在本机跑一个模型Run a local model
先交付一份能复查的部署计划，再在自己的授权环境中验证；容量数字和软件平台都要有范围。Prepare a checkable deployment plan before validating in your own authorized environment; scope both memory estimates and platform claims.
展开输入、步骤与验收Open inputs, steps and acceptance
输入Input
准备操作系统/架构、任务、实际可用内存、候选仓与修订、模型许可和网络限制。练习输入使用虚构的 10 GiB 预算，不代表一台真实机器。Prepare OS/architecture, task, available memory, candidate repo/revision, model licence and network constraints. The exercise assumes a fictional 10 GiB budget, not a real machine.
交付Deliverable
交付版本与许可记录、容量假设、待确认项；实际运行结果单独记录，未运行就保留未运行。Deliver version/licence records, capacity assumptions and unknowns. Record execution evidence separately; keep unexecuted steps unexecuted.
按用途找资源Find resources by role
资源是按用途举例，先展开条目的条件；没有必选组合。Resources illustrate roles. Inspect their conditions first; there is no mandatory combination.
写清任务与边界Define task and limits
把输入语言、输出长度、联网限制和是否允许保存日志写进工作表。先用虚构输入，不把生产数据当首次练习。Record language, output length, network limits and logging permission. Start with fictional input rather than production data.
确认模型与运行器是一组Match model and runtime
保留模型仓、固定修订、文件名、许可 URL；核对运行器的该模型/格式/平台版本说明。GGUF 或接口名称相同不能替代兼容说明。Retain repository, pinned revision, filename and licence URL. Check runtime support for that model/format/platform version; an extension or API name is not compatibility evidence.
分开算权重、缓存与预留Separate weights, cache and reserve
用容量工具填写驻留权重、缓存结构、总缓存长度和并发。先确认公式适用架构，文件体积不直接当驻留内存。Budget resident weights, cache structure, total cached tokens and concurrency. Confirm architecture scope first; file size is not resident memory.
把验证与计划分开保存Separate validation from planning
计划通过后，在自己的环境以短输入检查加载、输出和实际占用，再按需求增加上下文/并发。记录失败与回退，不把估算值填成观测值。After planning, check loading, output and observed allocation with short input in your own environment, then adjust context/concurrency. Keep failures and fallback steps; do not report estimates as observations.
怎样判断交付合格How to check the deliverable
- 模型/运行器/文件修订可定位，许可与用途分别记录。Model/runtime/file revisions are identifiable, with licence and intended use recorded separately.
- GiB、缓存长度、并发、架构范围与预留没有混成一个固定百分比。GiB, cached length, concurrency, architecture scope and reserve are not replaced by one fixed percentage.
- 参考计划的 9 GiB 估算不写成已成功运行。The reference plan’s 9 GiB estimate is not reported as successful execution.
失败后查哪里When it fails
填写上下文与并发预算Budget context and concurrency
输入样例Sample input · 参考结果Designed reference · 空白工作表Blank worksheet
修一个代码问题Fix a code issue
用一个故意写错的小函数练习复现、最小修改、验证与回退；参考修复属于本站代码，不是编码产品实测。Use a deliberately wrong small function to practice reproduction, minimal edits, checks and rollback. The reference is site-authored code, not a coding-product test.
展开输入、步骤与验收Open inputs, steps and acceptance
输入Input
下载 add-bug.mjs 与输入 JSON。约定 add 接受两个有限数字并求和；本练习不扩展成金额计算库或数据清洗功能。Download add-bug.mjs and the input JSON. The contract adds two finite numbers; this exercise does not expand into a money library or data-cleaning feature.
交付Deliverable
交付复现步骤、最小补丁、前后检查结果和回退方法。保留 bug 输入，不只给一段修改后代码。Deliver reproduction steps, a minimal patch, before/after checks and rollback instructions. Preserve the failing input, not just the repaired code.
