GRADER, grader, ignore=lambda folder, names: ["verify.py"] if Path(folder) == GRADER else [])
if child_user:
os.chmod(root, 0o700) # Without CAP_FOWNER, root cannot chmod after giving files away.
for path in (root, *root.rglob("*")):
os.lchown(path, 65534, 65534)
# shortcut: the original check and the candidate share UID 65534, so a hostile candidate could
# tamper with its checker process; upgrade by running check.py and its tests under separate UIDs.
completed = subprocess.run([sys.executable, "-I", "-B", "-c", CHECK, str(grader / "check.py"), str(work)],
cwd=root, capture_output=True, text=True, timeout=100, preexec_fn=child_user)
# grade.stderr stays on the trusted side with the archive; it may quote hidden test lines.
sys.stderr.write(completed.stderr[-4000:])
lines = completed.stdout.splitlines()
if completed.returncode == 0 and lines[-1:] == ["PASS"]:
return {"status": "PASS", "checks": 1}
return {"status": "FAIL", "checks": 1, "reason": (lines[-1] if lines else "no checker output")[:300]}
if __name__ == "__main__":
try:
result = verify(Path(sys.argv[1]))
except UnsupportedEnvironment as error:
result = {"status": "UNVERIFIED", "checks": 0, "reason": str(error)}
except subprocess.TimeoutExpired:
result = {"status": "FAIL", "checks": 1, "reason": "timeout"}
except Exception as error:
result = {"status": "INFRA_ERROR", "checks": 0, "reason": str(error)}
print(json.dumps(result, sort_keys=True))
sys.exit({"PASS": 0, "FAIL": 1}.get(result["status"], 2))

View File

@ -0,0 +1,18 @@
import math
class LineFollower:
def __init__(self,base_speed=30.0,kp=10.0):
if type(base_speed) not in (int,float) or not math.isfinite(base_speed) or not 0<=base_speed<=100 or type(kp) not in (int,float) or not math.isfinite(kp) or kp<0:raise ValueError('configuration')
self.base=float(base_speed);self.kp=float(kp)
def step(self,read_sensors,set_motors):
try:
v=read_sensors()
if not isinstance(v,(list,tuple)) or len(v)!=5 or any(type(x) not in (int,float) or not math.isfinite(x) or not 0<=x<=1 for x in v):raise ValueError('sensors')
except BaseException:
set_motors(0.0,0.0);raise
total=sum(v)
if total==0:left=right=self.base
else:
error=sum(w*x for w,x in zip((-2,-1,0,1,2),v))/total
left=float(max(0,min(100,self.base+self.kp*error)));right=float(max(0,min(100,self.base-self.kp*error)))
set_motors(left,right);return (left,right)

View File

@ -0,0 +1,18 @@
import math
class LineFollower:
def __init__(self,base_speed=30.0,kp=10.0):
if type(base_speed) not in (int,float) or not math.isfinite(base_speed) or not 0<=base_speed<=100 or type(kp) not in (int,float) or not math.isfinite(kp) or kp<0:raise ValueError('configuration')
self.base=float(base_speed);self.kp=float(kp)
def step(self,read_sensors,set_motors):
try:
v=read_sensors()
if not isinstance(v,(list,tuple)) or len(v)!=5 or any(type(x) not in (int,float) or not math.isfinite(x) or not 0<=x<=1 for x in v):raise ValueError('sensors')
except BaseException:
set_motors(0.0,0.0);raise
total=sum(v)
if total==0:left=right=0.0
else:
error=v.index(max(v))-2
left=float(max(0,min(100,self.base+self.kp*error)));right=float(max(0,min(100,self.base-self.kp*error)))
set_motors(left,right);return (left,right)

View File

@ -0,0 +1,16 @@
TARGETS=['follower.py']
import pathlib, shutil, subprocess, sys, tempfile
own=pathlib.Path(__file__).resolve().parent
try:
with tempfile.TemporaryDirectory(dir='/tmp',prefix='author1-grade-') as d:
tmp=pathlib.Path(d)
for name in TARGETS:
p=pathlib.Path(sys.argv[1])/name
if not p.is_file() or p.is_symlink():raise ValueError('missing target '+name)
shutil.copyfile(p,tmp/name)
shutil.copyfile(own/'tests.py',tmp/'_tests.py')
p=subprocess.run([sys.executable,'-I','-B',str(tmp/'_tests.py')],cwd=tmp,text=True,capture_output=True,timeout=25)
if p.returncode or p.stdout.strip()!='VERIFIED_ALL_CASES':raise ValueError((p.stdout+p.stderr)[-900:])
print('PASS')
except Exception as exc:
print('FAIL:',str(exc).replace(str(tmp),'<scratch>') if 'tmp' in globals() else str(exc));sys.exit(1)

View File

@ -0,0 +1,18 @@
import math
class LineFollower:
def __init__(self,base_speed=30.0,kp=10.0):
if type(base_speed) not in (int,float) or not math.isfinite(base_speed) or not 0<=base_speed<=100 or type(kp) not in (int,float) or not math.isfinite(kp) or kp<0:raise ValueError('configuration')
self.base=float(base_speed);self.kp=float(kp)
def step(self,read_sensors,set_motors):
try:
v=read_sensors()
if not isinstance(v,(list,tuple)) or len(v)!=5 or any(type(x) not in (int,float) or not math.isfinite(x) or not 0<=x<=1 for x in v):raise ValueError('sensors')
except BaseException:
set_motors(0.0,0.0);raise
total=sum(v)
if total==0:left=right=0.0
else:
error=sum(w*x for w,x in zip((-2,-1,0,1,2),v))/total
left=float(max(0,min(100,self.base+self.kp*error)));right=float(max(0,min(100,self.base-self.kp*error)))
set_motors(left,right);return (left,right)

View File

@ -0,0 +1,2 @@
"""Implement README.md."""
raise NotImplementedError

View File

@ -0,0 +1,3 @@
当前目录是待实现的空工程。请按 README.md 完成五路灰度循线比例控制,并自行验证。
最终只提交并评分以下文件:`follower.py`。

View File

@ -0,0 +1,95 @@
# 任务规格:S500_a1_35c5ce8f8465
本规格按[任务制作规范](../../docs/task-authoring.md)记录路线A导入审阅;不得提供给Agent。
## 1. 身份与来源
| 字段 | 填写 |
| --- | --- |
| task-id / 任务版本 | S500_a1_35c5ce8f8465 / 1 |
| 作者 / 业务评审人 / 独立评审人 | 本批Codex(导入);业务评审未单独指定 / Codex review_1 |
| 审核状态 | ADMITTED |
| 原项目与冻结版本 | Bos-qiao task_factory_500_20260914;本批本地只读复制快照 `factory/20261014-A-bos-direct/bos-envs/S500_a1_35c5ce8f8465/`;原task.json SHA-256 `c32741ce2a74fe2d7ca2832f2d09cd9f7b0e4069554175112c8498689439667d` |
| 来源路线 | A 导入成熟任务;原创新建软件任务(bos.origin未声明;依据公开README的新建空工程声明,不标为历史现场) |
| 来源 ID | `S500_a1_35c5ce8f8465`;source_qid=`q_e359a3d235c5ce8f8465` |
| 来源与授权 | 原来源 `/mnt/data/ming01.wang/Bos-qiao/envs/S500_a1_35c5ce8f8465`;依用户2026-10-12接入/试跑授权与本次作业A授权,只保留导入软件题材料与控制侧证据 |
| 隐私处理 | 无真实用户会话原文;不读/复制私有凭据;公开材料仅任务软件契约和部署路径 |
| 替身测试 | 软件:实现归一化五路灰度加权中心比例控制及异常停机;空标准工程为原新建需求。 用合成输入、受控回调/PTY或本机只读软件接口可复现;不依赖实体外设 |
| split 分组键 | `bos-qiao/task_factory_500_20260914/envs/S500_a1_35c5ce8f8465`(同源变体使用同组,不把重复采样算新题) |
| 来源组 / 根因族 | Bos-qiao/q_e359a3d235c5ce8f8465 / `five-channel-weighted-line-proportional-control` |
| 与已有任务的关系 | 独立问题;在本批与既有题的公开需求/根因对照中未发现仅参数或问法替换的重复。 |
| 验证范围 | 下列验收矩阵中的确定软件行为;family=`motor-sensor-control`;最终仅冻结 `follower.py` |
| 未验证范围 | 未单独覆盖tuple输入、停止回调同时异常和所有非法有限数边界;不证明真实传感器、电机、PID闭环或车体循线。 |
实际问题说明:实现归一化五路灰度加权中心比例控制及异常停机;空标准工程为原新建需求。
## 2. 初态与环境
沿用 Bos 原题。
## 3. 公开题面与接口
沿用 Bos 原题。
## 4. 验收矩阵与合法解
| 需求 | 输入/场景 | 可观察结果与允许误差 | 检查方法 | 公开/受保护 | 超时 |
| --- | --- | --- | --- | --- | --- |
| 加权中心与限幅 | base/kp=(30,10)/(0,100)/(100,100)/(30,.25);中心/左右/全1和每组180个种子600向量 | error=(-2*v0-v1+v3+2*v4)/sum;left/right按公开公式限到0..100;float二元tuple | 独立公式+math.isclose(abs_tol=1e-10,默认rel_tol=1e-9) | 规则公开;具体输入/期望受保护 | 评分总120秒;内部25/30秒 |
| 丢线与回调事务 | 五个0;正常每个向量 | 丢线返回(0.0,0.0);每次先read一次再set一次,回传同一数值对 | 记录顺序与类型 | 规则公开;具体输入/期望受保护 | 评分总120秒;内部25/30秒 |
| 读数/构造校验 | 长度0/4/6、bool、2、NaN、字符串;base bool/-1/101、kp负/inf | ValueError;非法读数先且仅发送停止;非法构造拒绝 | 异常类型及回调列表 | 规则公开;具体输入/期望受保护 | 评分总120秒;内部25/30秒 |
| 异常传播 | read抛RuntimeError(sensor);set抛RuntimeError(output) | read失败先停止一次;set失败原消息传播且不重复发送 | 受控回调事件断言 | 规则公开;具体输入/期望受保护 | 评分总120秒;内部25/30秒 |
- 合法替代解:valid_alternative将权重求和改为展开式;其他等价加权中心计算/限幅实现均可。验收不要求私有字段名或具体类内部结构。
- 不允许的捷径:cheat丢线继续base速度由全0向量拒绝;cheat_winner_takes_all忽略灰度权重由多路非零随机向量拒绝。
- 独立复算与冻结:check.py将声明目标复制到干净/tmp再运行/编译;正式期望来自公开约定的独立字节/数学/受控事件计算,不运行gold求答案,不复用Agent中间产物。判据附件为 `check.py`, `tests.py`。
- 判断范围与覆盖限制:未单独覆盖tuple输入、停止回调同时异常和所有非法有限数边界;不证明真实传感器、电机、PID闭环或车体循线。
- grader输出:统一verify.py返回一行JSON;PASS/FAIL checks=1、退出0/1,环境/基础设施错误为UNVERIFIED/INFRA_ERROR、退出2;内部行为断言全部完成才PASS。
- 隔离复核:沿用固定runtime的Agent/评分分容器、candidate冻结只读与UID65534探针;Docker校准回执才证明正式探针及清理。verify.py内已有同UID原checker与候选共进程的shortcut,不能据此宣称已具备独立UID的恶意checker防篡改。
<!-- shortcut: 现有Bos评分包装器的check.py与候选同UID且共享checker进程,正式UID探针不等于完整恶意候选防篡改;升级条件为独立checker身份/进程边界合同准入。 -->
- 历史答案可见性:Agent仅得到prompt与template,spec/grader/gold不放入公开工作区;历史归档不由本题挂载。spec不进入task digest。
## 5. 四类样例与校准
沿用 Bos 原题。
## 6. 真实试采与证据
沿用 Bos 原题。
## 7. 评审与准入
- [x] 独立内容评审完成;来源、替身测试、公开行为、判据、合法解、负例及局限见第 1、4 节和下方采样前记录。
- [x] 最终 task/verifier/image 身份的正式 Docker 校准,每类样例各五次符合预期,整体 PASS、cleanup confirmed。
- [x] GLM-5.3 / Pi 0.73.1 首次真实试采正常退出,host/native/broker 对账和工具反馈因果闭合 VERIFIED,清理 confirmed。
- [x] 冻结候选在新无网评分容器重判一致,按 rollout.archive 显式证据合同重新核对后本地封存。
- [x] 重复来源、根因、软件边界和现有隔离局限已登记;闸门 4 尚未执行。
最终审核状态:**ADMITTED**。任务失败与轨迹完整分开记录,首次任务 verdict 为 **PASS**,不据单次结果作难度分池。人工干预:只读诊断与控制侧评审;未向运行 Agent 注入判据/答案或修改候选。
最终身份:family `motor-sensor-control`;task digest `0ed00682361ce51e2203a67d5e1450dcc09271a9d6cb597355526e08a91dd51e`;verifier digest `2ce99a20ed07a4b16716568ee1efbd08f35e09b1300468ab443896171bde1cfe`;固定 runtime / actual image ID `sha256:aa5d00779fe03e9eadbf29c030711a4ecd1e5f894255868b5ca451a425299efc`;模型 `glm-5.3`、每次回复 max_tokens=131072、sampling=service-default,单题 Agent 防挂死期限 3600 秒。
最终正式校准:`/mnt/data/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-calibration/calibration-d763da3f46b74ff58520368d15513d80`;本地原执行路径 `/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-calibration/calibration-d763da3f46b74ff58520368d15513d80`。首次 controller:`/mnt/data/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct/pi-f32fb8ca5c9c4104babe93130e681c93`;原路径 `/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct/pi-f32fb8ca5c9c4104babe93130e681c93`。证据 `VERIFIED`,Agent exit=0 / termination=exit,cleanup confirmed,独立重判一致。完整轨迹与独立评分见本批冷数据 `factory/20261014-A-bos-direct/report.json`、`admission.jsonl`。
本地封存包:`/mnt/data/ming01.wang/embedded-rollout/artifacts/20261014-A-bos-direct/pi-f32fb8ca5c9c4104babe93130e681c93.tar.gz`;SHA-256 `76ec111fa20e32abfb3eb89053043b224e3be30f2861d21221156c115fb8a5c7`,426374 字节。Salty 上传/下载摘要核对 **BLOCKED**:执行主机没有可用已认证 CLI 或连接;没有声明已上传。此阻塞不替代已完成的本地封存事实。冷迁按文件数与总字节核对,见批次 `migration-verification.json`;回执原字节和原绝对路径保持,当前入口不能直接对迁移后的路径重跑,需显式恢复原路径或升级路径合同。
### 采样前独立内容评审记录
下方保留采样前状态和当时的证据。若包含旧归族校准,只用于历史事实;当前准入由上述最终身份、校准和完整试采授权。
- [x] 2026-10-14,Codex review_1已逐项静态核对task/prompt/template、check及附件、gold/valid/全部cheat;未执行来源脚本、未改来源或平台。
- [x] 公开README与隐藏行为逐项对照;矩阵内受检正确行为均有公开依据,初态桩无gold泄露,目标自包含,软件替身范围明确。
- [x] root cause与来源组记录,已对照当前任务公开主题确认解题路径不同;合法替代解按行为而非内部补丁评审。
- [x] 固定镜像正式五次/样例校准整体PASS且cleanup confirmed(仅根据已见实际回执)。
- [ ] 闸门3真实试采的工具轨迹、独立评分、VERIFIED对账与清理:本评审尚未核对;由批次controller补齐,不能据本规格宣称ADMITTED。
- [ ] 闸门4≥5次同配置难度分池:本批作业A不执行。
样例合同:gold=PASS、template=FAIL、noop=FAIL、cheat=FAIL、cheat_winner_takes_all=FAIL、valid_alternative=PASS;各类需正式重复5次,不能以静态审阅代替。
校准证据:
- 可信本地开发:`/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-dev/calibration-8b0a7d0045a94c70937e19fa93a0d422`;verdict=PASS,cleanup=confirmed,trials=30;task_digest=`0ed00682361ce51e2203a67d5e1450dcc09271a9d6cb597355526e08a91dd51e`;实际image_id=`None`。
- 固定镜像Docker:`/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-calibration/calibration-d763da3f46b74ff58520368d15513d80`;verdict=PASS,cleanup=confirmed,trials=30;task_digest=`0ed00682361ce51e2203a67d5e1450dcc09271a9d6cb597355526e08a91dd51e`;实际image_id=`sha256:aa5d00779fe03e9eadbf29c030711a4ecd1e5f894255868b5ca451a425299efc`。
独立结论:APPROVED_FOR_CALIBRATION/PILOT(题目内容),人工生命周期状态为CALIBRATED,非ADMITTED。任务导入摘要 `0ed00682361ce51e2203a67d5e1450dcc09271a9d6cb597355526e08a91dd51e`;正式镜像固定为 `ccr-29eug8s3-pub.cnc.bj.baidubce.com/algo/bench-runtime@sha256:aa5d00779fe03e9eadbf29c030711a4ecd1e5f894255868b5ca451a425299efc`。限制:未单独覆盖tuple输入、停止回调同时异常和所有非法有限数边界;不证明真实传感器、电机、PID闭环或车体循线。 后续以最终正式校准与真实试采证据更新本节。

