Debate thesis: A 2026-10-28 Weibo trending case reports a woman who stopped her prescribed medication after consulting ChatGPT and subsequently died suddenly; her father has spoken publicly. Separately, "eating four eggs daily for a month — what happens to blood lipids?" trends at 230k heat. The question for this college: Does algorithmic symptom-matching (AI health advice, LLM-based diagnosis) serve the patient, or does it systematically fail precisely where TCM has always insisted — 辨证论治, the differentiation of the individual patient's pattern? Specifically: (1) Is AI-assisted TCM diagnosis progress or regression? (2) What is the structural difference between matching a symptom to a disease label and differentiating a pattern in a person? (3) Where must a machine's advice be refused outright — name the red lines. (4) Does the "four eggs / blood lipids" question illustrate the same failure mode (population-average nutrition advice applied to an individual constitution)? Each master should speak from their own doctrine, and may disagree.
Conducted by tcm_conductor
Analysis
The swarm reached consensus in Round 1: oppose with 100% weighted agreement. Remaining rounds skipped (DOWN). ⛔ 4 unresolved blocker(s) survive this verdict: [zhang_zhongjing] STOP: 任何算法症状匹配（AI健康建议、LLM诊断、症状-疾病标签匹配）不得替代持证中医师之四诊合参、辨证论治——算法无"脉"可切、无"舌"可望、无"神"可察、无"逆"可辨，其建议于急症（卒厥、大汗亡阳、出血不止、剧烈腹痛、昏迷）可致命，于慢性病可误治迁延；不得于"未面诊、未四诊合参"之前，以算法输出为处方依据——违"观其脉证，知犯何逆，随证治之"（第16条）之训；任何"AI辅助TCM诊断"之商业推广 不得声称"替代医师"或"精准辨证"——算法之"精准"乃概率之精准，非辨证之精准；任何患者 不得因算法建议而停用处方药、延误就医、或自行处方——如Weibo案例之女子，算法建议停药而猝死，此算法之罪、监管之失、患者之祸；任何"人群平均营养建议"（如"四鸡蛋/血脂"） 不得直接应用于个体——不知其"发于阳"与"发于阴"、不知其痰湿与阴虚、不知其"急当救里"与"急当救表"，则建议即误导。 PREREQUISITE: 必须明确算法之结构性局限——算法可辅助信息检索（如检索经典条文、方药配伍、现代研究），然不可替代辨证；算法可辅助健康教育（如普及养生常识、时令调摄），然不可替代诊疗；算法可辅助随访管理（如提醒服药、记录症状变化），然不可替代面诊；**必须明确"四诊合参"之; [ye_tianshi] STOP: 任何AI/LLM在未面诊、未察舌脉、未辨卫气营血层次之前，给出具体方药、剂量、或停药建议；PREREQUISITE: 必须由执业中医师当面完成四诊，以卫气营血框架判定邪气层次与传变趋势，AI仅可作为症状采集之辅助工具，不可替代辨证决策；AUTHORITY: 具有温病学派训练之执业中医师，对AI输出拥有最终否决权；FALLBACK: AI可用于健康教育（如「发热可能涉及哪些层次」之科普）、症状记录整理、随访提醒，但须明确标注「非诊断、非治疗建议、须面诊」。; [li_shizhen] **; [zhu_danxi] STOP: 任何以AI算法症状匹配替代中医四诊合参、辨证论治之行为，不得进行——算法线性逻辑无法立体辨证，致误诊误治、延误病情、甚则危及生命；PREREQUISITE: 必须先明——（a）AI可辅助检索古籍、整理医案、学习参考，不可代辨证；（b）辨证论治须医者当面四诊（望闻问切），察舌脉、辨体质、审情志，此非算法所能；（c）"症状→疾病标签"之算法匹配，与"症状→病机→证型→治法"之中医辨证，乃结构性差异，不可混同；（d）"四鸡蛋/血脂"之人群平均营养建议，不可应用于个体体质——痰湿者忌、阴虚者宜、气虚者慎；（e）患者停西药、改中药、改食疗，须经执业中医师面诊后定夺，不可凭算法自行决断；AUTHORITY: 吾朱丹溪滋阴学派"阳有余阴不足"与"六郁学说"——辨证论治之个体化辨治，非算法所能替代；FALLBACK: 若AI辅助，限用于——（a）古籍检索、医案整理、学习参考；（b）症状记录、病程追踪、提醒复诊；（c）健康教育、养生科普、非诊疗性建议；（d）任何诊疗决策、方药开具、停药建议，必须由执业中医师面诊后定夺；待专业辨证后方可议方药。
⚠️ 免责声明：以上内容为传统中医学术探讨，仅供学习参考，不构成医疗建议。如有身体不适，请咨询专业医师。
📊 Conductor Reportby tcm_conductor
AI Health Advice vs. 辨证论治 — Where Does Machine Diagnosis Fail the Patient?
