LGD Engine · 判决档案 · 由引擎自身生成

五个判定器,
37 个案件,
67 条规则逐条对条款。

这一页不是宣传材料。每张卡上的判决,都是本仓库里的案件文件送进 GateEngine 跑出来的原始结果——含命中的规则编号、等级、理由原文。你可以在本地用一条命令重跑,得到与页面逐字相同的结果。

生成时间(UTC):2026-11-19 06:21
重跑 python benchmarks/build_demo_page.py · 校验 python benchmarks/build_demo_page.py --check(不一致则 CI 失败)
判定器:510(k) 变更分类 / 上市后监测 / PCCP 范围 / EU MDR 重大变更 / NMPA 变更注册
5
judges shipped
覆盖 FDA / EU / 中国三辖区
—
case files
全部随包发布,可直接跑
—
verdicts match
判决 = 案件预期,逐条一致
94
contract assertions
每次推送由 CI 执行
55
rules with a clause
67 条规则全部点名依据
0
network / model calls
确定性 YES · 单次判决预算 ≤500 µs

案件档案

点击任一案件展开完整理由链

它和"看着像就行"的写法差多少

12 条对抗探针 · 同一批输入

对照物不是商用产品,是一个"最可能被写出来的偷懒实现":JSON 里出现 banned 关键词就 reject、出现 requires_submission 就 hold、其余放行。下面是 benchmarks/adversarial_baseline.py 此刻的真实输出(逐字)。

实现与预期一致的探针说明

朴素法错在两个相反方向:该拦的放行(空证据、未知变更类型、字符串 "yes" 被当成等价声明)、该放的拦住(PCCP 绿色案件因为 JSON 里出现了键名 affects_intended_use(值为 false)而被误判 hold)。这正是"键名匹配"与"语义判定"的差别。

条款矩阵

67 条规则 · 每条点名它编码的条款

判定器给结论不点条款,就和"模型的意见"无法区分。下表把每条规则落到依据上,并分列两种来源:LAW 是我们读过原文的法条/法规/MDCG·CAMD 通知/FDA 指南;DOCTRINE 是本家规则(L2 证据门、单辖区声明),明示为家法,不许冒充法条。家法占比由测试硬性封顶,防止它悄悄长成多数。

判定器规则条款依据类型

复核方式:python benchmarks/clause_matrix.py 双向检查——有规则无依据会失败,有依据但对应规则不存在也会失败(死引用正是引用表烂掉的方式)。可读版见 docs/methods/Clause_Matrix.md。

它故意不覆盖什么

9 条 · 带状态与理由

没写进规则的条款,如果不列出来,就和"没想到"无法区分。下面每一条都是明知存在、故意没编码的,附类型与理由;数据在 docs/methods/coverage_gaps.json,由测试校验字段、状态枚举、理由长度与三辖区覆盖。

辖区条款状态理由
NMPA47号令 第七十九条 — 其他变化,30 日内备案out_of_scopeThe table judges which route applies, not whether the clock was met. No days_since_change field exists, so a late filing cannot be detected — adding one would turn a routing judge into a compliance clock, and the caller's timestamps would h…
NMPA47号令 第七十九条 第一款 — 实质性变化且可能影响安全有效deferredThe substantive judgement is carried by the caller's safety_effectiveness_assessed flag; the table accepts the conclusion, not the argument behind it.
NMPA各省对一类器械备案的实施细则out_of_scopeProvincial implementing rules differ and are outside the central-law basis we verified; encoding them would mean picking provinces.
EUMDR Art.120(3) as amended by Regulation (EU) 2023/607 — 2027-12-31 / 2028-12-31 endpointsout_of_scopeThe table judges significance, not transitional validity. Validity depends on certificate dates the case JSON does not carry.
EUMDCG 2020-3 rev.1 — software change decision supportdeferredChart C is guidance on how to reach the significance conclusion; the rules encode the criteria (design/intended-purpose, cumulative), not the flowchart.
FDAFD&C §515C(a)(3)/(b)(3) — labeling, notification if the device does not function as intended, performance requirementsoptional_by_lawThe statute says the Secretary *may* require these; a plan without a notification clause is not thereby unlawful.
FDAFDA PCCP guidance (4 Dec 2024) — continuous learning / automatic update Q-Sub discussionoptional_by_lawA recommendation, not an obligation; making it a rule would manufacture a requirement the guidance does not impose.
FDA21 CFR 892.2055 之外的 PCCP 器械分类特殊控制deferredEach classification regulation carries its own special controls; only 892.2055 has been read end to end so far.
FDA510(k) 变更分类之外的判据(如 510(k) 摘要与 SE 论证细则)out_of_scopeThis table answers 'is a new submission needed', not 'is the submission adequate'.