View File

@ -0,0 +1,36 @@
{
"id": "S500_a1_35c5ce8f8465",
"version": "1",
"family": "motor-sensor-control",
"source": "bos-qiao/task_factory_500_20260914/envs/S500_a1_35c5ce8f8465",
"template": "template",
"verifier": "verify.py",
"candidate_files": [
"follower.py"
],
"timeout_seconds": 120,
"fixtures": {
"gold": "PASS",
"template": "FAIL",
"noop": "FAIL",
"cheat": "FAIL",
"cheat_winner_takes_all": "FAIL",
"valid_alternative": "PASS"
},
"grader_files": [
"check.py",
"tests.py"
],
"bos": {
"source_qid": "q_e359a3d235c5ce8f8465",
"origin": null,
"image": "python@sha256:78387bc3881b8273120a12ebe6c1ab22b018ccc2c9adf565ae1ac9b536e184ea",
"budget_s": 600,
"fixture_scripts": {
"gold": "gold.sh",
"cheat": "cheat_lost_line_forward.sh",
"cheat_winner_takes_all": "cheat_winner_takes_all.sh",
"valid_alternative": "valid_alternative.sh"
}
}
}

View File

@ -0,0 +1,8 @@
# 五路灰度循线比例控制
这是依据原始新建代码需求创建的空标准工程,不是历史故障现场恢复。
实现 follower.py 的 LineFollower(base_speed=30.0,kp=10.0) 与 step(read_sensors,set_motors)->(left,right)。这是原新建循线控制代码需求的软件控制层,5路灰度传感器由回调提供,不读取物理GPIO;set_motors为调用者实际驱动可接的输出接口。
左到右5路read_sensors()读数:0为白色、1为黑色,中间值为归一化黑线强度;每次step只读一次。输入为长度5的list/tuple,每项有限int/float且0<=v<=1,拒绝bool。非法读数抛ValueError,且先调用set_motors(0.0,0.0)一次以停止;read_sensors本身异常也先停止再原样传播。若停止回调也异常,则传播停止回调异常。
明确采用加权中心+比例控制,不声称未定义的PID:权重[-2,-1,0,1,2],total=sum(values),total==0表示丢线、输出(0.0,0.0);否则error=sum(weight*value)/total,left=clamp(base_speed+kp*error,0,100),right=clamp(base_speed-kp*error,0,100)。base_speed有限数[0,100]、kp有限非负数,拒绝bool;构造非法抛ValueError。
step调用set_motors(left,right)恰好一次后返回同一对float值;不在调用成功前伪造返回;set_motors异常原样传播,不额外重复发送。所有速度以百分比计,额外硬件配置/实际循线效果不在验收内。可传入合成5路序列和记录输出的list回调自测中心/左/右/丢线。
只依赖标准库;不访问硬件或网络。可在当前目录编写自己的自测;额外文件不影响判分。

View File

@ -0,0 +1,2 @@
"""Implement README.md."""
raise NotImplementedError

View File

@ -0,0 +1,40 @@
import sys, pathlib, random, math, json, struct
sys.path.insert(0,str(pathlib.Path(__file__).resolve().parent))
import follower as f
r=random.Random(600)
for base,kp in [(30,10),(0,100),(100,100),(30,.25)]:
c=f.LineFollower(base,kp)
for v in [[0,0,0,0,0],[1,0,0,0,0],[0,0,1,0,0],[0,0,0,0,1],[1,1,1,1,1]]+[[r.random() for _ in range(5)] for _ in range(180)]:
ev=[]
def read():ev.append('read');return v
def set_(l,r):ev.append((l,r))
got=c.step(read,set_);den=sum(v)
if den==0:exp=(0.,0.)
else:
err=(-2*v[0]-v[1]+v[3]+2*v[4])/den;exp=(max(0.,min(100.,base+kp*err)),max(0.,min(100.,base-kp*err)))
assert ev==['read',got] and type(got) is tuple and len(got)==2 and all(type(x) is float for x in got)
assert all(math.isclose(x,y,abs_tol=1e-10) for x,y in zip(got,exp))
for bad in [[],[0]*4,[0]*6,[True,0,0,0,0],[2,0,0,0,0],[float('nan'),0,0,0,0],'00100']:
ev=[]
try:f.LineFollower().step(lambda:bad,lambda l,r:ev.append((l,r)))
except ValueError:pass
else:raise AssertionError('bad readings')
assert ev==[(0.,0.)]
for base,kp in [(True,1),(-1,1),(101,1),(1,-1),(1,float('inf'))]:
try:f.LineFollower(base,kp)
except ValueError:pass
else:raise AssertionError('bad config')
ev=[]
def broken():raise RuntimeError('sensor')
try:f.LineFollower().step(broken,lambda l,r:ev.append((l,r)))
except RuntimeError as ex:assert str(ex)=='sensor'
else:raise AssertionError('sensor error')
assert ev==[(0.,0.)]
ev=[]
def output(l,r):ev.append((l,r));raise RuntimeError('output')
try:f.LineFollower().step(lambda:[0,0,1,0,0],output)
except RuntimeError as ex:assert str(ex)=='output'
else:raise AssertionError('output error')
assert ev==[(30.,30.)]
print('VERIFIED_ALL_CASES')

View File

@ -0,0 +1,18 @@
import math
class LineFollower:
def __init__(self,base_speed=30.0,kp=10.0):
if type(base_speed) not in (int,float) or not math.isfinite(base_speed) or not 0<=base_speed<=100 or type(kp) not in (int,float) or not math.isfinite(kp) or kp<0:raise ValueError('configuration')
self.base=float(base_speed);self.kp=float(kp)
def step(self,read_sensors,set_motors):
try:
v=read_sensors()
if not isinstance(v,(list,tuple)) or len(v)!=5 or any(type(x) not in (int,float) or not math.isfinite(x) or not 0<=x<=1 for x in v):raise ValueError('sensors')
except BaseException:
set_motors(0.0,0.0);raise
total=sum(v)
if total==0:left=right=0.0
else:
error=(-2*v[0]-v[1]+v[3]+2*v[4])/total
left=float(max(0,min(100,self.base+self.kp*error)));right=float(max(0,min(100,self.base-self.kp*error)))
set_motors(left,right);return (left,right)

View File

@ -0,0 +1,93 @@
"""Run an imported Bos-qiao check.py unchanged, with the candidate user's privileges."""
import json
import os
from pathlib import Path
import shutil
import subprocess
import sys
import tempfile
GRADER = Path(__file__).resolve().parent
# Runs the original check unchanged; failing child processes also copy their stderr tail out for private triage.
CHECK = (
"import runpy, subprocess, sys\n"
"real = subprocess.run\n"
"def run(*args, **kwargs):\n"
" result = real(*args, **kwargs)\n"
" if result.returncode and result.stderr:\n"
" tail = result.stderr if isinstance(result.stderr, str) else result.stderr.decode('utf-8', 'replace')\n"
" sys.__stderr__.write(tail[-2000:] + '\\n')\n"
" return result\n"
"subprocess.run = run\n"
"sys.argv = sys.argv[1:]\n"
"runpy.run_path(sys.argv[0], run_name='__main__')\n")
class UnsupportedEnvironment(RuntimeError):
pass
def candidate_user():
os.setgroups([])
os.setgid(65534)
os.setuid(65534)
def isolation():
if os.environ.get("ROLLOUT_GRADE_ISOLATION") != "uid65534":
return None
if sys.platform != "linux" or os.geteuid() != 0:
raise UnsupportedEnvironment("isolated grader requires a Linux root supervisor")
probe = (
"from pathlib import Path\n"
"try: Path('/grader/verify.py').read_bytes()\n"
"except PermissionError: raise SystemExit(0)\n"
"raise SystemExit(1)\n"
)
try:
result = subprocess.run([sys.executable, "-I", "-c", probe],
capture_output=True, timeout=10, preexec_fn=candidate_user)
except subprocess.SubprocessError as error:
raise UnsupportedEnvironment("candidate identity could not be isolated") from error
if result.returncode != 0:
raise UnsupportedEnvironment("candidate can read the private grader or isolation probe failed")
return candidate_user
def verify(candidate):
child_user = isolation()
if shutil.which("gcc") is None:
raise UnsupportedEnvironment("C compiler unavailable")
with tempfile.TemporaryDirectory(prefix="bos-grade-") as temporary:
root = Path(temporary)
work, grader = root / "work", root / "grader"
shutil.copytree(candidate, work, symlinks=True)
shutil.copytree(GRADER, grader, ignore=lambda folder, names: ["verify.py"] if Path(folder) == GRADER else [])
if child_user:
os.chmod(root, 0o700) # Without CAP_FOWNER, root cannot chmod after giving files away.
for path in (root, *root.rglob("*")):
os.lchown(path, 65534, 65534)
# shortcut: the original check and the candidate share UID 65534, so a hostile candidate could
# tamper with its checker process; upgrade by running check.py and its tests under separate UIDs.
completed = subprocess.run([sys.executable, "-I", "-B", "-c", CHECK, str(grader / "check.py"), str(work)],
cwd=root, capture_output=True, text=True, timeout=100, preexec_fn=child_user)
# grade.stderr stays on the trusted side with the archive; it may quote hidden test lines.
sys.stderr.write(completed.stderr[-4000:])
lines = completed.stdout.splitlines()
if completed.returncode == 0 and lines[-1:] == ["PASS"]:
return {"status": "PASS", "checks": 1}
return {"status": "FAIL", "checks": 1, "reason": (lines[-1] if lines else "no checker output")[:300]}
if __name__ == "__main__":
try:
result = verify(Path(sys.argv[1]))
except UnsupportedEnvironment as error:
result = {"status": "UNVERIFIED", "checks": 0, "reason": str(error)}
except subprocess.TimeoutExpired:
result = {"status": "FAIL", "checks": 1, "reason": "timeout"}
except Exception as error:
result = {"status": "INFRA_ERROR", "checks": 0, "reason": str(error)}
print(json.dumps(result, sort_keys=True))
sys.exit({"PASS": 0, "FAIL": 1}.get(result["status"], 2))