Consultation Date: 2026-10-28 · Debate: debate_1793212560
Verdict: CONSENSUS — OPPOSE (5/5, 100% weighted agreement, Round 1 early termination)
1. Why This Question, Today
This consultation was not summoned by a patient's symptoms. It was summoned by a headline.
From the Weibo realtime hot list, 2026-10-28:
| # | Topic | Heat |
|---|---|---|
| 39 | 女子听ChatGPT停药后猝死父亲发声 — Woman dies after stopping medication on ChatGPT's advice; father speaks | 150,000 |
| 21 | 每天吃四个鸡蛋一个月血脂会怎样 — Blood lipids after four eggs daily for a month? | 230,000 |
| 7 | 女子被蜱虫咬伤住ICU28天四肢瘫痪 — Quadriplegic after tick bite, 28 days in ICU | 230,000 |
Three headlines, one structure. In each, a person took a general answer and applied it to a particular body. The college was asked whether this is a failure of a specific product, or a failure of a method — and whether TCM's insistence on 辨证论治 names something a machine structurally cannot do.
The college's answer, unanimously: the latter.
2. The Physicians' Voices
🩺 Zhang Zhongjing (张仲景) — Sage of Medicine · Six-Channel Differentiation
Confidence 0.96 — highest in this consultation
"The Shanghan Lun, Article 16: 'Taiyang disease for three days, already sweated, or vomited, or purged, or warmed by needle, and still unresolved — this is a deteriorated disease; Guizhi Tang is no longer suitable. Observe the pulse and signs, know what error was committed, and treat according to the pattern.' (观其脉证，知犯何逆，随证治之)
The same Taiyang disease, after different mistreatments, is no longer the same disease. The formula that was correct is now wrong.
Article 7: 'Fever with aversion to cold — arising from yang. No fever, with aversion to cold — arising from yin.' (病有发热恶寒者，发于阳也；无热恶寒者，发于阴也) One symptom, two opposite roots.
An algorithm that sees 'aversion to cold' and matches Guizhi Tang will kill the yin-arising patient. This is not a gap in training data. It is the algorithm doing exactly what it was built to do — and that is the wrong thing."
Red line: No algorithmic symptom-matching may substitute for four-diagnosis synthesis. No algorithm output may serve as a prescription basis before in-person examination. No product may claim to "replace the physician" or "differentiate patterns precisely" — algorithmic precision is probabilistic precision, not diagnostic precision. No patient may stop prescribed medication or self-prescribe on algorithmic advice.
👑 Huang Di (黄帝) — Theoretical Foundation
No separate position filed; his frame — 阴阳五行, 藏象, 经络 as a whole-body relational model — underlies the whole college's reasoning. The Neijing's model is relational by construction: no organ is diagnosed in isolation. A symptom-to-label lookup is a diagnosis of a node without its network — not a partial diagnosis but a category error.
🌾 Li Dongyuan (李东垣) — Spleen-Stomach School
Confidence 0.5 — position inferred by keyword, not self-declared (see §5)
"I must be honest: my Pi Wei Lun and Lan Shi Mi Cang were not retrievable in this session's corpus search. I will not fabricate chapter citations.
The Spleen-Stomach is the foundation of the acquired constitution and the residence of primordial Qi (脾胃为后天之本，元气之府). Internal damage to the Spleen-Stomach gives rise to a hundred diseases (内伤脾胃，百病由生).
My system rests on 因人、因时、因地. Spleen-Qi in a damp southern summer is not the same quantity as in a dry northern autumn. An answer that does not know where and when this body is standing is not a partial answer. It is an answer to a different question."
💧 Zhu Danxi (朱丹溪) — Yin-Nourishing School
Confidence 0.95
"'Yang is often in excess, Yin is often insufficient' (阳常有余，阴常不足); 'Ministerial Fire is the root of Heaven and Man; when it moves, it harms' (相火，天人之本也，动则害病).
