LGD Change Gate(医械 AI 变更判定)
Decide whether a change to a medical-device AI or ML function needs formal change control, and return a pass, hold or reject verdict with auditable reasons. Medical-device AI/ML change-control judgment per the LGD three laws (有籍 Recorded / 有证 Evidenced / 有门禁 Gated). Trigger when the user asks whether a device-software or algorithm change requires change control, asks for a risk verdict (pass/hold/reject), asks for change-level classification (minor/major/critical), or asks for an LGD-theory judgment with auditable reasons. Applies to regulated environments (FDA 510(k)/PMA, EU MDR, ISO 13485) and other AI/ML medical-device change evaluations.
LGD Change-Gate Judgment (lgd-change-gate)
When to use (triggers)
- The user describes a change to a medical-device AI/ML function (algorithm update, data change, model retrain, threshold adjustment, deployment change) and asks whether change control is needed.
- The user asks for a change verdict: pass / hold / reject.
- The user asks for change-level classification: minor / major / critical.
- The user asks for an LGD-theory judgment (three laws: Recorded / Evidenced / Gated) with auditable reasons.
Do NOT trigger: ordinary software changes outside medical devices, non-AI document edits, or questions without a change substance.
Core flow
Execute each step to evidence, never on impression.
Step 1 — Record the change (有籍 Recorded)
Collect the change object's "registry" facts:
- Change object: device/system/software component name and identifier (model, version, registration number).
- Change content: what changed (input, model, weights, code, configuration, data, deployment).
- Change initiator and time.
- Whether a baseline record exists (premarket submission, DHF, existing change-request record).
If the baseline is missing or the object cannot be identified → hold, reason: 无籍不可判定 (unregistered cannot be judged).
Step 2 — Verify evidence (有证 Evidenced)
Collect and verify evidence for the change claim:
- Necessity evidence: why the change is needed (defect, clinical need, regulatory requirement).
- Impact-analysis evidence: scope of impact on safety, effectiveness, and intended use.
- Verification evidence: test plan, data, results; coverage of old-vs-new behavior comparison.
- Evidence provenance and freshness: traceable, unexpired, authoritative.
If evidence is insufficient or unverifiable → hold, reason: 无证不裁决 (unevidenced no ruling).
Step 3 — Pass the gate (有门禁 Gated)
Pass gates by change level:
- minor (low risk, e.g. comments, docs, log adjustments): registration + verification record; pass condition = complete verification record.
- major (medium risk, e.g. algorithm parameter adjustment, performance tuning): impact analysis + verification + regression; pass condition = all three laws pass and verification passes.
- critical (high risk, e.g. core algorithm change, intended-use change, safety/effectiveness impact): full change-control process + clinical/regulatory assessment + higher-level approval; pass condition = full evidence-chain closure with approval record.
Gate rules:
| Condition | Verdict |
|---|---|
| All three laws pass (Recorded, Evidenced, Gated with closed evidence chain) | pass |
| Any law's evidence missing or pending | hold |
| Safety/effectiveness risk with clearly insufficient evidence, or intent to evade regulation | reject |
Output format
Structured output:
verdict: pass | hold | reject
level: minor | major | critical
reasons:
- recorded: ...
- evidenced: ...
- gated: ...
citations:
- evidence/record source...
Delivery standards
- Every verdict must include reasons (one per law) and citations (traceable sources).
- When evidence is insufficient, mark "pending" explicitly; never substitute assumptions for evidence.
- When citing regulatory clauses, quote the primary source and note its origin.
- Final verdict is exactly pass/hold/reject — no fuzzy middle state.
Local engine (optional deterministic check)
The bundled Python engine (engine/change_gate.py + engine/lgd_engine/) implements the same three-law gate deterministically. To run it:
python engine/change_gate.py --case <case.json>
python engine/change_gate.py --domain 510k --case <case.json>
A case JSON example:
{
"change_type": "material_nonconforming",
"device_risk_class": "II",
"safety_impact": "none",
"predicate_unchanged": true
}
The engine returns verdict (go/hold/reject) with per_judge, reasons, and law_enforcements (L1-RECORDED / L2-EVIDENCED / L3-GATED).
Triggers / 触发词
这些是用户真实会说的话(中 / 英对照),命中任一即应加载本技能:
- 「算法改了一版,要不要走变更」 / "we changed the algorithm, does it need review"
- 「模型重新训练需要重新申报吗」 / "does retraining the model require a new filing"
- 「这次改动算重大还是轻微」 / "is this a major or minor change"
- 「帮我出变更风险判定结论」 / "give me a change risk verdict"
- 「阈值调整能不能直接上线」 / "can we ship the threshold tweak as is"
- 「审评时怎么解释这次算法更新」 / "how do we justify this update to reviewers"
Trigger when the user says any of the above, in Chinese or English.
Ecosystem / 生态入口
This skill is one door of the LGD ecosystem — every entry below is publicly reachable. 本技能是 LGD 生态 的一个入口,下列地址均对公网可达。
- Ecosystem hub / 生态总入口: https://zhaoxinghua09-cell.github.io/lgd-hub/
- LGD theory (canonical) / 理论正本: https://github.com/zhaoxinghua09-cell/LGD
- Formal paper (DOI) / 形式化论文: https://doi.org/10.5281/zenodo.22456647
- Protocol & reference implementation / 协议与参考实现: https://github.com/zhaoxinghua09-cell/uibc-core
- Source of this skill / 本包源码: https://github.com/zhaoxinghua09-cell/agent-skills/tree/main/skills/lgd-change-gate
- All published skills / 全部已发布技能: https://github.com/zhaoxinghua09-cell/agent-skills
- AI-readable index / AI 说明书: https://github.com/zhaoxinghua09-cell/agent-skills/tree/main/skills/lgd-change-gate/llms.txt
Looking for something else? / 还想要别的?
Tell your AI assistant: "Find me an LGD skill that can " — the hub lists every published skill with a one-line purpose. 跟你的 AI 助手说一句「去 LGD 生态超市找能……的技能」,总入口按用途列全了所有已发布技能。
Data handling / 数据处理边界
- 默认不向任何外部地址发送数据:扫描结果、报告、命中明细一律只写本地或只打印到当前会话。
- 如需外发(告警、工单、邮件、IM),必须由人先明确指定接收方,并在每次发送前逐次确认;本技能不预设、也不硬编码任何接收端。
- 只处理调用者明确指定的目标路径,不遍历无关目录,不跟随符号链接,不读取目标之外的凭据文件。
- 命中内容在输出时按需脱敏,避免把敏感原值写进日志或对话记录。
- Downloads
- Security audit
- Review
- Last updated
- 1w ago
- Current version
- v0.1.7
- License
- MIT-0