View File

@ -0,0 +1,24 @@
import math
class ChassisDriver:
def __init__(self,write):
if not callable(write):raise ValueError('write')
self.write=write;self.last_command=(0.0,0.0)
def _apply(self,speed,left,right):
if type(speed) not in (int,float) or not math.isfinite(speed) :raise ValueError('speed')
speed=max(0,min(100,speed));cmd=(float(speed)*left,float(speed)*right);self.write(*cmd);self.last_command=cmd
def forward(self,speed):
"""Command both wheels forward."""
self._apply(speed,1,1)
def backward(self,speed):
"""Command both wheels backward."""
self._apply(speed,-1,-1)
def turn_left(self,speed):
"""Rotate left in place."""
self._apply(speed,-1,1)
def turn_right(self,speed):
"""Rotate right in place."""
self._apply(speed,1,-1)
def stop(self):
"""Stop both wheels."""
self._apply(0,1,1)

View File

@ -0,0 +1,24 @@
import math
class ChassisDriver:
def __init__(self,write):
if not callable(write):raise ValueError('write')
self.write=write;self.last_command=(0.0,0.0)
def _apply(self,speed,left,right):
if type(speed) not in (int,float) or not math.isfinite(speed) or not 0<=speed<=100:raise ValueError('speed')
cmd=(float(speed)*left,float(speed)*right);self.write(*cmd);self.last_command=cmd
def forward(self,speed):
"""Command both wheels forward."""
self._apply(speed,1,1)
def backward(self,speed):
"""Command both wheels backward."""
self._apply(speed,-1,-1)
def turn_left(self,speed):
"""Rotate left in place."""
self._apply(speed,0,1)
def turn_right(self,speed):
"""Rotate right in place."""
self._apply(speed,1,-1)
def stop(self):
"""Stop both wheels."""
self._apply(0,1,1)

View File

@ -0,0 +1,16 @@
TARGETS=['chassis.py']
import pathlib, shutil, subprocess, sys, tempfile
own=pathlib.Path(__file__).resolve().parent
try:
with tempfile.TemporaryDirectory(dir='/tmp',prefix='author1-grade-') as d:
tmp=pathlib.Path(d)
for name in TARGETS:
p=pathlib.Path(sys.argv[1])/name
if not p.is_file() or p.is_symlink():raise ValueError('missing target '+name)
shutil.copyfile(p,tmp/name)
shutil.copyfile(own/'tests.py',tmp/'_tests.py')
p=subprocess.run([sys.executable,'-I','-B',str(tmp/'_tests.py')],cwd=tmp,text=True,capture_output=True,timeout=25)
if p.returncode or p.stdout.strip()!='VERIFIED_ALL_CASES':raise ValueError((p.stdout+p.stderr)[-900:])
print('PASS')
except Exception as exc:
print('FAIL:',str(exc).replace(str(tmp),'<scratch>') if 'tmp' in globals() else str(exc));sys.exit(1)

View File

@ -0,0 +1,24 @@
import math
class ChassisDriver:
def __init__(self,write):
if not callable(write):raise ValueError('write')
self.write=write;self.last_command=(0.0,0.0)
def _apply(self,speed,left,right):
if type(speed) not in (int,float) or not math.isfinite(speed) or not 0<=speed<=100:raise ValueError('speed')
cmd=(float(speed)*left,float(speed)*right);self.write(*cmd);self.last_command=cmd
def forward(self,speed):
"""Command both wheels forward."""
self._apply(speed,1,1)
def backward(self,speed):
"""Command both wheels backward."""
self._apply(speed,-1,-1)
def turn_left(self,speed):
"""Rotate left in place."""
self._apply(speed,-1,1)
def turn_right(self,speed):
"""Rotate right in place."""
self._apply(speed,1,-1)
def stop(self):
"""Stop both wheels."""
self._apply(0,1,1)

View File

@ -0,0 +1,2 @@
"""Implement README.md."""
raise NotImplementedError

View File

@ -0,0 +1,3 @@
当前目录是待实现的空工程。请按 README.md 完成差速底盘命令映射与输出事务,并自行验证。
最终只提交并评分以下文件:`chassis.py`。

View File

@ -0,0 +1,92 @@
# 任务规格:S500_a1_42207d48c1f6
规范:[任务制作规范](../../docs/task-authoring.md)。本规格仅供可信作者与评审,不进入 template 或 Agent 题面。
## 1. 身份与来源
| 字段 | 填写 |
| --- | --- |
| task-id / 任务版本 | S500_a1_42207d48c1f6 / 1 |
| 作者 / 业务评审人 / 独立评审人 | 本批 Codex / 原 Bos 公开契约业务审查 / Codex runner_audit |
| 审核状态 | ADMITTED |
| 原项目与冻结版本 | Bos-qiao;本批本地复制快照 `factory/20261014-A-bos-direct/bos-envs/S500_a1_42207d48c1f6/`;来源 task.json SHA-256 `b9727adaf021d2c8cfb91a332fe4b636b35031f30d8cf464b07e94b36be8fdb2` |
| 来源路线 | A 导入成熟任务;原创构题;bos.origin 缺失(null),公开 README 明确为依据原始新建代码需求创建空工程,不是历史故障现场恢复;不把 source_qid 当作历史初态证据。 |
| 来源 ID | S500_a1_42207d48c1f6;source_qid=q_5ce038e542207d48c1f6 |
| 来源与授权 | Bos `envs/S500_a1_42207d48c1f6`;用户 2026-10-12 项目接入与试跑授权,2026-10-14 明确授权本批 Bos direct 导入;只保留任务资产,不复制来源工具或凭据 |
| 隐私处理 | 未引入真实会话原文、个人信息或密钥;源码/判据为本题软件工程资产 |
| 替身测试 | software;注入 write(left,right) 记录回调与失败回调后,方向映射、非法速度拒绝和状态事务仍可重现,不要求真实底盘。 |
| split 分组键 | bos-qiao/envs/S500_a1_42207d48c1f6;同源衍生题应共用此键 |
| 来源组 / 根因族 | S500_a1_42207d48c1f6 / differential-chassis-command-transaction;采集 family=motor-sensor-control |
| 与已有任务的关系 | 本批新题;ID 与公开契约摘要无完全重复,未发现与原 20 题候选/gold 的精确重复;本题机制是 differential-chassis-command-transaction,跨候选语义去重仍由批次汇总复核 |
| 验证范围 | 有符号轮速百分比映射、一次输出调用、float类型、last_command事务与实例独立性。 |
| 未验证范围 | 厂商串口报文、电机电气、板端运行、实体转向与实时性。 |
实际问题说明:空 chassis.py 未实现 ChassisDriver;应将前进/后退/原地转向映射为明确的两轮float命令,先成功输出再更新状态。
## 2. 初态与环境
沿用 Bos 原题。
## 3. 公开题面与接口
沿用 Bos 原题。
## 4. 验收矩阵与合法解
| 需求 | 输入/场景 | 可观察结果与允许误差 | 检查方法 | 公开/受保护 | 超时 |
| --- | --- | --- | --- | --- | --- |
| 四方向映射 | speed=0、100、0.001、33.3及固定种子3545的每方向100个随机速度 | forward=(s,s)、backward=(-s,-s)、left=(-s,s)、right=(s,-s),每次write恰一次且数值均float | 记录回调,独立符号乘积逐项与last_command比对 | README公开行为;tests.py保护输入 | 25秒整体 |
| 参数校验 | -1、100.001、bool、None、字符串、NaN、inf;不可调用write构造参数 | ValueError;不得调用write或改last_command | 保存事件/状态后逐项验证异常与不变性 | README公开;隐藏非法输入 | 25秒整体 |
| 成功/失败事务 | stop();backward(7)输出回调抛RuntimeError;两个实例交错调用 | stop=(0.0,0.0);失败只发生一次(-7.0,-7.0)回调且状态仍初值;实例独立 | 真实受控回调记录、异常捕获与状态断言 | README公开;tests.py保护 | 25秒整体 |
| 公开方法文档 | forward/backward/turn_left/turn_right/stop | 每个公开方法存在非空docstring | 反射检查__doc__,不指定实现文本 | README公开 | 25秒整体 |
- 合法替代解:valid_alternative/chassis.py 用tuple符号生成替代直接tuple表达式;接口和状态提交顺序一致。其他映射表/分支实现也可通过;替代例结构差异较小。
- 不允许的捷径与拦截:cheat 把非法速度钳入[0,100],由-1/100.001必须ValueError且事件不变抓住;cheat_turn_arc 把左转写成(0,s),由独立(-s,s)期望抓住;noop/template 未实现导入失败。
- 判据独立性:tests.py直接由公开方向符号计算float(s)*sign,使用记录list和失败回调检查次数与状态,不导入gold作期望。
- 判断范围:全部必需行为通过才可 PASS;只编译/导入成功不够。grader JSON 为 PASS/FAIL/UNVERIFIED/INFRA_ERROR,checks 为非负整数,PASS/FAIL 的 checks>0,退出码 0/1/2/2。
- 隔离复核:Agent 工作区只有 template 与公开题面;gold、判据与 spec 不挂载 Agent。正式 verify.py 探测 UID 65534 无法读取 /grader/verify.py,并由 controller 单独无网评分。当前 Bos 包装器会将原 check/tests 与候选放在同 UID 的临时评分工作区;这不证明候选代码无法读取或篡改该临时 checker。
- shortcut: Bos 原 check/tests 与候选共享 UID 65534,尚不能证明任意恶意候选的全面隔离;升级为不同 checker/candidate 身份并增加抗篡改行为探针后重新准入。
- 历史答案可见性:Agent 不挂载 factory、runs、grader、gold、Salty 历史归档或本规格;本题采样尚未完成,实际挂载与终态仍需 controller 证据复核。
- 已知覆盖限制:回调异常只实际测试RuntimeError类型,尚未逐项验证任意异常对象身份;合法替代解仅一个且结构差异小;不检查真实硬件输出。
## 5. 四类样例与校准
沿用 Bos 原题。
## 6. 真实试采与证据
沿用 Bos 原题。
## 7. 评审与准入
- [x] 独立内容评审完成;来源、替身测试、公开行为、判据、合法解、负例及局限见第 1、4 节和下方采样前记录。
- [x] 最终 task/verifier/image 身份的正式 Docker 校准,每类样例各五次符合预期,整体 PASS、cleanup confirmed。
- [x] GLM-5.3 / Pi 0.73.1 首次真实试采正常退出,host/native/broker 对账和工具反馈因果闭合 VERIFIED,清理 confirmed。
- [x] 冻结候选在新无网评分容器重判一致,按 rollout.archive 显式证据合同重新核对后本地封存。
- [x] 重复来源、根因、软件边界和现有隔离局限已登记;闸门 4 尚未执行。
最终审核状态:**ADMITTED**。任务失败与轨迹完整分开记录,首次任务 verdict 为 **PASS**,不据单次结果作难度分池。人工干预:只读诊断与控制侧评审;未向运行 Agent 注入判据/答案或修改候选。
最终身份:family `motor-sensor-control`;task digest `2aef30d0e388782ffc272d731db8a14933233847d8e15b47fd3c5ac45802977e`;verifier digest `2ce99a20ed07a4b16716568ee1efbd08f35e09b1300468ab443896171bde1cfe`;固定 runtime / actual image ID `sha256:aa5d00779fe03e9eadbf29c030711a4ecd1e5f894255868b5ca451a425299efc`;模型 `glm-5.3`、每次回复 max_tokens=131072、sampling=service-default,单题 Agent 防挂死期限 3600 秒。
最终正式校准:`/mnt/data/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-calibration/calibration-4a0706b6725f41f3b32ce52f462e7a28`;本地原执行路径 `/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-calibration/calibration-4a0706b6725f41f3b32ce52f462e7a28`。首次 controller:`/mnt/data/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct/pi-ece866e4e0ac4f0793072cb500a461df`;原路径 `/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct/pi-ece866e4e0ac4f0793072cb500a461df`。证据 `VERIFIED`,Agent exit=0 / termination=exit,cleanup confirmed,独立重判一致。完整轨迹与独立评分见本批冷数据 `factory/20261014-A-bos-direct/report.json`、`admission.jsonl`。
本地封存包:`/mnt/data/ming01.wang/embedded-rollout/artifacts/20261014-A-bos-direct/pi-ece866e4e0ac4f0793072cb500a461df.tar.gz`;SHA-256 `418ff58ab800b2cfe30d4bc5bba89ce3c3faba122507c7d317bc71acc7b63b11`,427516 字节。Salty 上传/下载摘要核对 **BLOCKED**:执行主机没有可用已认证 CLI 或连接;没有声明已上传。此阻塞不替代已完成的本地封存事实。冷迁按文件数与总字节核对,见批次 `migration-verification.json`;回执原字节和原绝对路径保持,当前入口不能直接对迁移后的路径重跑,需显式恢复原路径或升级路径合同。
### 采样前独立内容评审记录
下方保留采样前状态和当时的证据。若包含旧归族校准,只用于历史事实;当前准入由上述最终身份、校准和完整试采授权。
- [x] 已逐项对照实际 prompt/template/check/tests/gold/valid/cheat 核对公开行为、完整初态、输入、独立复算和合法解;结论:未发现未公开的必需行为或限定唯一补丁的判据。
- [x] 正式环境每类五次符合预期、整体 PASS、cleanup confirmed;仅记录以下真实证据。
- [ ] 真实试采及工具因果、终态、独立评分和清理尚未完成;不得标 ADMITTED。
- [x] 来源与根因分组已记录,未验证范围、同 UID 限制及合法解审查边界已明确。
- [x] 独立只读行为评审已完成(2026-10-14,Codex runner_audit);评审没有执行来源脚本或候选、修改平台或来源、读取凭据。
审核记录:2026-10-14,CALIBRATED。task digest `2aef30d0e388782ffc272d731db8a14933233847d8e15b47fd3c5ac45802977e`;固定正式镜像 `ccr-29eug8s3-pub.cnc.bj.baidubce.com/algo/bench-runtime@sha256:aa5d00779fe03e9eadbf29c030711a4ecd1e5f894255868b5ca451a425299efc`。样例类别与预期沿用 task.json 声明,不把负例 FAIL 写成任务能力失败。
- local:`/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-dev/calibration-f4ad54e65f2f4423b55510b447061a3a`;verdict=PASS;cleanup=confirmed;30 次。
- docker:`/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-calibration/calibration-4a0706b6725f41f3b32ce52f462e7a28`;verdict=PASS;cleanup=confirmed;30 次。
- 评审台账:`factory/20261014-A-bos-direct/review-2.jsonl`。允许在正式校准 PASS 后进入一次真实试采;试采失败不单独否决有效任务。
只在闸门 2、独立评审和闸门 3 证据齐备后由批次控制端更新准入状态;本规格不替代运行回执。

