目录 ／ 文档-技术白皮书 ／ 44-EFT.WP.Data.ModelCards v1.0
I. 章节目的与范围
。禁用中文，覆盖结构/类型/正则/依赖/引用锚点/计量校核/冻结切分/泄漏防控/公平性与鲁棒性最小检查；产物直接用于发布前阻断检查与门户自动校验。键名统一 snake_case；跨卷引用采用“卷名 vX.Y:锚点”；数学表达用反引号并加括号，Lint 规则集与规范性 JSON Schema提供模型卡的II. 规范性工件（发布必备）
artifacts:
- path: "schema/model_card.schema.json" # 规范性 JSON Schema
- path: "schema/lint_rules.yaml" # 规范性 Lint 规则（可扩展）
- path: "schema/examples/minimal.yaml" # 最小可用示例
- path: "schema/examples/full.yaml" # 全字段示例
以上工件须在 export_manifest.artifacts[] 中登记并附 sha256；引用锚点与本卷口径一致。III. 规范性 JSON Schema（核心摘录）
references[] 正则强制“卷名 vX.Y:锚点”；evaluation.protocol.splits 固定为 "frozen"；resources 数量单位由计量章统一校核。IV. Lint 规则（规范性）
version: "v1.0"
rules:
# 结构与版本
- id: STRUCT.REQUIRED
when: "$"
assert: "has_keys(model_id,title,version,task,io_schema,architecture,training_data,preprocess,optimization,hyperparams,evaluation,calibration,robustness,fairness,safety,deployment,resources,export_manifest)"
level: error
- id: VERSION.SEMVER
when: "$.version"
assert: "matches('^v\\d+\\.\\d+(\\.\\d+)?$')"
level: error
# 冻结切分与引用口径
- id: EVAL.SPLITS_FROZEN
when: "$.evaluation.protocol.splits"
assert: "value == 'frozen'"
level: error
- id: REFERENCES.FORMAT
when: "$.export_manifest.references[*]"
assert: "matches('^[^:]+ v\\d+\\.\\d+:[A-Z].+$')"
level: error
# 计量与单位
- id: METROLOGY.SI_AND_CHECKDIM
when: "$.resources"
assert: "exists($.metrology) ? ($.metrology.units=='SI' and $.metrology.check_dim==true) : true"
level: error
see: ["EFT.WP.Core.Metrology v1.0:check_dim"]
# I/O 合规
- id: IO.SOFTMAX_RANGE
when: "$.io_schema.outputs[?(@.semantics=='softmax')]"
assert: "all(_.range == '[0,1]' for _ in items)"
level: error
- id: IO.SHAPE_NONEMPTY
when: "$.io_schema.inputs[*].shape"
assert: "matches('^\\(')"
level: error
# 训练数据绑定与泄漏防控
- id: TD.REFS_REQUIRED
when: "$.training_data"
assert: "len(refs) > 0 and splits_ref"
level: error
- id: TD.LEAKAGE_GUARDS
when: "$.training_data.leakage_guards"
assert: "contains_any(['per-object','per-timewindow','per-scene'])"
level: warn
# 公平性与鲁棒性最小检查
- id: FAIRNESS.THRESHOLD
when: "$.fairness.threshold"
assert: "value <= 0.10"
level: warn
- id: ROBUSTNESS.DROP_MAX
when: "$.robustness.thresholds.drop_rel_max"
assert: "value <= 0.20"
level: warn
# 文本与记号护栏
- id: MATH.NO_CHINESE
when: "$"
assert: "no_chinese_in_math()"
level: warn
- id: SYMBOLS.CONFLICT
when: "$"
assert: "not_mixed(['T_fil','T_trans']) and not_mixed(['n','n_eff'])"
level: error
。阻断项以上规则与 Schema 联合执行；STRUCT.REQUIRED/EVAL.SPLITS_FROZEN/REFERENCES.FORMAT/METROLOGY.SI_AND_CHECKDIM/SYMBOLS.CONFLICT 为V. 失败样例与诊断（节选）
fail_examples:
- case: "invalid reference"
input: {export_manifest:{references:["Core.DataSpec:EXPORT"]}}
expect: {rule:"REFERENCES.FORMAT", level:"error",
fix:"Use 'EFT.WP.Core.DataSpec v1.0:EXPORT'"}
- case: "splits not frozen"
input: {evaluation:{protocol:{splits:"custom"}}}
expect: {rule:"EVAL.SPLITS_FROZEN", level:"error",
fix:"Set evaluation.protocol.splits to 'frozen'"}
- case: "missing metrology check"
input: {resources:{M_param:25.6, FLOPs:4.1e9, T_inf:3.8}}