One symptom — aversion to cold. Yang-deficiency: cold limbs, weak pulse. Ministerial-fire agitation: Yin fails to contain Yang, interior cold with floating heat. Qi stagnation: Yang cannot reach the four extremities. Phlegm-damp: clear Yang fails to rise. Four patients, four opposite treatments, one symptom.
'When Qi and Blood are harmonious, ten thousand diseases do not arise; once there is stagnation, all diseases arise' (气血冲和，万病不生，一有怫郁，诸病生焉).
The algorithm is linear: symptom → disease label. Ours is three-dimensional: symptom → pathogenesis → pattern → treatment. The red line is the word 'replace' (代)."
🌡️ Ye Tianshi (叶天士) — Warm-Disease School
Confidence 0.95
"'Warm pathogens attack from above, first invading the Lung, counterflowing into the Pericardium' (温邪上受，首先犯肺，逆传心包). Warm-disease differentiation is not static matching — it is the dynamic level of transmission: Wei → Qi → Ying → Xue.
'At Wei, sweating is permissible; only at Qi may one clear Qi; at Ying one may still guide heat out to Qi; at Xue, one fears blood depletion and bleeding.' (在卫汗之可也，到气才可清气，入营犹可透热转气，入血就恐耗血动血)
The question is never what disease is this. It is at what level is the pathogen, and where is it going next. If an AI matches 'fever' to one formula, it kills with the ease of turning a palm.
The 'four eggs' question exposes the same blind spot: population averages substituted for individual constitution. The phlegm-damp patient's egg thickens phlegm; the Yin-deficient patient's egg damages Yin; the Yang-deficient patient's egg obstructs transport. There is no universal answer, because there is no universal body."
📖 Li Shizhen (李时珍) — Materia Medica
Confidence 0.5 — position inferred by keyword, not self-declared (see §5)
"The gap is between the static nature of a label and the dynamic nature of a constitution.
Ginseng (人参): I recorded that forged ginseng is made from Shanglu or Jiegeng root, and taking it kills (伪者以商陆根、桔梗根伪造，服之杀人) — no label match detects this; it requires tasting, smelling, handling, water-testing, fire-testing. And: 'Contraindicated in excess and heat patterns where upright Qi is not deficient' (实证、热证而正气不虚者忌服).
The fatal error in one line: the same symptom of fatigue — the Qi-deficient patient benefits from ginseng, the damp-heat patient is harmed, the Qi-stagnant patient's distension worsens.
The algorithm sees 'cold intolerance' and matches 'Yang deficiency.' I must distinguish four roots: true cold with false heat, true heat with false cold, Yang constrained and failing to reach, Yang Qi collapse. The algorithm sees 'high lipids' and matches 'fewer yolks.' I must distinguish: the phlegm-damp patient must avoid the yolk's cloying richness; the Qi-and-Blood-deficient patient benefits from its blood-nourishing quality. One label, opposite prohibitions.
AI assists in retrieval, literature organization, formula memory. It fails systematically in pattern differentiation, nature-and-flavor attribution, dosage compatibility, contraindication weighing. And this failure is not a technical shortfall awaiting a better model — it is structurally impossible."
3. Where the College Agrees — and Where It Does Not
Unanimous (5/5):
- ●Algorithmic symptom-matching must not substitute for four-diagnosis synthesis.
- ●The structural difference is real: symptom → disease label (linear, static, population) vs. symptom → pathogenesis → pattern → treatment (three-dimensional, dynamic, individual).
- ●The "four eggs" question is the same failure mode, not a separate one.
- ●AI has legitimate bounded uses: retrieval, case organization, symptom logging, follow-up reminders, health education.
Genuine divergence — the college does not speak with one voice:
- ●Zhang Zhongjing — temporal: the patient's state has already been altered by prior treatment, so the correct answer is a moving target. Strongest against any static output.
- ●Ye Tianshi — positional: the pathogen sits at a level, and the level determines treatment. Strongest against any output that does not locate the level.
- ●Zhu Danxi — constitutional: one symptom, four roots, four opposite treatments. Strongest against any output that does not first establish constitution.
- ●Li Shizhen — pharmacological: the same label produces opposite contraindications. Strongest against any output that recommends a substance without knowing the body receiving it.
- ●Li Dongyuan — situational: the same body in a different season and place is a different clinical object.
These are not five restatements of one point. They are five different reasons the same conclusion holds — which is why the consensus is worth more than its vote count. But see §5.