View File

@ -0,0 +1,36 @@
{
"id": "S500_a1_42207d48c1f6",
"version": "1",
"family": "motor-sensor-control",
"source": "bos-qiao/task_factory_500_20260914/envs/S500_a1_42207d48c1f6",
"template": "template",
"verifier": "verify.py",
"candidate_files": [
"chassis.py"
],
"timeout_seconds": 120,
"fixtures": {
"gold": "PASS",
"template": "FAIL",
"noop": "FAIL",
"cheat": "FAIL",
"cheat_turn_arc": "FAIL",
"valid_alternative": "PASS"
},
"grader_files": [
"check.py",
"tests.py"
],
"bos": {
"source_qid": "q_5ce038e542207d48c1f6",
"origin": null,
"image": "python@sha256:78387bc3881b8273120a12ebe6c1ab22b018ccc2c9adf565ae1ac9b536e184ea",
"budget_s": 600,
"fixture_scripts": {
"gold": "gold.sh",
"cheat": "cheat_clamp_bad_speed.sh",
"cheat_turn_arc": "cheat_turn_arc.sh",
"valid_alternative": "valid_alternative.sh"
}
}
}

View File

@ -0,0 +1,7 @@
# 差速底盘命令映射与输出事务
这是依据原始新建代码需求创建的空标准工程,不是历史故障现场恢复。
实现 chassis.py 的 ChassisDriver(write) 类与 forward(speed)、backward(speed)、turn_left(speed)、turn_right(speed)、stop()。每个公开方法有说明功能的docstring。
这是通用底盘软件命令层:write(left,right) 为调用者注入的输出接口,参数是有符号轮速百分比[-100,100];这里不虚构任何厂商串口报文,也不证明实体电机动作。公开选择差速底盘、原地转向:forward(s)输出(s,s)、backward(s)输出(-s,-s)、turn_left(s)输出(-s,s)、turn_right(s)输出(s,-s),stop输出(0.0,0.0)。每次方法调用只调用write一次,数值输出为float。
speed仅接受有限int/float,范围[0,100],bool非法;非法抛ValueError且不得调用write或改状态。构造write非callable抛ValueError。last_command公开属性初始(0.0,0.0);每次write成功后更新为本次tuple;回调抛异常则原样传播且last_command不改变。各实例独立,不使用真实时钟和网络。用自建记录list的回调自测。
只依赖标准库;不访问硬件或网络。可在当前目录编写自己的自测;额外文件不影响判分。

View File

@ -0,0 +1,2 @@
"""Implement README.md."""
raise NotImplementedError

View File

@ -0,0 +1,32 @@
import sys, pathlib, random, math, json, struct
sys.path.insert(0,str(pathlib.Path(__file__).resolve().parent))
import chassis as c
r=random.Random(3545);events=[];d=c.ChassisDriver(lambda l,r:events.append((l,r)))
assert d.last_command==(0.,0.)
for name,mult in [('forward',(1,1)),('backward',(-1,-1)),('turn_left',(-1,1)),('turn_right',(1,-1))]:
assert getattr(c.ChassisDriver,name).__doc__
for v in [0,100,.001,33.3]+[r.uniform(0,100) for _ in range(100)]:
before=len(events);getattr(d,name)(v);exp=(float(v)*mult[0],float(v)*mult[1]);assert len(events)==before+1 and events[-1]==exp and d.last_command==exp
assert all(type(x) is float for x in events[-1])
for bad in [-1,100.001,True,None,'3',float('nan'),float('inf')]:
before=list(events);state=d.last_command
try:getattr(d,name)(bad)
except ValueError:pass
else:raise AssertionError('invalid speed')
assert before==events and d.last_command==state
assert c.ChassisDriver.stop.__doc__;d.stop();assert events[-1]==(0.,0.) and d.last_command==(0.,0.)
state=d.last_command
for bad in [None,3,'write']:
try:c.ChassisDriver(bad)
except ValueError:pass
else:raise AssertionError('bad output callback')
other=c.ChassisDriver(lambda *_:None);other.forward(9);assert d.last_command==state
calls=[]
def fail(l,r):calls.append((l,r));raise RuntimeError('driver')
f=c.ChassisDriver(fail)
try:f.backward(7)
except RuntimeError:pass
else:raise AssertionError('callback exception')
assert calls==[(-7.,-7.)] and f.last_command==(0.,0.)
print('VERIFIED_ALL_CASES')

View File

@ -0,0 +1,24 @@
import math
class ChassisDriver:
def __init__(self,write):
if not callable(write):raise ValueError('write')
self.write=write;self.last_command=(0.0,0.0)
def _apply(self,speed,left,right):
if type(speed) not in (int,float) or not math.isfinite(speed) or not 0<=speed<=100:raise ValueError('speed')
cmd=tuple(float(speed)*sign for sign in (left,right));self.write(*cmd);self.last_command=cmd
def forward(self,speed):
"""Command both wheels forward."""
self._apply(speed,1,1)
def backward(self,speed):
"""Command both wheels backward."""
self._apply(speed,-1,-1)
def turn_left(self,speed):
"""Rotate left in place."""
self._apply(speed,-1,1)
def turn_right(self,speed):
"""Rotate right in place."""
self._apply(speed,1,-1)
def stop(self):
"""Stop both wheels."""
self._apply(0,1,1)

View File

@ -0,0 +1,93 @@
"""Run an imported Bos-qiao check.py unchanged, with the candidate user's privileges."""
import json
import os
from pathlib import Path
import shutil
import subprocess
import sys
import tempfile
GRADER = Path(__file__).resolve().parent
# Runs the original check unchanged; failing child processes also copy their stderr tail out for private triage.
CHECK = (
"import runpy, subprocess, sys\n"
"real = subprocess.run\n"
"def run(*args, **kwargs):\n"
" result = real(*args, **kwargs)\n"
" if result.returncode and result.stderr:\n"
" tail = result.stderr if isinstance(result.stderr, str) else result.stderr.decode('utf-8', 'replace')\n"
" sys.__stderr__.write(tail[-2000:] + '\\n')\n"
" return result\n"
"subprocess.run = run\n"
"sys.argv = sys.argv[1:]\n"
"runpy.run_path(sys.argv[0], run_name='__main__')\n")
class UnsupportedEnvironment(RuntimeError):
pass
def candidate_user():
os.setgroups([])
os.setgid(65534)
os.setuid(65534)
def isolation():
if os.environ.get("ROLLOUT_GRADE_ISOLATION") != "uid65534":
return None
if sys.platform != "linux" or os.geteuid() != 0:
raise UnsupportedEnvironment("isolated grader requires a Linux root supervisor")
probe = (
"from pathlib import Path\n"
"try: Path('/grader/verify.py').read_bytes()\n"
"except PermissionError: raise SystemExit(0)\n"
"raise SystemExit(1)\n"
)
try:
result = subprocess.run([sys.executable, "-I", "-c", probe],
capture_output=True, timeout=10, preexec_fn=candidate_user)
except subprocess.SubprocessError as error:
raise UnsupportedEnvironment("candidate identity could not be isolated") from error
if result.returncode != 0:
raise UnsupportedEnvironment("candidate can read the private grader or isolation probe failed")
return candidate_user
def verify(candidate):
child_user = isolation()
if shutil.which("gcc") is None:
raise UnsupportedEnvironment("C compiler unavailable")
with tempfile.TemporaryDirectory(prefix="bos-grade-") as temporary:
root = Path(temporary)
work, grader = root / "work", root / "grader"
shutil.copytree(candidate, work, symlinks=True)
shutil.copytree(GRADER, grader, ignore=lambda folder, names: ["verify.py"] if Path(folder) == GRADER else [])
if child_user:
os.chmod(root, 0o700) # Without CAP_FOWNER, root cannot chmod after giving files away.
for path in (root, *root.rglob("*")):
os.lchown(path, 65534, 65534)
# shortcut: the original check and the candidate share UID 65534, so a hostile candidate could
# tamper with its checker process; upgrade by running check.py and its tests under separate UIDs.
completed = subprocess.run([sys.executable, "-I", "-B", "-c", CHECK, str(grader / "check.py"), str(work)],
cwd=root, capture_output=True, text=True, timeout=100, preexec_fn=child_user)
# grade.stderr stays on the trusted side with the archive; it may quote hidden test lines.
sys.stderr.write(completed.stderr[-4000:])
lines = completed.stdout.splitlines()
if completed.returncode == 0 and lines[-1:] == ["PASS"]:
return {"status": "PASS", "checks": 1}
return {"status": "FAIL", "checks": 1, "reason": (lines[-1] if lines else "no checker output")[:300]}
if __name__ == "__main__":
try:
result = verify(Path(sys.argv[1]))
except UnsupportedEnvironment as error:
result = {"status": "UNVERIFIED", "checks": 0, "reason": str(error)}
except subprocess.TimeoutExpired:
result = {"status": "FAIL", "checks": 1, "reason": "timeout"}
except Exception as error:
result = {"status": "INFRA_ERROR", "checks": 0, "reason": str(error)}
print(json.dumps(result, sort_keys=True))
sys.exit({"PASS": 0, "FAIL": 1}.get(result["status"], 2))

View File

@ -0,0 +1,33 @@
import struct,math,functools,operator
def xor(b):return functools.reduce(operator.xor,b,0)
def encode(x_mm_s=0,z_deg_s_1000=0):
if any(type(v) is not int or not -32768<=v<=32767 for v in (x_mm_s,z_deg_s_1000)):raise ValueError('range')
b=bytes([123,0])+struct.pack('>hhh',x_mm_s,0,z_deg_s_1000)+bytes(1)
return b+bytes([xor(b),125])
def decode(frame):
if len(frame)!=24 or frame[0]!=123 or frame[-1]!=125 or xor(frame[:22])!=frame[22]:return None
return dict(x_mm_s=int.from_bytes(frame[2:4],'big',signed=True),z_deg_s=int.from_bytes(frame[6:8],'big',signed=True)/1000.0,left_enc=int.from_bytes(frame[8:10],'big',signed=True),right_enc=int.from_bytes(frame[10:12],'big',signed=True),gas_ppm=int.from_bytes(frame[20:22],'big'),sensor_ok=frame[12]==252)
class StreamDecoder:
def __init__(self):self.buffer=b''
def feed(self,chunk):
self.buffer+=chunk;out=[]
while self.buffer:
i=self.buffer.find(b'\x7b')
if i<0:self.buffer=b'';break
self.buffer=self.buffer[i:]
if len(self.buffer)<24:break
d=decode(self.buffer[:24])
if d is None:self.buffer=self.buffer[1:]
else:out.append(d);self.buffer=self.buffer[24:]
return out
class Watchdog:
def __init__(self,send):self.send=send;self.last=None;self.now=None;self.armed=False
def validate(self,t):
if type(t) not in (int,float) or not math.isfinite(t) or t<0 or (self.now is not None and t<self.now):raise ValueError('time')
def command(self,now_ms,x=0,z=0):
self.validate(now_ms);b=encode(x,z);self.send(b);self.now=now_ms;self.last=now_ms;self.armed=True
def tick(self,now_ms):
self.validate(now_ms)
if self.armed and now_ms-self.last>=2000:self.send(encode());self.armed=True
self.now=now_ms