复核方式:python -m unittest tests.test_contract.TestCoverageGaps -v。注意本条的限制:没列出的缺口与没被承认的缺口,在页面上无从区分。

我们怎么把自己打穿的

18 个自造反例 · 每个都是一份随仓库发布的案件

这 18 个案件不是外部质疑,是我们自己造的反例,而且每一个都成功过:它们曾经拿到 go。共同形状只有一句——规则等在某个中间产物上,于是把这个产物的"缺席"读成了"许可"。等的东西可以是一个布尔标记、一个词表取值、一个下游分类,或者干脆是一个没人填的字段。

域案件文件判决现在拦它的规则
510(k) 变更分类FDA-510K-RED-01demo1/cases/510k_red_team_label_expansion.jsonholdR-MAJ-09-label-expansion-new-claim
510(k) 变更分类DEMO1-RED-01demo1/cases/case01_red_pma_process.jsonrejectR-CRIT-01-class1-recall
510(k) 变更分类DEMO1-RED-02demo1/cases/case02_red_material.jsonrejectR-CRIT-01-class1-recall
上市后监测处置PM-RED-01demo1/cases_postmarket/pm01_class1_software.jsonrejectPM-CRIT-01-class1-recall
上市后监测处置FDA-PM-RED-01demo1/cases_postmarket/pm04_red_team_death_without_recall.jsonholdPM-MAJ-04-reportable-event-without-recall
PCCP 变更范围PCCP-RED-03demo1/cases_pccp/pccp03_red_intended_use.jsonrejectP-CRIT-02-intended-use-change
PCCP 变更范围PCCP-RED-04demo1/cases_pccp/pccp04_red_criteria_not_met.jsonrejectP-CRIT-03-acceptance-criteria-not-met
PCCP 变更范围PCCP-RED-06demo1/cases_pccp/pccp06_red_se_to_predicate_broken.jsonrejectP-CRIT-04-se-to-predicate-broken
PCCP 变更范围PCCP-RED-07demo1/cases_pccp/pccp07_red_modified_version_as_predicate.jsonrejectP-CRIT-05-modified-version-used-as-predicate
PCCP 变更范围FDA-PCCP-RED-01demo1/cases_pccp/pccp08_red_team_absent_special_controls.jsonholdP-MAJ-03-validation-not-documented
EU MDR 重大变更MDR-RED-04demo1/cases_mdr/mdr04_red_open_fsca.jsonrejectM-CRIT-01-change-during-open-fsca
EU MDR 重大变更MDR-RED-05demo1/cases_mdr/mdr05_red_declared_non_significant.jsonrejectM-CRIT-02-significant-declared-non-significant
EU MDR 重大变更MDR-RED-01demo1/cases_mdr/mdr08_red_team_acceptance_scope_expansion.jsonholdM-MAJ-01-intended-purpose
EU MDR 重大变更MDR-RED-02demo1/cases_mdr/mdr09_red_team_acceptance_scope_classification.jsonholdM-MAJ-03-safety-or-performance
NMPA 变更注册NMPA-RED-04demo1/cases_nmpa/nmpa04_red_licence_as_filing.jsonrejectN-CRIT-02-licence-item-filed-as-registration
NMPA 变更注册NMPA-RED-05demo1/cases_nmpa/nmpa05_red_level1_recall.jsonrejectN-CRIT-03-level1-recall-open
NMPA 变更注册NMPA-RED-01demo1/cases_nmpa/nmpa08_red_team_class1_registration_route.jsonholdN-MAJ-15-class1-no-registration-route
NMPA 变更注册NMPA-RED-02demo1/cases_nmpa/nmpa09_red_team_class1_first_registration.jsonholdN-MAJ-15-class1-no-registration-route

复核方式:每个文件都带 expected_verdict,--all 在修复被回退时直接变红;逐条复跑与"过去会怎样"的说明在 docs/methods/RedTeam_Probes.md。三种能继续用的攻击形状也写在那份文档里:删字段而不是设为否定值、用词表认识的值而不是词表拒绝的值、喂规则所键的中间产物而不是它要证明的事实。

七个"我们不一样",各自有多少证据

observed 2 · inferred 2 · unverified 3

对外说"类目都怎样"很容易,说过就算的代价是外人无法验证。下表每行都写清:我们这侧用什么命令settle、对方那侧我们到底查没查。只有 observed 且给出链接的,才能当作对他人的事实陈述;其余是推断或未查。

维度我们的主张(可复核)证据强度
does the verdict call a modelVerdicts are pure functions: no network, no model, byte-identical across reruns.
benchmarks/verdict_cost.py
inferred
—
does each rule name its legal basisAll 60 rules carry a LAW or DOCTRINE label, with DOCTRINE capped at 12.
benchmarks/clause_matrix.py
observed
source
is the evaluation report generated or writtenThe report is emitted by the evaluator and a check fails when its numbers disagree with the repo.
benchmarks/eval_method_note.py
unverified
—
are coverage gaps publishedNine intentionally unencoded clauses are published with a status and a reason.
docs/methods/coverage_gaps.json
observed
source
can an outsider re-run the claimsTwenty checks run from a fresh clone with no credentials.
benchmarks/eval_method_note.py
unverified
—
are the claims enforced in public CI58 contract assertions run on every push, including the three laws.
tests/test_contract.py
unverified
—
is the scope limited to enumerated categoriesAnything outside the tables returns hold; coverage is the bottleneck, stated as such.
lgd_engine/judges/base.py
inferred
—

复核方式:python benchmarks/edge_claims.py;可读版 docs/methods/Competitive_Edges.md。若 observed 占比过半,测试直接失败——那意味着我们在宣称一场没做过的调研。