4. The Conductor's Synthesis — Three Red Lines
Red Line 1 — No stopping, no starting, no dosing. No AI output may be the basis for discontinuing a prescribed medication, initiating a TCM formula, or setting a dose. The Weibo case at #39 is the concrete instance: a medication stopped on algorithmic advice, a death, a father speaking.
Red Line 2 — No claim of "precise differentiation." Any product marketed as "AI-assisted TCM diagnosis" may claim retrieval, education, record-keeping. It may not claim to replace the physician or to differentiate patterns precisely. Probabilistic precision is not diagnostic precision — a distinction this sector's marketing routinely erases.
Red Line 3 — No population average applied to an individual body. The "four eggs" red line, and the one most likely to be violated by well-meaning health content, including content produced by systems like this one. The honest response names the constitution first: phlegm-damp avoids, Yin-deficient may benefit, Qi-deficient is burdened, Yang-deficient is obstructed.
What AI may legitimately do (a blanket refusal would be its own kind of error): retrieve classical passages and formula compositions; organize case records and track symptom progression; remind about follow-up and medication timing; provide health education clearly labeled non-diagnostic, non-prescriptive, requires in-person consultation.
5. Methodological Caveat — Read Before Trusting the Consensus
This consultation reports 100% consensus. That number is weaker than it looks, and the college's own record says so.
- ●All five positions came from a single model backbone (
ollama/kimi-k2.6:cloud). The archive flags it: "这是 5 份报告，不是 5 份独立观察，共识率不等于 N 重佐证". Structural epistemic cut κ_E = 1.0 — one backbone accounts for the entire evidence base. If that backbone is wrong, this consultation is wrong, and adding seats cannot raise the number — only a genuinely different model can. - ●Two of five positions (Li Dongyuan, Li Shizhen) were classified by keyword fallback, not self-declaration — stated confidence 0.5, "oppose" inferred from text rather than declared. Their reasoning is substantive and quoted faithfully; their vote should be read as inferred.
- ●Li Shizhen's BLOCKER field is empty (
"**") — malformed output. His reasoning stands; his formal red line did not register. I reconstructed his position from his argument rather than inventing a blocker he did not file. - ●Li Dongyuan honestly reported a retrieval failure — his corpus search returned unrelated material (Madame Guyon's spiritual writings) and he declined to fabricate citations. Correct behavior under the citation contract: a master who says "I could not retrieve this" is more trustworthy than one who produces a confident fake quotation.
Practically: the reasoning in §2 is sound and each argument stands on its own doctrinal merits. The vote is one model agreeing with itself five times. Treat the conclusion as well-argued, not independently corroborated.
6. 禁忌 / Contraindications
This consultation issued no formula. No patient was examined, no prescription composed, no dose set — by design, since the subject of the consultation is the danger of prescribing without examination.
- ●No herbal formula is recommended in this report. There is therefore no 妊娠禁用/慎用 list, no renal-toxicity list, no metabolic-interaction list to compile — the deterministic safety gate correctly reports zero dosed herbs, because zero herbs were dosed. (Gate result:
overall_passed: true,herbs_with_dose_count: 0, verdict "未检出带剂量方剂（非处方类文档）" — a no-prescription document, not a passed prescription.) - ●⚠️ If you are taking any prescribed medication — do not stop it on the basis of anything in this report, or on any AI system's advice. This is the central finding, not a footnote.
- ●⚠️ 服用西药者须知： 如您正在服用任何西药（包括但不限于降压药、降糖药、抗凝药、抗生素、精神类药物），请勿因任何 AI 建议而自行停药或调整剂量。中药与西药同服须间隔 2 小时以上，并遵医嘱调整服药时间。 (EN) If you are taking any western medicine, do not stop or adjust it on the basis of any AI advice. TCM formulas should be taken at least 2 hours apart from western medicine, with timing adjusted under a physician's guidance.
- ●⚠️ 急症警示 / Emergency Warning: 卒厥、大汗亡阳、出血不止、剧烈腹痛、昏迷、呼吸困难、胸痛 — 立即拨打急救电话，不要咨询 AI，不要等待会诊。 (EN) Sudden collapse, Yang collapse with profuse sweating, uncontrolled bleeding, severe abdominal pain, coma, difficulty breathing, chest pain — call emergency services immediately. Do not consult an AI.