按用途找资源Find resources by role
资源是按用途举例，先展开条目的条件；没有必选组合。Resources illustrate roles. Inspect their conditions first; there is no mandatory combination.
先得到可复现的差异Reproduce a difference
确认文件与运行环境，把预期和实际结果分开写。例如 2+3 预期为 5，错误版本返回 -1。这里只描述随包代码，不归因于某个 AI 工具。Record file/environment and separate expected from observed output. For 2+3, expect 5; the bundled wrong version returns -1. This concerns the supplied code, not an AI product.
限定修改范围Bound the edit
只修函数中的错误运算，不重构目录、引入依赖或改动不相关文件。使用外部助手前先核对会上传哪些代码与日志。Fix the wrong operation without reorganizing directories, adding dependencies or touching unrelated files. Check code/log uploads before using an external assistant.
检查补丁与结果Check patch and results
读 diff，再用正数、负数和零的输入检查。测试要能区分错误与正确实现；不要把两边都通过的例子当修复证据。Inspect the diff, then check positive, negative and zero inputs. Cases must distinguish wrong and repaired implementations; a case both pass is insufficient evidence.
留下可撤回的交付Keep a reversible deliverable
保存原文件或补丁，写出如何恢复与哪些行为未覆盖。下载的参考文件供对照，不应替代对自己改动的检查。Keep the original or a patch, with rollback instructions and untested behavior. Downloaded reference code is not a substitute for checking your own edits.
怎样判断交付合格How to check the deliverable
- 至少一个复现输入在错误版本失败、修复版本通过。At least one reproduction case distinguishes the wrong and repaired versions.
- 补丁只更改约定范围，未添加远程调用或依赖。The patch stays within scope and adds no remote calls or dependencies.
- 检查结果、运行范围和回退方法均可复查。Check results, execution scope and rollback remain inspectable.
失败后查哪里When it fails
继续完整修复流程Continue the full repair workflow
输入样例Sample input · 参考结果Designed reference · 空白工作表Blank worksheet · 错误代码Wrong code · 参考修复Reference repair
核对资料并写简报Check sources and write a brief
用三条虚构公告练习保留日期、对象、预览状态与未知项。查找、整理和核实各自交付，不用更多链接替代证据。Use three fictional announcements to retain dates, eligible users, preview status and unknowns. Searching, organizing and checking have separate outputs; more links do not replace support.
展开输入、步骤与验收Open inputs, steps and acceptance
输入Input
输入 JSON 包含同一虚构产品的三条公告。只以包内原文做第一次练习，不能把虚构产品当真实服务搜索后补写事实。The input JSON contains three announcements for one fictional product. Use only those texts initially; do not search for the fictional product and invent real-world facts.
交付Deliverable
交付一页简报和结论—出处表，标出日期、对象、支持位置、冲突与未确认项。Deliver a one-page brief and claim-to-source record with dates, eligible users, locations, conflicts and unknowns.
按用途找资源Find resources by role
资源是按用途举例，先展开条目的条件；没有必选组合。Resources illustrate roles. Inspect their conditions first; there is no mandatory combination.
先写可回答的问题Write an answerable question
本例只问预览对象、是否延期、公众何时可用。将问题和未知项记录下来，避免从公告标题直接推全面上线。Ask who receives the preview, whether it was delayed, and when public availability begins. Record unknowns; a headline does not establish general release.
给资料分工Assign a purpose to each source/tool
搜索工具找原文线索；资料阅读工具整理已收集来源；学术文献工具只在问题涉及论文时使用。三种角色不要求每次全部加入。Search tools locate originals; source-reading tools organize collected material; scholarly tools are relevant only to paper-related questions. All three are not mandatory for every task.
逐条链接到原文位置Locate support for every claim
本例用 A/B/C 和段落号定位。真实材料保留原始 URL、访问日与必要引文；预览、延期和范围必须一起保留。Use A/B/C and paragraph IDs in the exercise. For real sources, retain original URLs, access dates and necessary passages; preserve preview status, delay and scope together.