View File

@ -0,0 +1,33 @@
import struct,math,functools,operator
def xor(b):return functools.reduce(operator.xor,b,0)
def encode(x_mm_s=0,z_deg_s_1000=0):
if any(type(v) is not int or not -32768<=v<=32767 for v in (x_mm_s,z_deg_s_1000)):raise ValueError('range')
b=bytes([123,0])+struct.pack('>hhh',x_mm_s,0,z_deg_s_1000)+bytes(1)
return b+bytes([xor(b),125])
def decode(frame):
if len(frame)!=24 or frame[0]!=123 or frame[-1]!=125 or xor(frame[:22])!=frame[22]:return None
return dict(x_mm_s=int.from_bytes(frame[2:4],'big',signed=True),z_deg_s=int.from_bytes(frame[6:8],'big',signed=True)/1000.0,left_enc=int.from_bytes(frame[8:10],'big',signed=True),right_enc=int.from_bytes(frame[10:12],'big',signed=True),gas_ppm=int.from_bytes(frame[15:17],'big'),sensor_ok=frame[12]==252)
class StreamDecoder:
def __init__(self):self.buffer=b''
def feed(self,chunk):
self.buffer+=chunk;out=[]
while self.buffer:
i=self.buffer.find(b'\x7b')
if i<0:self.buffer=b'';break
self.buffer=self.buffer[i:]
if len(self.buffer)<24:break
d=decode(self.buffer[:24])
if d is None:self.buffer=self.buffer[1:]
else:out.append(d);self.buffer=self.buffer[24:]
return out
class Watchdog:
def __init__(self,send):self.send=send;self.last=None;self.now=None;self.armed=False
def validate(self,t):
if type(t) not in (int,float) or not math.isfinite(t) or t<0 or (self.now is not None and t<self.now):raise ValueError('time')
def command(self,now_ms,x=0,z=0):
self.validate(now_ms);b=encode(x,z);self.send(b);self.now=now_ms;self.last=now_ms;self.armed=True
def tick(self,now_ms):
self.validate(now_ms)
if self.armed and now_ms-self.last>=2000:self.send(encode());self.armed=False
self.now=now_ms

View File

@ -0,0 +1,16 @@
TARGETS=['gas_bridge.py']
import pathlib, shutil, subprocess, sys, tempfile
own=pathlib.Path(__file__).resolve().parent
try:
with tempfile.TemporaryDirectory(dir='/tmp',prefix='author1-grade-') as d:
tmp=pathlib.Path(d)
for name in TARGETS:
p=pathlib.Path(sys.argv[1])/name
if not p.is_file() or p.is_symlink():raise ValueError('missing target '+name)
shutil.copyfile(p,tmp/name)
shutil.copyfile(own/'tests.py',tmp/'_tests.py')
p=subprocess.run([sys.executable,'-I','-B',str(tmp/'_tests.py')],cwd=tmp,text=True,capture_output=True,timeout=25)
if p.returncode or p.stdout.strip()!='VERIFIED_ALL_CASES':raise ValueError((p.stdout+p.stderr)[-900:])
print('PASS')
except Exception as exc:
print('FAIL:',str(exc).replace(str(tmp),'<scratch>') if 'tmp' in globals() else str(exc));sys.exit(1)

View File

@ -0,0 +1,33 @@
import struct,math,functools,operator
def xor(b):return functools.reduce(operator.xor,b,0)
def encode(x_mm_s=0,z_deg_s_1000=0):
if any(type(v) is not int or not -32768<=v<=32767 for v in (x_mm_s,z_deg_s_1000)):raise ValueError('range')
b=bytes([123,0])+struct.pack('>hhh',x_mm_s,0,z_deg_s_1000)+bytes(1)
return b+bytes([xor(b),125])
def decode(frame):
if len(frame)!=24 or frame[0]!=123 or frame[-1]!=125 or xor(frame[:22])!=frame[22]:return None
return dict(x_mm_s=int.from_bytes(frame[2:4],'big',signed=True),z_deg_s=int.from_bytes(frame[6:8],'big',signed=True)/1000.0,left_enc=int.from_bytes(frame[8:10],'big',signed=True),right_enc=int.from_bytes(frame[10:12],'big',signed=True),gas_ppm=int.from_bytes(frame[20:22],'big'),sensor_ok=frame[12]==252)
class StreamDecoder:
def __init__(self):self.buffer=b''
def feed(self,chunk):
self.buffer+=chunk;out=[]
while self.buffer:
i=self.buffer.find(b'\x7b')
if i<0:self.buffer=b'';break
self.buffer=self.buffer[i:]
if len(self.buffer)<24:break
d=decode(self.buffer[:24])
if d is None:self.buffer=self.buffer[1:]
else:out.append(d);self.buffer=self.buffer[24:]
return out
class Watchdog:
def __init__(self,send):self.send=send;self.last=None;self.now=None;self.armed=False
def validate(self,t):
if type(t) not in (int,float) or not math.isfinite(t) or t<0 or (self.now is not None and t<self.now):raise ValueError('time')
def command(self,now_ms,x=0,z=0):
self.validate(now_ms);b=encode(x,z);self.send(b);self.now=now_ms;self.last=now_ms;self.armed=True
def tick(self,now_ms):
self.validate(now_ms)
if self.armed and now_ms-self.last>=2000:self.send(encode());self.armed=False
self.now=now_ms

View File

@ -0,0 +1,2 @@
"""Implement README.md."""
raise NotImplementedError

View File

@ -0,0 +1,3 @@
当前目录是待实现的空工程。请按 README.md 完成气体遥测流重同步与软件命令看门狗,并自行验证。
最终只提交并评分以下文件:`gas_bridge.py`。

View File

@ -0,0 +1,96 @@
# 任务规格:S500_a1_993e8b5b0064
本规格按[任务制作规范](../../docs/task-authoring.md)记录路线A导入审阅;不得提供给Agent。
## 1. 身份与来源
| 字段 | 填写 |
| --- | --- |
| task-id / 任务版本 | S500_a1_993e8b5b0064 / 1 |
| 作者 / 业务评审人 / 独立评审人 | 本批Codex(导入);业务评审未单独指定 / Codex review_1 |
| 审核状态 | ADMITTED |
| 原项目与冻结版本 | Bos-qiao task_factory_500_20260914;本批本地只读复制快照 `factory/20261014-A-bos-direct/bos-envs/S500_a1_993e8b5b0064/`;原task.json SHA-256 `4a918c56cd9c8deaa89ed112e7eef1bc821cafe434a14cc8c9b3a75fa1c6b68b` |
| 来源路线 | A 导入成熟任务;原创新建软件任务(bos.origin未声明;依据公开README的新建空工程声明,不标为历史现场) |
| 来源 ID | `S500_a1_993e8b5b0064`;source_qid=`q_35b94b12993e8b5b0064` |
| 来源与授权 | 原来源 `/mnt/data/ming01.wang/Bos-qiao/envs/S500_a1_993e8b5b0064`;依用户2026-10-12接入/试跑授权与本次作业A授权,只保留导入软件题材料与控制侧证据 |
| 隐私处理 | 无真实用户会话原文;不读/复制私有凭据;公开材料仅任务软件契约和部署路径 |
| 替身测试 | 软件:实现控制帧XOR、气体浓度真值字段、噪声分片流和一次性软件看门狗;空标准工程为新建需求。 用合成输入、受控回调/PTY或本机只读软件接口可复现;不依赖实体外设 |
| split 分组键 | `bos-qiao/task_factory_500_20260914/envs/S500_a1_993e8b5b0064`(同源变体使用同组,不把重复采样算新题) |
| 来源组 / 根因族 | Bos-qiao/q_35b94b12993e8b5b0064 / `gas-frame-resync-and-command-watchdog` |
| 与已有任务的关系 | 与既有frame-decoder同属帧软件,但新增明确24字节遥测字段、噪声重同步、分片状态及软件看门狗,解题路径不同;不是只换帧常量。 |
| 验证范围 | 下列验收矩阵中的确定软件行为;family=`motor-sensor-control`;最终仅冻结 `gas_bridge.py` |
| 未验证范围 | encode除7条与两端点外未随机遍历全部int16;send回调异常和非法command后的全部内部状态未独立断言。UART/STM32实体超时、实际气体测量均未验证。 |
实际问题说明:实现控制帧XOR、气体浓度真值字段、噪声分片流和一次性软件看门狗;空标准工程为新建需求。
## 2. 初态与环境
沿用 Bos 原题。
## 3. 公开题面与接口
沿用 Bos 原题。
## 4. 验收矩阵与合法解
| 需求 | 输入/场景 | 可观察结果与允许误差 | 检查方法 | 公开/受保护 | 超时 |
| --- | --- | --- | --- | --- | --- |
| 控制帧精确编码 | 公开7条命令;X/Z=(-32768,32767)/(32767,-32768) | 11字节帧与公开hex或struct+reduce(xor)完全相等;back100校验为18 | 独立字节构建/XOR | 规则公开;具体输入/期望受保护 | 评分总120秒;内部25/30秒 |
| 24字节遥测 | 种子901生成200帧;坏长度和翻转校验 | 读取2:4/6:8/8:10/10:12/20:22;Z除1000;sensor_ok由byte12;非法None | 独立struct.unpack与字节整数复算 | 规则公开;具体输入/期望受保护 | 评分总120秒;内部25/30秒 |
| 流重同步与隔离 | noise/garbage插入;随机1..47字节分片;坏帧后真帧;两个实例分片交错 | 合法帧按序全输出;空feed无输出;坏帧逐字节恢复;实例不共享缓冲 | 结果list与独立期望逐项相等 | 规则公开;具体输入/期望受保护 | 评分总120秒;内部25/30秒 |
| 看门狗边界与布防 | tick从未command;command9000;tick10999.9/11000/11001;command12000、tick14000 | 从未command不发;<2000不发;=2000发停止一次;后续不重复;新command重新布防 | send事件列表与公开encode帧 | 规则公开;具体输入/期望受保护 | 评分总120秒;内部25/30秒 |
| 非法输入/时间 | 时间倒退13999、-1/bool/NaN/inf;X=32768/-32769/bool/1.2 | ValueError;非法tick不新增发送 | 类型异常及事件列表 | 规则公开;具体输入/期望受保护 | 评分总120秒;内部25/30秒 |
- 合法替代解:valid_alternative对浓度改为byte20*256+byte21,并改变缓冲切片形式;任何等价解析、逐字节状态机和一次性看门狗可通过。
- 不允许的捷径:cheat超时每次tick重复stop由11000/11001抓住;cheat_stale_gas_copy读取15:17,由随机帧20:22真值抓住。
- 独立复算与冻结:check.py将声明目标复制到干净/tmp再运行/编译;正式期望来自公开约定的独立字节/数学/受控事件计算,不运行gold求答案,不复用Agent中间产物。判据附件为 `check.py`, `tests.py`。
- 判断范围与覆盖限制:encode除7条与两端点外未随机遍历全部int16;send回调异常和非法command后的全部内部状态未独立断言。UART/STM32实体超时、实际气体测量均未验证。
- grader输出:统一verify.py返回一行JSON;PASS/FAIL checks=1、退出0/1,环境/基础设施错误为UNVERIFIED/INFRA_ERROR、退出2;内部行为断言全部完成才PASS。
- 隔离复核:沿用固定runtime的Agent/评分分容器、candidate冻结只读与UID65534探针;Docker校准回执才证明正式探针及清理。verify.py内已有同UID原checker与候选共进程的shortcut,不能据此宣称已具备独立UID的恶意checker防篡改。
<!-- shortcut: 现有Bos评分包装器的check.py与候选同UID且共享checker进程,正式UID探针不等于完整恶意候选防篡改;升级条件为独立checker身份/进程边界合同准入。 -->
- 历史答案可见性:Agent仅得到prompt与template,spec/grader/gold不放入公开工作区;历史归档不由本题挂载。spec不进入task digest。
## 5. 四类样例与校准
沿用 Bos 原题。
## 6. 真实试采与证据
沿用 Bos 原题。
## 7. 评审与准入
- [x] 独立内容评审完成;来源、替身测试、公开行为、判据、合法解、负例及局限见第 1、4 节和下方采样前记录。
- [x] 最终 task/verifier/image 身份的正式 Docker 校准,每类样例各五次符合预期,整体 PASS、cleanup confirmed。
- [x] GLM-5.3 / Pi 0.73.1 首次真实试采正常退出,host/native/broker 对账和工具反馈因果闭合 VERIFIED,清理 confirmed。
- [x] 冻结候选在新无网评分容器重判一致,按 rollout.archive 显式证据合同重新核对后本地封存。
- [x] 重复来源、根因、软件边界和现有隔离局限已登记;闸门 4 尚未执行。
最终审核状态:**ADMITTED**。任务失败与轨迹完整分开记录,首次任务 verdict 为 **PASS**,不据单次结果作难度分池。人工干预:只读诊断与控制侧评审;未向运行 Agent 注入判据/答案或修改候选。
最终身份:family `motor-sensor-control`;task digest `06a952ce8fa8924570be7bfbf8d9a47aa5b68a852c44f85cf293e12dbdfb3a70`;verifier digest `2ce99a20ed07a4b16716568ee1efbd08f35e09b1300468ab443896171bde1cfe`;固定 runtime / actual image ID `sha256:aa5d00779fe03e9eadbf29c030711a4ecd1e5f894255868b5ca451a425299efc`;模型 `glm-5.3`、每次回复 max_tokens=131072、sampling=service-default,单题 Agent 防挂死期限 3600 秒。
最终正式校准:`/mnt/data/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-calibration/calibration-8d8e27c12d024751a37bbb680e62bec6`;本地原执行路径 `/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-calibration/calibration-8d8e27c12d024751a37bbb680e62bec6`。首次 controller:`/mnt/data/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct/pi-f18d7c7068e840d29b6074f1ac0f6501`;原路径 `/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct/pi-f18d7c7068e840d29b6074f1ac0f6501`。证据 `VERIFIED`,Agent exit=0 / termination=exit,cleanup confirmed,独立重判一致。完整轨迹与独立评分见本批冷数据 `factory/20261014-A-bos-direct/report.json`、`admission.jsonl`。
本地封存包:`/mnt/data/ming01.wang/embedded-rollout/artifacts/20261014-A-bos-direct/pi-f18d7c7068e840d29b6074f1ac0f6501.tar.gz`;SHA-256 `1a025987128718b48fbd9ac542ac606baac4bcaa758cb48ff4017b04d747bc3a`,2284082 字节。Salty 上传/下载摘要核对 **BLOCKED**:执行主机没有可用已认证 CLI 或连接;没有声明已上传。此阻塞不替代已完成的本地封存事实。冷迁按文件数与总字节核对,见批次 `migration-verification.json`;回执原字节和原绝对路径保持,当前入口不能直接对迁移后的路径重跑,需显式恢复原路径或升级路径合同。
### 采样前独立内容评审记录
下方保留采样前状态和当时的证据。若包含旧归族校准,只用于历史事实;当前准入由上述最终身份、校准和完整试采授权。
- [x] 2026-10-14,Codex review_1已逐项静态核对task/prompt/template、check及附件、gold/valid/全部cheat;未执行来源脚本、未改来源或平台。
- [x] 公开README与隐藏行为逐项对照;矩阵内受检正确行为均有公开依据,初态桩无gold泄露,目标自包含,软件替身范围明确。
- [x] root cause与来源组记录,已对照当前任务公开主题确认解题路径不同;合法替代解按行为而非内部补丁评审。
- [x] 固定镜像正式五次/样例校准整体PASS且cleanup confirmed(仅根据已见实际回执)。
- [ ] 闸门3真实试采的工具轨迹、独立评分、VERIFIED对账与清理:本评审尚未核对;由批次controller补齐,不能据本规格宣称ADMITTED。
- [ ] 闸门4≥5次同配置难度分池:本批作业A不执行。
样例合同:gold=PASS、template=FAIL、noop=FAIL、cheat=FAIL、cheat_stale_gas_copy=FAIL、valid_alternative=PASS;各类需正式重复5次,不能以静态审阅代替。
校准证据:
- 可信本地开发:`/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-dev/calibration-0ae379a6a98d440794693767c5ba43a1`;verdict=PASS,cleanup=confirmed,trials=30;task_digest=`06a952ce8fa8924570be7bfbf8d9a47aa5b68a852c44f85cf293e12dbdfb3a70`;实际image_id=`None`。
- 固定镜像Docker:`/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-calibration/calibration-8d8e27c12d024751a37bbb680e62bec6`;verdict=PASS,cleanup=confirmed,trials=30;task_digest=`06a952ce8fa8924570be7bfbf8d9a47aa5b68a852c44f85cf293e12dbdfb3a70`;实际image_id=`sha256:aa5d00779fe03e9eadbf29c030711a4ecd1e5f894255868b5ca451a425299efc`。
独立结论:APPROVED_FOR_CALIBRATION/PILOT(题目内容),人工生命周期状态为CALIBRATED,非ADMITTED。任务导入摘要 `06a952ce8fa8924570be7bfbf8d9a47aa5b68a852c44f85cf293e12dbdfb3a70`;正式镜像固定为 `ccr-29eug8s3-pub.cnc.bj.baidubce.com/algo/bench-runtime@sha256:aa5d00779fe03e9eadbf29c030711a4ecd1e5f894255868b5ca451a425299efc`。限制:encode除7条与两端点外未随机遍历全部int16;send回调异常和非法command后的全部内部状态未独立断言。UART/STM32实体超时、实际气体测量均未验证。 后续以最终正式校准与真实试采证据更新本节。

