菜单
Use when a production AI feature needs a version-specific model card covering purpose, data, evaluation, limitations, bias, intended use, prohibited use and operational pins; use AI architecture for system design and evaluation artefacts as evidence inputs.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/peterbamuhigire/srs-skills --skill 12-ai-model-card命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | 12-ai-model-card |
| description | Use when a production AI feature needs a version-specific model card covering purpose, data, evaluation, limitations, bias, intended use, prohibited use and operational pins; use AI architecture for system design and evaluation artefacts as evidence inputs. |
| metadata | {"portable":true,"compatible_with":["claude-code","codex"]} |
| Artefact | Source or provider | Required? | Missing behaviour |
|---|---|---|---|
| Deployed configuration pins and provider evidence | AI architecture, prompt registry and provider sources | Required | Stop if the deployable cannot be identified reproducibly. |
| Evaluation, red-team, data and incident evidence | Current project evidence | Required | Mark each missing check not assessed and block certification claims. |
| Artefact | Consumer | Observable acceptance condition |
|---|---|---|
| Versioned AI Model Card | Users, buyers, risk reviewers, auditors and operations | Every claim cites deployment-specific evidence; limitations and prohibited uses are explicit; missing assessments are not reported as passes. |
| Evidence | Consumer | Acceptance condition |
|---|---|---|
| Source and decision trace | Reviewer and downstream owner | Each material statement cites an approved input, named decision or qualified open issue. |
| Completed verification record | Release or phase gate owner | Every applicable check records pass/fail; unavailable checks remain not assessed. |
Read-only is the default for analysis, review, evaluation and planning. Read and search access to authorised project artefacts are required. Editing is limited to an explicitly requested project deliverable. Execution may run document, syntax or validation checks. Network access is used only for facts that require current verification. Do not publish, spend, change production, approve policy, or claim certification without explicit authority.
If any required capability is unavailable, return the narrowest useful qualified Versioned AI Model Card draft plus a gap register showing the missing item, affected sections, risk and owner. Never convert an unassessed check into a pass.
| Choice | Action | Failure or risk avoided |
|---|---|---|
| Evidence describes deployed feature and version | Include with source/date | Card remains auditable |
| Only provider-level evidence exists | Qualify applicability and require system evaluation | Generic claims do not certify the product |
Produces the per-feature model card that buyers, auditors, and Responsible-AI reviewers will read. Anchored in Mitchell et al. (2019) and the EU AI Act Annex IV technical documentation requirements.
A model card refers to a deployed combination: { feature, base model + version, prompt registry tag, retrieval index version, post-processing rules }. Pin all four.
Describe what the feature does, who uses it, what decisions it informs, and the buyer outcome. Reference the AI FR ID.
For hosted base models, cite the provider's training-data disclosure with a date. For our retrieval data, summarise: sources (cite the AI Data Spec), volumes, languages, freshness, exclusions. For fine-tunes, list training set, size, labelling provenance, holdout set.
Attach the most recent eval report for the feature: golden-set pass rate, factuality, abstention precision, citation rate, hallucination rate, judge-LLM score, latency P95, cost per call. Each metric carries the date and the eval set ID.
Latest red-team result by category: prompt injection, jailbreak, data exfiltration, cross-tenant leak, PII surfacing, hallucination probe, bias surfacing. Pass / fail per category; any open finding with severity and remediation date.
List known limitations as concrete failure modes:
List bias risks identified (gender, race, age, disability, geography, language) and the mitigation in place (eval set composition, abstain rule, content filter, human-in-the-loop).
Intended use: business contexts the feature supports. Out-of-scope: contexts where it must not be used (e.g. medical decisions, legal advice, hiring, lending). Mirror the safety rules in the PRD spec.
Model + version, prompt registry tag, retrieval index version, gateway config tag, content-filter version, eval suite tag.
Map each Annex IV element to the section of this model card that satisfies it:
| Annex IV item | Where covered |
|---|---|
| General description of the AI system | Purpose |
| Elements of the AI system and process of its development | Training/Grounding + Operational Pins |
| Data sets used | Training/Grounding + AI Data Spec |
| Validation and testing procedures | Evaluation + Red-team |
| Risk management measures | Limitations + Bias + Operational Pins |
| Human oversight | Intended Use + AI FR clause |
| Changes through the lifecycle | Versioning section |
Model_Card_<feature>_v<version>.md. Date-stamped. Owner named.