expect: {rule:"METROLOGY.SI_AND_CHECKDIM", level:"error",
fix:"Add metrology {units:'SI', check_dim:true}"}
。修复建议Lint 报告需包含 rule、path、message 与VI. 机器可读最小示例（通过校验）
model_id: "eift.vision.cls.resnet50"
title: "ResNet50 Classifier"
version: "v1.0"
task: "classification"
io_schema:
inputs: [{name:"image", shape:"(H,W,3)", dtype:"uint8", range:"[0,255]"}]
outputs: [{name:"probs", shape:"(K,)", dtype:"float32", range:"[0,1]", semantics:"softmax"}]
batching: {mode:"dynamic", max_batch:128}
architecture: {version:"v1.0", backbone:"resnet50", topology:[{name:"stem", type:"conv"}]}
training_data: {refs:["EFT.WP.Data.DatasetCards v1.0:Ch.11"], splits_ref:"eift.obs.demo@v1.0"}
preprocess: {pipeline_id:"img-prep-v1", steps:[], parameter_lock:true}
optimization: {objective:{name:"cross_entropy"}}
hyperparams: {batch_size:256, epochs:200}
evaluation:
protocol: {splits:"frozen", seeds:[0,1,2,3,4], repeats:5}
metrics: {classification:["f1_macro","roc_auc","ece","brier"]}
calibration: {method:"temperature", params:{t:1.7}}
robustness: {thresholds:{drop_rel_max:0.10}}
fairness: {axes:["class","region"], gap_metric:"abs_diff", threshold:0.05}
safety: {allowed_use:["academic"], prohibited_use:["surveillance"]}
deployment: {forms:["rest"], devices:["A100"], latency_targets_ms:{p50:5}}
resources: {M_param:25.6, FLOPs:4.1e9, T_inf:3.8}
metrology: {units:"SI", check_dim:true}
export_manifest:
version: "v1.0"
artifacts: [{path:"model_card.yaml", sha256:"..."}]
references: ["EFT.WP.Core.DataSpec v1.0:EXPORT","EFT.WP.Core.Metrology v1.0:check_dim"]
VII. 与导出清单的耦合（规范性）
export_manifest:
artifacts:
- {path:"schema/model_card.schema.json", sha256:"..."}
- {path:"schema/lint_rules.yaml", sha256:"..."}
- {path:"schema/examples/minimal.yaml", sha256:"..."}
references:
- "EFT.WP.Core.DataSpec v1.0:EXPORT"
- "EFT.WP.Core.Metrology v1.0:check_dim"
必须列出并可校验；引用携带“卷名 vX.Y:锚点”。阻断件Schema 与 Lint 作为VIII. 验证接口（实现绑定 Ixx-?，返回统一结构）
def validate_model_card(card: dict) -> dict: ...
def lint_model_card(card: dict, rules: dict) -> dict: ...
def check_units(card: dict) -> dict: ... # uses Core.Metrology v1.0:check_dim
def verify_references(card: dict) -> dict: ...# regex + anchor reachability
返回形如 {"ok": bool, "errors":[...], "warnings":[...], "metrics":{...}}，用于门户/CI。IX. 本章合规自检
- model_card.schema.json 与 lint_rules.yaml 已生成、入库并在 export_manifest 登记 sha256。
- Schema 强制 evaluation.protocol.splits="frozen" 与 export_manifest.references[] 的锚点正则；Lint 阻断短码/无版本/无锚点引用。
- metrology.units="SI"、check_dim=true 生效；resources、I/O 约束与单位口径一致；softmax 输出区间校核通过。
- 训练数据引用与 splits_ref 完整；泄漏防控与最小公平/鲁棒规则启用；符号冲突与数学禁中文规则通过。
- 最小示例能在本章 Schema 与 Lint 下一次通过；CI/门户验证接口已集成并返回统一结构。
版权与许可（CC BY 4.0）
版权声明：除另有说明外，《能量丝理论》（含文本、图表、插图、符号与公式）的著作权由作者（“屠广林”先生）享有。
许可方式：本作品采用 Creative Commons 署名 4.0 国际许可协议（CC BY 4.0）进行许可；在注明作者与来源的前提下，允许为商业或非商业目的进行复制、转载、节选、改编与再分发。
署名格式（建议）：作者：“屠广林”；作品：《能量丝理论》；来源：energyfilament.org；许可证：CC BY 4.0。
首次发布： 2025-11-11｜当前版本：v5.1
协议链接：https://creativecommons.org/licenses/by/4.0/