View File

@ -0,0 +1,36 @@
{
"id": "S500_a1_993e8b5b0064",
"version": "1",
"family": "motor-sensor-control",
"source": "bos-qiao/task_factory_500_20260914/envs/S500_a1_993e8b5b0064",
"template": "template",
"verifier": "verify.py",
"candidate_files": [
"gas_bridge.py"
],
"timeout_seconds": 120,
"fixtures": {
"gold": "PASS",
"template": "FAIL",
"noop": "FAIL",
"cheat": "FAIL",
"cheat_stale_gas_copy": "FAIL",
"valid_alternative": "PASS"
},
"grader_files": [
"check.py",
"tests.py"
],
"bos": {
"source_qid": "q_35b94b12993e8b5b0064",
"origin": null,
"image": "python@sha256:78387bc3881b8273120a12ebe6c1ab22b018ccc2c9adf565ae1ac9b536e184ea",
"budget_s": 600,
"fixture_scripts": {
"gold": "gold.sh",
"cheat": "cheat_repeated_stop.sh",
"cheat_stale_gas_copy": "cheat_stale_gas_copy.sh",
"valid_alternative": "valid_alternative.sh"
}
}
}

View File

@ -0,0 +1,11 @@
# 气体遥测流重同步与软件命令看门狗
这是依据原始新建代码需求创建的空标准工程,不是历史故障现场恢复。
实现 gas_bridge.py。encode(x_mm_s=0,z_deg_s_1000=0)->bytes 构建11字节控制帧:7B 00、int16大端X、两字节0、int16大端Z、00、前9字节XOR、7D;X/Z仅接受int16(不接受bool),非法抛ValueError。
decode(frame:bytes)->dict|None 验证长度24、头7B、尾7D、byte22为byte0..21的XOR;非法帧返回None。结果字段 x_mm_s 来自2:4有符号大端,z_deg_s来自6:8有符号大端除1000.0,left_enc/right_enc来自8:10/10:12有符号大端累计计数,gas_ppm来自20:22无符号大端,sensor_ok为byte12==0xFC。byte15/16的历史浓度副本不是真值来源,不能将其替代byte20/21。
StreamDecoder.feed(chunk:bytes)->list[dict]:保留分片,忽略噪声;遇坏24字节候选逐字节向后重同步,输出所有完整合法帧,保留未完成帧(头字节可能存在于载荷内,不因此丢弃完整合法帧)。不同实例无共享状态。
Watchdog(send) 有 command(now_ms,x=0,z=0)、tick(now_ms)。采用传入单调非负有限数(拒绝bool)毫秒,不访问真实时钟。command验证时间与参数后立即send(encode(...))并记录时间/重新布防;tick在距最近command >=2000ms时发送停止帧一次,后续tick不重复;从未command不发送;新的command重新布防。相同时间允许,倒退或非法时间抛ValueError且不改变状态/不发送。send为受控回调。本题只验证软件看门狗,不声称STM32实体超时已验证。
语言为Python标准库、UART名义配置115200 8N1(无设备操作)。按正式逐字节XOR公式冻结原7条速查;原后退100的校验e7错误,公开纠正为18:
forward500=7b0001f400000000008e7d,forward100=7b00006400000000001f7d,back500=7b00fe0c0000000000897d,back100=7b00ff9c0000000000187d,left10=7b00000000002710004c7d,right10=7b0000000000d8f000537d,stop=7b00000000000000007b7d。
全部7条均纳入测试,另测边界和流式重同步。
只依赖标准库;不访问硬件或网络。可在当前目录编写自己的自测;额外文件不影响判分。

View File

@ -0,0 +1,2 @@
"""Implement README.md."""
raise NotImplementedError

View File

@ -0,0 +1,38 @@
import sys, pathlib, random, math, json, struct
sys.path.insert(0,str(pathlib.Path(__file__).resolve().parent))
import gas_bridge as g
from functools import reduce
from operator import xor
for x,z,h in [(500,0,'7b0001f400000000008e7d'),(100,0,'7b00006400000000001f7d'),(-500,0,'7b00fe0c0000000000897d'),(-100,0,'7b00ff9c0000000000187d'),(0,10000,'7b00000000002710004c7d'),(0,-10000,'7b0000000000d8f000537d'),(0,0,'7b00000000000000007b7d')]:assert g.encode(x,z)==bytes.fromhex(h)
r=random.Random(901);frames=[];expected=[]
for _ in range(200):
b=bytearray(r.randbytes(22));b[0]=123;b[12]=r.choice([252,4]);b+=bytes([reduce(xor,b,0),125]);b=bytes(b)
exp=dict(x_mm_s=struct.unpack('>h',b[2:4])[0],z_deg_s=struct.unpack('>h',b[6:8])[0]/1000,left_enc=struct.unpack('>h',b[8:10])[0],right_enc=struct.unpack('>h',b[10:12])[0],gas_ppm=256*b[20]+b[21],sensor_ok=b[12]==252)
assert g.decode(b)==exp;frames.append(b);expected.append(exp)
assert g.decode(b[:-1]) is None and g.decode(b[:22]+bytes([b[22]^1])+b[23:]) is None
stream=b'noise'+b''.join(b'garbage'+f for f in frames)
d=g.StreamDecoder();got=[];i=0
while i<len(stream):
size=r.randint(1,47);got+=d.feed(stream[i:i+size]);i+=size
assert got==expected and d.feed(b'')==[]
d=g.StreamDecoder();bad=frames[0][:22]+bytes([frames[0][22]^1])+frames[0][23:]
assert d.feed(bad+frames[1])==[expected[1]]
a=g.StreamDecoder();b=g.StreamDecoder();a.feed(frames[0][:8]);assert b.feed(frames[1])==[expected[1]];assert a.feed(frames[0][8:])==[expected[0]]
ev=[];w=g.Watchdog(ev.append);w.tick(0);w.tick(9000);assert ev==[]
w.command(9000,100,0);w.tick(10999.9);assert ev==[g.encode(100,0)]
w.tick(11000);w.tick(11001);assert ev==[g.encode(100,0),g.encode()]
w.command(12000,-100,20);w.tick(14000);assert ev[-2:]==[g.encode(-100,20),g.encode()]
for bad in [13999,-1,True,float('nan'),float('inf')]:
old=list(ev)
try:w.tick(bad)
except ValueError:pass
else:raise AssertionError('bad time')
assert ev==old
for x,z in [(-32768,32767),(32767,-32768)]:
body=bytes([123,0])+struct.pack('>hhh',x,0,z)+bytes(1);assert g.encode(x,z)==body+bytes([reduce(xor,body,0),125])
for v in [32768,-32769,True,1.2]:
try:w.command(15000,v,0)
except ValueError:pass
else:raise AssertionError('bad command')
print('VERIFIED_ALL_CASES')

View File

@ -0,0 +1,33 @@
import struct,math,functools,operator
def xor(b):return functools.reduce(operator.xor,b,0)
def encode(x_mm_s=0,z_deg_s_1000=0):
if any(type(v) is not int or not -32768<=v<=32767 for v in (x_mm_s,z_deg_s_1000)):raise ValueError('range')
b=bytes([123,0])+struct.pack('>hhh',x_mm_s,0,z_deg_s_1000)+bytes(1)
return b+bytes([xor(b),125])
def decode(frame):
if len(frame)!=24 or frame[0]!=123 or frame[-1]!=125 or xor(frame[:22])!=frame[22]:return None
return dict(x_mm_s=int.from_bytes(frame[2:4],'big',signed=True),z_deg_s=int.from_bytes(frame[6:8],'big',signed=True)/1000.0,left_enc=int.from_bytes(frame[8:10],'big',signed=True),right_enc=int.from_bytes(frame[10:12],'big',signed=True),gas_ppm=frame[20]*256+frame[21],sensor_ok=frame[12]==252)
class StreamDecoder:
def __init__(self):self.buffer=b''
def feed(self,chunk):
self.buffer+=chunk;out=[]
while self.buffer:
i=self.buffer.find(b'\x7b')
if i<0:self.buffer=b'';break
self.buffer=self.buffer[i:]
if len(self.buffer)<24:break
d=decode(self.buffer[:24])
if d is None:self.buffer=bytes(self.buffer[1:])
else:out.append(d);self.buffer=self.buffer[24:]
return out
class Watchdog:
def __init__(self,send):self.send=send;self.last=None;self.now=None;self.armed=False
def validate(self,t):
if type(t) not in (int,float) or not math.isfinite(t) or t<0 or (self.now is not None and t<self.now):raise ValueError('time')
def command(self,now_ms,x=0,z=0):
self.validate(now_ms);b=encode(x,z);self.send(b);self.now=now_ms;self.last=now_ms;self.armed=True
def tick(self,now_ms):
self.validate(now_ms)
if self.armed and now_ms-self.last>=2000:self.send(encode());self.armed=False
self.now=now_ms