把未确认留在简报里Keep unknowns in the brief
关掉摘要回原文检查每个关键句。没有公开上线日期就写未给出，不把两份相同来源的答案当独立交叉证据。Return from summaries to originals for each key sentence. If no public date is given, keep it unknown. Two answers based on one source are not independent evidence.
怎样判断交付合格How to check the deliverable
- 每条结论均对应 A/B/C 的具体位置，没出现虚构的公众上线日或价格变化。Every claim locates A/B/C evidence without invented public dates or price changes.
- 预览范围与延期的时间关系清楚；新闻标题没有变成能力承诺。Preview scope and the timing of the delay are explicit; headlines do not become capability promises.
- 上传真实文档前已核对账号、数据处理与导出要求。Account, data-processing and export requirements are checked before uploading real documents.
失败后查哪里When it fails
继续研究简报的完整流程Continue the research-brief workflow
输入样例Sample input · 参考结果Designed reference · 空白工作表Blank worksheet
制作一分钟视频Make a one-minute video
从任务书、分镜、字幕和素材记录走到交付检查。示例是一份 60 秒计划，不冒充已经渲染的视频。Move from brief, storyboard, captions and asset records to delivery checks. The example is a 60-second plan, not a rendered video.
展开输入、步骤与验收Open inputs, steps and acceptance
输入Input
主题为“让结论能回到来源”的教学短片。输入给出八段时长和自制 SVG 卡片；没有第三方音乐、人物肖像或声音授权。The teaching film explains tracing a claim to its source. Inputs include eight durations and a site-authored SVG card; no third-party music, likeness or voice permissions are supplied.
交付Deliverable
交付任务书、分镜表、素材许可表、字幕稿、渲染参数和验收记录。只有实际导出并回放后，才能把视频状态改为完成。Deliver brief, storyboard, asset-rights ledger, caption script, render settings and checks. Mark a video complete only after actual export and playback.
按用途找资源Find resources by role
资源是按用途举例，先展开条目的条件；没有必选组合。Resources illustrate roles. Inspect their conditions first; there is no mandatory combination.
先确定事实和受众Set facts and audience
一句话说明观众需要带走什么。把事实句与演示句分开；未核实的数据不写进旁白。State what viewers should learn. Separate factual claims from illustration; do not put unverified numbers in narration.
按镜头分配时间与素材Allocate shots and assets
将八段时长相加，计划应为 60 秒；每段写画面、文字、声音与来源。现有 SVG 是原始教学素材，不能据此补上其他音乐的许可。Sum eight shot durations to 60 seconds. Record visual, text, sound and source per shot. The supplied SVG is original teaching material, not a licence for unrelated music.
检查生成与导出条件Check generation and export conditions
选工具前核对账号、素材输入、收费、保存格式和水印。目录目前只确认 Runway 账号内 Assets 可批量下载 ZIP，不保证具体模型分辨率或方案额度。Check accounts, asset input, billing, output format and watermarks before choosing tools. The directory confirms Runway Assets ZIP batch downloads only, not model resolution or plan quotas.
导出后分开回放检查Inspect the exported file
检查实际时长、画面/字幕同步、音量、来源与许可、播放器和目标平台要求。用实际文件记录结果；计划与预览不等于交付完成。Check actual duration, visual/caption sync, audio, evidence/rights, player and destination requirements. Record the actual file; a plan or preview is not completed delivery.
怎样判断交付合格How to check the deliverable
- 八段计划时长合计 60 秒，实际视频时长仍待导出后确认。Eight planned durations sum to 60 seconds; actual video duration awaits export.
- 每项素材有来源与许可，未知授权素材不进入发布包。Each asset has provenance and rights; assets with unknown permissions stay out of publication.
- 字幕、音量、容器/编码、平台要求与导出限制分别检查。Check captions, audio, container/codec, destination requirements and export restrictions separately.