View File

@ -0,0 +1,93 @@
"""Run an imported Bos-qiao check.py unchanged, with the candidate user's privileges."""
import json
import os
from pathlib import Path
import shutil
import subprocess
import sys
import tempfile
GRADER = Path(__file__).resolve().parent
# Runs the original check unchanged; failing child processes also copy their stderr tail out for private triage.
CHECK = (
"import runpy, subprocess, sys\n"
"real = subprocess.run\n"
"def run(*args, **kwargs):\n"
" result = real(*args, **kwargs)\n"
" if result.returncode and result.stderr:\n"
" tail = result.stderr if isinstance(result.stderr, str) else result.stderr.decode('utf-8', 'replace')\n"
" sys.__stderr__.write(tail[-2000:] + '\\n')\n"
" return result\n"
"subprocess.run = run\n"
"sys.argv = sys.argv[1:]\n"
"runpy.run_path(sys.argv[0], run_name='__main__')\n")
class UnsupportedEnvironment(RuntimeError):
pass
def candidate_user():
os.setgroups([])
os.setgid(65534)
os.setuid(65534)
def isolation():
if os.environ.get("ROLLOUT_GRADE_ISOLATION") != "uid65534":
return None
if sys.platform != "linux" or os.geteuid() != 0:
raise UnsupportedEnvironment("isolated grader requires a Linux root supervisor")
probe = (
"from pathlib import Path\n"
"try: Path('/grader/verify.py').read_bytes()\n"
"except PermissionError: raise SystemExit(0)\n"
"raise SystemExit(1)\n"
)
try:
result = subprocess.run([sys.executable, "-I", "-c", probe],
capture_output=True, timeout=10, preexec_fn=candidate_user)
except subprocess.SubprocessError as error:
raise UnsupportedEnvironment("candidate identity could not be isolated") from error
if result.returncode != 0:
raise UnsupportedEnvironment("candidate can read the private grader or isolation probe failed")
return candidate_user
def verify(candidate):
child_user = isolation()
if shutil.which("gcc") is None:
raise UnsupportedEnvironment("C compiler unavailable")
with tempfile.TemporaryDirectory(prefix="bos-grade-") as temporary:
root = Path(temporary)
work, grader = root / "work", root / "grader"
shutil.copytree(candidate, work, symlinks=True)
shutil.copytree(GRADER, grader, ignore=lambda folder, names: ["verify.py"] if Path(folder) == GRADER else [])
if child_user:
os.chmod(root, 0o700) # Without CAP_FOWNER, root cannot chmod after giving files away.
for path in (root, *root.rglob("*")):
os.lchown(path, 65534, 65534)
# shortcut: the original check and the candidate share UID 65534, so a hostile candidate could
# tamper with its checker process; upgrade by running check.py and its tests under separate UIDs.
completed = subprocess.run([sys.executable, "-I", "-B", "-c", CHECK, str(grader / "check.py"), str(work)],
cwd=root, capture_output=True, text=True, timeout=100, preexec_fn=child_user)
# grade.stderr stays on the trusted side with the archive; it may quote hidden test lines.
sys.stderr.write(completed.stderr[-4000:])
lines = completed.stdout.splitlines()
if completed.returncode == 0 and lines[-1:] == ["PASS"]:
return {"status": "PASS", "checks": 1}
return {"status": "FAIL", "checks": 1, "reason": (lines[-1] if lines else "no checker output")[:300]}
if __name__ == "__main__":
try:
result = verify(Path(sys.argv[1]))
except UnsupportedEnvironment as error:
result = {"status": "UNVERIFIED", "checks": 0, "reason": str(error)}
except subprocess.TimeoutExpired:
result = {"status": "FAIL", "checks": 1, "reason": "timeout"}
except Exception as error:
result = {"status": "INFRA_ERROR", "checks": 0, "reason": str(error)}
print(json.dumps(result, sort_keys=True))
sys.exit({"PASS": 0, "FAIL": 1}.get(result["status"], 2))

View File

@ -0,0 +1,12 @@
import math
def run(steps,get_distance_cm,set_speed,stop,sleep_ms):
if type(steps) is not int or steps<0:raise ValueError('steps')
try:
for _ in range(steps):
d=get_distance_cm()
if type(d) not in (int,float) or not math.isfinite(d) or d<0:stop()
elif d<=20:set_speed(-2.0,2.0)
else:set_speed(2.0,2.0)
sleep_ms(50)
finally:pass

View File

@ -0,0 +1,12 @@
import math
def run(steps,get_distance_cm,set_speed,stop,sleep_ms):
if type(steps) is not int or steps<0:raise ValueError('steps')
try:
for _ in range(steps):
d=get_distance_cm()
if type(d) not in (int,float) or not math.isfinite(d) or d<0:stop()
elif d<20:set_speed(-2.0,2.0)
else:set_speed(2.0,2.0)
sleep_ms(50)
finally:stop()

View File

@ -0,0 +1,16 @@
TARGETS=['controller.py']
import pathlib, shutil, subprocess, sys, tempfile
own=pathlib.Path(__file__).resolve().parent
try:
with tempfile.TemporaryDirectory(dir='/tmp',prefix='author1-grade-') as d:
tmp=pathlib.Path(d)
for name in TARGETS:
p=pathlib.Path(sys.argv[1])/name
if not p.is_file() or p.is_symlink():raise ValueError('missing target '+name)
shutil.copyfile(p,tmp/name)
shutil.copyfile(own/'tests.py',tmp/'_tests.py')
p=subprocess.run([sys.executable,'-I','-B',str(tmp/'_tests.py')],cwd=tmp,text=True,capture_output=True,timeout=25)
if p.returncode or p.stdout.strip()!='VERIFIED_ALL_CASES':raise ValueError((p.stdout+p.stderr)[-900:])
print('PASS')
except Exception as exc:
print('FAIL:',str(exc).replace(str(tmp),'<scratch>') if 'tmp' in globals() else str(exc));sys.exit(1)

View File

@ -0,0 +1,12 @@
import math
def run(steps,get_distance_cm,set_speed,stop,sleep_ms):
if type(steps) is not int or steps<0:raise ValueError('steps')
try:
for _ in range(steps):
d=get_distance_cm()
if type(d) not in (int,float) or not math.isfinite(d) or d<0:stop()
elif d<=20:set_speed(-2.0,2.0)
else:set_speed(2.0,2.0)
sleep_ms(50)
finally:stop()

View File

@ -0,0 +1,2 @@
"""Implement README.md."""
raise NotImplementedError

View File

@ -0,0 +1,3 @@
当前目录是待实现的空工程。请按 README.md 完成障碍阈值控制与安全退出,并自行验证。
最终只提交并评分以下文件:`controller.py`。

View File

@ -0,0 +1,95 @@
# 任务规格:S500_a1_9c6eba37e895
本规格按[任务制作规范](../../docs/task-authoring.md)记录路线A导入审阅;不得提供给Agent。
## 1. 身份与来源
| 字段 | 填写 |
| --- | --- |
| task-id / 任务版本 | S500_a1_9c6eba37e895 / 1 |
| 作者 / 业务评审人 / 独立评审人 | 本批Codex(导入);业务评审未单独指定 / Codex review_1 |
| 审核状态 | ADMITTED |
| 原项目与冻结版本 | Bos-qiao task_factory_500_20260914;本批本地只读复制快照 `factory/20261014-A-bos-direct/bos-envs/S500_a1_9c6eba37e895/`;原task.json SHA-256 `f8418fa5044048c83e7766397031e7b2a420e316701de663cf6093ceca26df25` |
| 来源路线 | A 导入成熟任务;原创新建软件任务(bos.origin未声明;依据公开README的新建空工程声明,不标为历史现场) |
| 来源 ID | `S500_a1_9c6eba37e895`;source_qid=`q_0e7395d39c6eba37e895` |
| 来源与授权 | 原来源 `/mnt/data/ming01.wang/Bos-qiao/envs/S500_a1_9c6eba37e895`;依用户2026-10-12接入/试跑授权与本次作业A授权,只保留导入软件题材料与控制侧证据 |
| 隐私处理 | 无真实用户会话原文;不读/复制私有凭据;公开材料仅任务软件契约和部署路径 |
| 替身测试 | 软件:实现20cm闭阈值策略、50ms回调周期与try/finally安全退出;为新建需求。 用合成输入、受控回调/PTY或本机只读软件接口可复现;不依赖实体外设 |
| split 分组键 | `bos-qiao/task_factory_500_20260914/envs/S500_a1_9c6eba37e895`(同源变体使用同组,不把重复采样算新题) |
| 来源组 / 根因族 | Bos-qiao/q_0e7395d39c6eba37e895 / `obstacle-threshold-and-final-stop` |
| 与已有任务的关系 | 独立问题;在本批与既有题的公开需求/根因对照中未发现仅参数或问法替换的重复。 |
| 验证范围 | 下列验收矩阵中的确定软件行为;family=`motor-sensor-control`;最终仅冻结 `controller.py` |
| 未验证范围 | 只测有限受控回调序列与异常,不证明避障效果、实际50ms调度、电机或距离传感器。 |
实际问题说明:实现20cm闭阈值策略、50ms回调周期与try/finally安全退出;为新建需求。
## 2. 初态与环境
沿用 Bos 原题。
## 3. 公开题面与接口
沿用 Bos 原题。
## 4. 验收矩阵与合法解
| 需求 | 输入/场景 | 可观察结果与允许误差 | 检查方法 | 公开/受保护 | 超时 |
| --- | --- | --- | --- | --- | --- |
| 阈值/周期顺序 | 20/20.0001/0/19.999,以及种子5613生成800个0..40距离 | <=20速度(-2,2),>20为(2,2);read→控制→sleep(50),结束stop | 独立构建完整事件序列 | 规则公开;具体输入/期望受保护 | 评分总120秒;内部25/30秒 |
| 无效距离/零周期 | -1/None/bool/NaN/inf;steps=0 | 无效距离周期stop仍sleep,结束额外stop;零周期stop一次 | 事件列表完全比较 | 规则公开;具体输入/期望受保护 | 评分总120秒;内部25/30秒 |
| 非法steps | -1/bool/1.2/字符串3 | ValueError且无任何回调 | 事件列表与异常断言 | 规则公开;具体输入/期望受保护 | 评分总120秒;内部25/30秒 |
| 异常与优先级 | read/speed/sleep抛异常;周期stop失败;两次stop均失败 | 尝试退出stop一次;异常传播;最终stop异常覆盖周期stop异常 | 异常消息+stop次数 | 规则公开;具体输入/期望受保护 | 评分总120秒;内部25/30秒 |
- 合法替代解:valid_alternative用条件表达式生成左右速度;显式finally或正确的等价清理逻辑均可,不依赖内部结构。
- 不允许的捷径:cheat省略final stop由空周期及完整序列拒绝;cheat_strict_threshold用d<20由d=20拒绝。
- 独立复算与冻结:check.py将声明目标复制到干净/tmp再运行/编译;正式期望来自公开约定的独立字节/数学/受控事件计算,不运行gold求答案,不复用Agent中间产物。判据附件为 `check.py`, `tests.py`。
- 判断范围与覆盖限制:只测有限受控回调序列与异常,不证明避障效果、实际50ms调度、电机或距离传感器。
- grader输出:统一verify.py返回一行JSON;PASS/FAIL checks=1、退出0/1,环境/基础设施错误为UNVERIFIED/INFRA_ERROR、退出2;内部行为断言全部完成才PASS。
- 隔离复核:沿用固定runtime的Agent/评分分容器、candidate冻结只读与UID65534探针;Docker校准回执才证明正式探针及清理。verify.py内已有同UID原checker与候选共进程的shortcut,不能据此宣称已具备独立UID的恶意checker防篡改。
<!-- shortcut: 现有Bos评分包装器的check.py与候选同UID且共享checker进程,正式UID探针不等于完整恶意候选防篡改;升级条件为独立checker身份/进程边界合同准入。 -->
- 历史答案可见性:Agent仅得到prompt与template,spec/grader/gold不放入公开工作区;历史归档不由本题挂载。spec不进入task digest。
## 5. 四类样例与校准
沿用 Bos 原题。
## 6. 真实试采与证据
沿用 Bos 原题。
## 7. 评审与准入
- [x] 独立内容评审完成;来源、替身测试、公开行为、判据、合法解、负例及局限见第 1、4 节和下方采样前记录。
- [x] 最终 task/verifier/image 身份的正式 Docker 校准,每类样例各五次符合预期,整体 PASS、cleanup confirmed。
- [x] GLM-5.3 / Pi 0.73.1 首次真实试采正常退出,host/native/broker 对账和工具反馈因果闭合 VERIFIED,清理 confirmed。
- [x] 冻结候选在新无网评分容器重判一致,按 rollout.archive 显式证据合同重新核对后本地封存。
- [x] 重复来源、根因、软件边界和现有隔离局限已登记;闸门 4 尚未执行。
最终审核状态:**ADMITTED**。任务失败与轨迹完整分开记录,首次任务 verdict 为 **PASS**,不据单次结果作难度分池。人工干预:只读诊断与控制侧评审;未向运行 Agent 注入判据/答案或修改候选。
最终身份:family `motor-sensor-control`;task digest `e2bf274e16351595cfff51edb3941259a7d8ee25d0edd70a36d259b4e5f952bf`;verifier digest `2ce99a20ed07a4b16716568ee1efbd08f35e09b1300468ab443896171bde1cfe`;固定 runtime / actual image ID `sha256:aa5d00779fe03e9eadbf29c030711a4ecd1e5f894255868b5ca451a425299efc`;模型 `glm-5.3`、每次回复 max_tokens=131072、sampling=service-default,单题 Agent 防挂死期限 3600 秒。
最终正式校准:`/mnt/data/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-calibration/calibration-bf395089efa84b7aadee9f8f391768cd`;本地原执行路径 `/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-calibration/calibration-bf395089efa84b7aadee9f8f391768cd`。首次 controller:`/mnt/data/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct/pi-bd945f582a144fd281c1f92fe4002ad8`;原路径 `/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct/pi-bd945f582a144fd281c1f92fe4002ad8`。证据 `VERIFIED`,Agent exit=0 / termination=exit,cleanup confirmed,独立重判一致。完整轨迹与独立评分见本批冷数据 `factory/20261014-A-bos-direct/report.json`、`admission.jsonl`。
本地封存包:`/mnt/data/ming01.wang/embedded-rollout/artifacts/20261014-A-bos-direct/pi-bd945f582a144fd281c1f92fe4002ad8.tar.gz`;SHA-256 `1ca9292b23e838120e608bbf25a39b8639cfaa5d036cae2e81bb65973e051367`,351854 字节。Salty 上传/下载摘要核对 **BLOCKED**:执行主机没有可用已认证 CLI 或连接;没有声明已上传。此阻塞不替代已完成的本地封存事实。冷迁按文件数与总字节核对,见批次 `migration-verification.json`;回执原字节和原绝对路径保持,当前入口不能直接对迁移后的路径重跑,需显式恢复原路径或升级路径合同。
### 采样前独立内容评审记录
下方保留采样前状态和当时的证据。若包含旧归族校准,只用于历史事实;当前准入由上述最终身份、校准和完整试采授权。
- [x] 2026-10-14,Codex review_1已逐项静态核对task/prompt/template、check及附件、gold/valid/全部cheat;未执行来源脚本、未改来源或平台。
- [x] 公开README与隐藏行为逐项对照;矩阵内受检正确行为均有公开依据,初态桩无gold泄露,目标自包含,软件替身范围明确。
- [x] root cause与来源组记录,已对照当前任务公开主题确认解题路径不同;合法替代解按行为而非内部补丁评审。
- [x] 固定镜像正式五次/样例校准整体PASS且cleanup confirmed(仅根据已见实际回执)。
- [ ] 闸门3真实试采的工具轨迹、独立评分、VERIFIED对账与清理:本评审尚未核对;由批次controller补齐,不能据本规格宣称ADMITTED。
- [ ] 闸门4≥5次同配置难度分池:本批作业A不执行。
样例合同:gold=PASS、template=FAIL、noop=FAIL、cheat=FAIL、cheat_strict_threshold=FAIL、valid_alternative=PASS;各类需正式重复5次,不能以静态审阅代替。
校准证据:
- 可信本地开发:`/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-dev/calibration-eccc44994aba4a61a51ef8472f4d677b`;verdict=PASS,cleanup=confirmed,trials=30;task_digest=`e2bf274e16351595cfff51edb3941259a7d8ee25d0edd70a36d259b4e5f952bf`;实际image_id=`None`。
- 固定镜像Docker:`/home/ming01.wang/embedded-rollout/runs/20261014-A-bos-direct-calibration/calibration-bf395089efa84b7aadee9f8f391768cd`;verdict=PASS,cleanup=confirmed,trials=30;task_digest=`e2bf274e16351595cfff51edb3941259a7d8ee25d0edd70a36d259b4e5f952bf`;实际image_id=`sha256:aa5d00779fe03e9eadbf29c030711a4ecd1e5f894255868b5ca451a425299efc`。
独立结论:APPROVED_FOR_CALIBRATION/PILOT(题目内容),人工生命周期状态为CALIBRATED,非ADMITTED。任务导入摘要 `e2bf274e16351595cfff51edb3941259a7d8ee25d0edd70a36d259b4e5f952bf`;正式镜像固定为 `ccr-29eug8s3-pub.cnc.bj.baidubce.com/algo/bench-runtime@sha256:aa5d00779fe03e9eadbf29c030711a4ecd1e5f894255868b5ca451a425299efc`。限制:只测有限受控回调序列与异常,不证明避障效果、实际50ms调度、电机或距离传感器。 后续以最终正式校准与真实试采证据更新本节。

View File

@ -0,0 +1,36 @@
{
"id": "S500_a1_9c6eba37e895",
"version": "1",
"family": "motor-sensor-control",
"source": "bos-qiao/task_factory_500_20260914/envs/S500_a1_9c6eba37e895",
"template": "template",
"verifier": "verify.py",
"candidate_files": [
"controller.py"
],
"timeout_seconds": 120,
"fixtures": {
"gold": "PASS",
"template": "FAIL",
"noop": "FAIL",
"cheat": "FAIL",
"cheat_strict_threshold": "FAIL",
"valid_alternative": "PASS"
},
"grader_files": [
"check.py",
"tests.py"
],
"bos": {
"source_qid": "q_0e7395d39c6eba37e895",
"origin": null,
"image": "python@sha256:78387bc3881b8273120a12ebe6c1ab22b018ccc2c9adf565ae1ac9b536e184ea",
"budget_s": 600,
"fixture_scripts": {
"gold": "gold.sh",
"cheat": "cheat_no_final_stop.sh",
"cheat_strict_threshold": "cheat_strict_threshold.sh",
"valid_alternative": "valid_alternative.sh"
}
}
}

View File

@ -0,0 +1,8 @@
# 障碍阈值控制与安全退出
这是依据原始新建代码需求创建的空标准工程,不是历史故障现场恢复。
实现 controller.py 的 run(steps, get_distance_cm, set_speed, stop, sleep_ms)。steps 是非负整数;非法 steps(含 bool)抛 ValueError 且无回调。
每周期按顺序读一次距离:有限非负实数且 >20 cm 时 set_speed(2.0,2.0),0..20(含20)时原地左转 set_speed(-2.0,2.0),无效距离(负数、非数值、bool、NaN、无穷)调用 stop()。速度单位 cm/s。每周期最后 sleep_ms(50),不使用真实 sleep。循环正常完成后额外 stop() 一次,steps=0 也必须 stop。任何读数/控制/等待回调异常必须传播,但退出前仍 stop 一次(无效值导致的周期 stop 不代替退出 stop)。此处公开固定转弯策略与50ms周期;传感器和电机通过传入回调抽象。
只依赖标准库;不访问硬件或网络。可在当前目录编写自己的自测;额外文件不影响判分。
异常优先级遵循 try/finally:若周期 stop 抛异常,仍尝试退出 stop;若退出 stop 也抛异常,传播退出 stop 的异常(覆盖先前异常)。steps 非法时不执行退出 stop。

View File

@ -0,0 +1,2 @@
"""Implement README.md."""
raise NotImplementedError

View File

@ -0,0 +1,51 @@
import sys, pathlib, random, math, json, struct
sys.path.insert(0,str(pathlib.Path(__file__).resolve().parent))
import controller as c
r=random.Random(5613)
for ds in [[],[20,20.0001,0,19.999,-1,None,True,float('nan'),float('inf')],[r.uniform(0,40) for _ in range(800)]]:
events=[];it=iter(ds)
def read():events.append(('read',));return next(it)
def speed(l,r):events.append(('speed',l,r))
def stop():events.append(('stop',))
c.run(len(ds),read,speed,stop,lambda ms:events.append(('sleep',ms)))
exp=[]
for d in ds:
exp.append(('read',))
exp.append(('stop',) if type(d) not in (int,float) or not math.isfinite(d) or d<0 else ('speed',2.0 if d>20 else -2.0,2.0))
exp.append(('sleep',50))
exp.append(('stop',));assert events==exp, 'callback sequence'
for bad in [-1,True,1.2,'3']:
ev=[]
try:c.run(bad,lambda:ev.append(1),lambda *x:ev.append(2),lambda:ev.append(3),lambda x:ev.append(4))
except ValueError:pass
else:raise AssertionError('invalid steps')
assert not ev
for failure in ['read','speed','sleep']:
ev=[]
def f(kind):
def cb(*a):
if kind==failure:raise RuntimeError('sensor')
return cb
try:c.run(3,f('read') if failure=='read' else lambda:30,f('speed'),lambda:ev.append('stop'),f('sleep'))
except RuntimeError:pass
else:raise AssertionError('exception swallowed')
assert ev==['stop']
# Invalid-distance stop itself raises; final stop is still required.
ev=[]
def stopping():
ev.append('stop')
if len(ev)==1:raise RuntimeError('period-stop')
try:c.run(1,lambda:None,lambda *x:None,stopping,lambda _:None)
except RuntimeError as ex:assert str(ex)=='period-stop'
else:raise AssertionError('stop exception lost')
assert ev==['stop','stop']
ev=[]
def always_broken():
ev.append('stop');raise RuntimeError('stop-'+str(len(ev)))
try:c.run(1,lambda:None,lambda *x:None,always_broken,lambda _:None)
except RuntimeError as ex:assert str(ex)=='stop-2'
else:raise AssertionError('final exception priority')
assert ev==['stop','stop']
print('VERIFIED_ALL_CASES')

View File

@ -0,0 +1,11 @@
import math
def run(steps,get_distance_cm,set_speed,stop,sleep_ms):
if type(steps) is not int or steps<0:raise ValueError('steps')
try:
for _ in range(steps):
d=get_distance_cm()
if type(d) not in (int,float) or not math.isfinite(d) or d<0:stop()
else:set_speed(-2.0 if d<=20 else 2.0,2.0)
sleep_ms(50)
finally:stop()

View File

@ -0,0 +1,93 @@
"""Run an imported Bos-qiao check.py unchanged, with the candidate user's privileges."""
import json
import os
from pathlib import Path
import shutil
import subprocess
import sys
import tempfile
GRADER = Path(__file__).resolve().parent
# Runs the original check unchanged; failing child processes also copy their stderr tail out for private triage.
CHECK = (
"import runpy, subprocess, sys\n"
"real = subprocess.run\n"
"def run(*args, **kwargs):\n"
" result = real(*args, **kwargs)\n"
" if result.returncode and result.stderr:\n"
" tail = result.stderr if isinstance(result.stderr, str) else result.stderr.decode('utf-8', 'replace')\n"
" sys.__stderr__.write(tail[-2000:] + '\\n')\n"
" return result\n"
"subprocess.run = run\n"
"sys.argv = sys.argv[1:]\n"
"runpy.run_path(sys.argv[0], run_name='__main__')\n")
class UnsupportedEnvironment(RuntimeError):
pass
def candidate_user():
os.setgroups([])
os.setgid(65534)
os.setuid(65534)
def isolation():
if os.environ.get("ROLLOUT_GRADE_ISOLATION") != "uid65534":
return None
if sys.platform != "linux" or os.geteuid() != 0:
raise UnsupportedEnvironment("isolated grader requires a Linux root supervisor")
probe = (
"from pathlib import Path\n"
"try: Path('/grader/verify.py').read_bytes()\n"
"except PermissionError: raise SystemExit(0)\n"
"raise SystemExit(1)\n"
)
try:
result = subprocess.run([sys.executable, "-I", "-c", probe],
capture_output=True, timeout=10, preexec_fn=candidate_user)
except subprocess.SubprocessError as error:
raise UnsupportedEnvironment("candidate identity could not be isolated") from error
if result.returncode != 0:
raise UnsupportedEnvironment("candidate can read the private grader or isolation probe failed")
return candidate_user
def verify(candidate):
child_user = isolation()
if shutil.which("gcc") is None:
raise UnsupportedEnvironment("C compiler unavailable")
with tempfile.TemporaryDirectory(prefix="bos-grade-") as temporary:
root = Path(temporary)
work, grader = root / "work", root / "grader"
shutil.copytree(candidate, work, symlinks=True)
shutil.copytree(GRADER, grader, ignore=lambda folder, names: ["verify.py"] if Path(folder) == GRADER else [])
if child_user:
os.chmod(root, 0o700) # Without CAP_FOWNER, root cannot chmod after giving files away.
for path in (root, *root.rglob("*")):
os.lchown(path, 65534, 65534)
# shortcut: the original check and the candidate share UID 65534, so a hostile candidate could
# tamper with its checker process; upgrade by running check.py and its tests under separate UIDs.
completed = subprocess.run([sys.executable, "-I", "-B", "-c", CHECK, str(grader / "check.py"), str(work)],
cwd=root, capture_output=True, text=True, timeout=100, preexec_fn=child_user)
# grade.stderr stays on the trusted side with the archive; it may quote hidden test lines.
sys.stderr.write(completed.stderr[-4000:])
lines = completed.stdout.splitlines()
if completed.returncode == 0 and lines[-1:] == ["PASS"]:
return {"status": "PASS", "checks": 1}
return {"status": "FAIL", "checks": 1, "reason": (lines[-1] if lines else "no checker output")[:300]}
if __name__ == "__main__":
try:
result = verify(Path(sys.argv[1]))
except UnsupportedEnvironment as error:
result = {"status": "UNVERIFIED", "checks": 0, "reason": str(error)}
except subprocess.TimeoutExpired:
result = {"status": "FAIL", "checks": 1, "reason": "timeout"}
except Exception as error:
result = {"status": "INFRA_ERROR", "checks": 0, "reason": str(error)}
print(json.dumps(result, sort_keys=True))
sys.exit({"PASS": 0, "FAIL": 1}.get(result["status"], 2))

Some files were not shown because too many files have changed in this diff Show More