失败后查哪里When it fails
继续一分钟视频的完整流程Continue the one-minute-video workflow
输入样例Sample input · 参考结果Designed reference · 空白工作表Blank worksheet · 原始图卡与署名Original card and attribution
AI 上线前安全评审AI pre-launch security review
把评审范围、数据流、权限边界与回退条件写成可复查的记录；随包日志和检查表是虚构教学材料，不是一次真实渗透测试的结论。Record scope, data flow, permission boundaries and rollback so they can be re-checked. The bundled log lines and checklist are fictional teaching material, not the result of a real penetration test.
展开输入、步骤与验收Open inputs, steps and acceptance
输入Input
一个虚构检索问答助手的系统说明、三条含两式虚构密钥的日志行，以及账号、工具权限、留存三类边界。随附工时与费用只是虚构估算，不指向任何真实系统或事件响应价格。A system description of a fictional retrieval assistant, three log lines carrying two fictional key formats, and three boundaries: accounts, tool permissions and retention. The hours and cost figures are fictional estimates and describe no real system or incident price.
交付Deliverable
交付范围表、数据流与边界记录、检查项台账（依据、结果、复现方式）、未决风险与回退方案。只有实际执行并留下证据的检查才能标为通过。Deliver a scope table, data-flow and boundary record, a check ledger (basis, result, reproduction) and open risks with rollback. Mark a check passed only when it was actually run and evidenced.
按用途找资源Find resources by role
资源是按用途举例，先展开条目的条件；没有必选组合。Resources illustrate roles. Inspect their conditions first; there is no mandatory combination.
先固定评审范围与资产Fix scope and assets
写明被测版本、涉及的数据类别、谁能看到生产提示词与日志、哪些能力会走到外部。练习范围要标注为教学范围，不冒充对真实系统的审计结论。State the version under review, the data classes involved, who can see production prompts and logs, and which capabilities reach outside. Label the exercise scope as teaching scope; it is not an audit of a real system.
画数据流并标注边界Map the data flow and its boundaries
按用户输入、应用、模型提供方、追踪与存储四段，分别写是否离开本机、是否留存、留存多久。厂商隐私表述记作文档声明并附日期，不记作已验证行为。For input, application, model provider and tracing/storage, record whether data leaves the machine, whether it is retained and for how long. Keep vendor privacy statements as dated documentation claims, not verified behaviour.
用可复查的检查项走一轮Run one round of re-checkable checks
覆盖提示注入、越权工具调用、密钥进入日志、依赖已知漏洞与导出删除路径。每项写依据（官方页面、代码或实测）、结果和复现步骤；没跑就留空，不填推测。Cover prompt injection, over-permitted tool calls, secrets reaching logs, known dependency vulnerabilities, and export/deletion paths. Record the basis (official page, code or an actual run), the result and the reproduction; leave unrun items blank instead of guessing.
交付风险表与上线条件Deliver the risk register and release conditions
每条风险给缓解、责任人、验证方式与回退，并把未覆盖和不适用分开写。是否可上线按范围与证据判断，不折算总分或等级。Give each risk a mitigation, owner, verification method and rollback, keeping not covered distinct from not applicable. Release readiness follows scope and evidence, not an overall score or grade.
怎样判断交付合格How to check the deliverable
- 每条结论都能回到具体位置（文件与行、配置项或带日期的官方页面），没有把看起来稳妥当证据。Every conclusion returns to a specific location (file and line, configuration key or dated official page); it looks safe is not recorded as evidence.
- 随包日志里两处虚构密钥（正文中的 sk- 值与 URL 查询串里的 token 值）在脱敏结果中都不可读，未覆盖的形式明确标注未覆盖。Both fictional secrets in the bundled logs — the sk- value in the message and the token value in a URL query — are unreadable after masking, and uncovered formats stay labelled uncovered.
- 依赖公告、模型许可与数据留存各有来源与核验日期；厂商宣传语与实测结果没有混写。Dependency advisories, model licences and data retention each carry their own source and check date; marketing statements are not mixed with observed results.
失败后查哪里When it fails
继续 agent 护栏清单Continue with the agent-guardrails checklist
输入样例Sample input · 参考结果Designed reference · 空白工作表Blank worksheet · 漏脱敏样例Under-masked code · 参考脱敏Reference masking
搭一条评测流水线Build an evaluation pipeline
从冻结数据集、运行记录、指标口径到回归门槛，先写清分母与版本；随包数字是虚构估算，不是任何模型或产品的实测得分。Freeze the dataset, record runs, define metrics and set a regression gate, with the denominator and versions explicit first. The bundled numbers are fictional estimates, not measured scores for any model or product.
展开输入、步骤与验收Open inputs, steps and acceptance
输入Input
一个虚构评测集的十条样本（其中两条未判定）、指标定义草案和一份标注为虚构的成本估算。练习不含真实用户数据，也不声称跑过任何模型或服务商。The input has ten samples from a fictional dataset (two left unjudged), a draft metric definition and a cost estimate labelled fictional. There is no real user data, and nothing here claims that a model or provider was run.
交付Deliverable
交付版本固定的数据集、指标口径（分子、分母、未判定处理）、运行记录、回归门槛与失败样本清单；未执行的运行保持未执行。Deliver a version-pinned dataset, metric definitions (numerator, denominator and unjudged handling), run records, a regression gate and the failing cases. Runs that were not executed stay unexecuted.
按用途找资源Find resources by role
资源是按用途举例，先展开条目的条件；没有必选组合。Resources illustrate roles. Inspect their conditions first; there is no mandatory combination.
先冻结数据集与口径Freeze the dataset and definitions
写明样本来源、去重规则、每条的正确判定标准和什么算不可判定，并给数据集一个版本号。合成样本要标注，不从线上日志静默取样。State sample provenance, deduplication, the correct-answer rule per case and what counts as unjudgeable, then give the dataset a version. Label synthetic cases; do not silently harvest production logs.
给每次运行留下可比的记录Keep run records comparable
每次运行记数据集版本、提示词与代码修订、模型标识与参数、判定所用标准。缺字段就写缺，不用事后说明把两次运行抹平成一样。Record the dataset version, prompt and code revision, model identifier and parameters and the judging standard for each run. Note missing fields instead of backfilling notes that make two runs look identical.
算指标时守住分母Guard the denominator when scoring
先决定未判定、报错与超时样本是否进入分母，再算准确率等指标，并把计数和百分比一起给出。跳过比例偏高时先查原因，而不是先调门槛。Decide first whether unjudged, errored and timed-out rows enter the denominator, then compute accuracy and report counts together with percentages. If many rows are skipped, investigate the cause before moving the threshold.
设回归门槛并保留失败样本Set a regression gate and keep failing cases
门槛写成相对上一基线的差值和允许波动区间，失败样本单独归档供复查。成本与时长用虚构估算并标明，不写成厂商承诺或实测吞吐。State the gate as a delta from the previous baseline with an allowed band, and archive failing cases for review. Mark cost and duration as fictional estimates rather than vendor promises or measured throughput.
怎样判断交付合格How to check the deliverable
- 每个指标都有定义、分子与分母规则和数据来源；未判定样本单独计数，既不算通过也不算失败。Every metric has a definition, numerator and denominator rule and data source; unjudged rows are counted separately, never as a pass or a fail.
- 两次运行可以对比：数据集版本、模型标识与判定标准都在记录里，改动可追。Two runs are comparable: dataset version, model identifier and judging standard are in the record, so changes are traceable.
- 随包评分函数在含未判定样本的输入上给出与参考解一致的数字，全部未判定时返回无法计算。The bundled scoring function matches the reference on inputs that contain unjudged rows, and returns not-computable when nothing was judged.
失败后查哪里When it fails
继续 agent 评测流程Continue with the agent-evaluation workflow
输入样例Sample input · 参考结果Designed reference · 空白工作表Blank worksheet · 错误计分样例Wrong scoring · 参考计分Reference scoring