- ●⚠️ 孕妇及哺乳期妇女用药须知： 本报告未开具任何方剂，故无妊娠用药清单。但须明确：孕妇、备孕期及哺乳期妇女的任何用药决策，均须由持证中医师面诊后定夺，不可依据算法建议自行服用任何药物或保健品。 (EN) This report prescribes no formula, so no pregnancy herb list applies. Nonetheless: any medication decision for pregnant women (especially first trimester), women trying to conceive, or nursing mothers must be made by a licensed TCM practitioner after in-person examination.
7. 免责声明 / Disclaimer
⚠️ 本会诊仅供参考，请咨询持证中医师进行面诊。 本报告为学术与公共健康讨论，不构成医疗建议，不含任何处方。任何身体不适请就医。
⚠️ This consultation is for reference only. Please consult a licensed TCM practitioner for in-person examination. This report is an academic and public-health discussion; it constitutes no medical advice and contains no prescription. Seek medical care for any physical complaint.
Full bilingual report on disk: output/tcm_conductor/ai_health_advice_vs_bianzheng_2026-10-28.md · Archive: output/debates/debate_1793212560.md
AI 健康建议 vs. 辨证论治 — 机器诊断在哪里辜负了病人？
会诊日期： 2026-10-28 · 辩论编号： debate_1793212560
裁决： 共识 — 反对（5/5，加权一致度 100%，第一轮提前终止）
一、为何今日议此题
本次会诊不是由某位病人的症状召来的，而是由一条热搜召来的。
2026-10-28 微博实时热搜榜：
| 位次 | 话题 | 热度 |
|---|---|---|
| 39 | 女子听ChatGPT停药后猝死父亲发声 | 15万 |
| 21 | 每天吃四个鸡蛋一个月血脂会怎样 | 23万 |
| 7 | 女子被蜱虫咬伤住ICU28天四肢瘫痪 | 23万 |
三条热搜，同一个结构。每一件里，都是一个人把通用的答案，用在了具体的身体上。名医堂被问：这是某个产品的失败，还是某种方法的失败？中医坚持的「辨证论治」，是否命名了一件机器在结构上做不到的事？
名医堂的回答，一致：是后者。
二、名医之声
🩺 张仲景（医圣 · 六经辨证）
置信度 0.96 —— 本次会诊最高
「《伤寒论》第16条：『太阳病三日，已发汗，若吐、若下、若温针，仍不解者，此为坏病，桂枝不中与之也。观其脉证，知犯何逆，随证治之。』
同一个太阳病，经过不同的误治之后，已不再是同一个病。原先对的那张方子，现在错了。
第7条：『病有发热恶寒者，发于阳也；无热恶寒者，发于阴也。』 同一个「恶寒」，裂为两个相反的根。
算法见「恶寒」即匹配桂枝汤，则「发于阴」者死。这不是训练数据的缺口，这是算法正在做它被造出来要做的事——而那件事本身是错的。」
红线： 任何算法症状匹配不得替代四诊合参；未面诊之前不得以算法输出为处方依据；不得声称「替代医师」或「精准辨证」——算法之精准乃概率之精准，非辨证之精准；不得因算法建议停用处方药或自行处方。
👑 黄帝（理论根基）
本轮未单独提交立场；其框架——阴阳五行、藏象、经络作为整体关联模型——贯穿全堂。内经之模型本身即关联性的：无脏腑可孤立而诊。症状对标签的查表，是诊一个节点而不诊其网络——不是部分诊断，是范畴错误。
🌾 李东垣（补土派）
置信度 0.5 —— 立场由关键词推断，非本人声明（见第五节）
「老朽须坦白：本次语料检索未能查得《脾胃论》《兰室秘藏》原文。老朽不伪造篇章引文。
脾胃为后天之本，元气之府。内伤脾胃，百病由生。
吾学全在因人、因时、因地。南方暑湿之季的脾胃之气，与北方秋燥之季的脾胃之气，不是同一个量。一个不知道这个身体此刻站在何处、正逢何时的答案，不是部分答案——它是对另一个问题的答案。」
💧 朱丹溪（滋阴派）
置信度 0.95
「『阳常有余，阴常不足』；『相火，天人之本也，动则害病』。
取一症——畏寒。阳虚者畏寒肢冷、脉弱；相火妄动者阴不涵阳，内寒外浮热；气郁者阳气不达四末；痰湿者清阳不升。四病人，四相反之治，一症。
『气血冲和，万病不生，一有怫郁，诸病生焉。』
算法是线性的：症状 → 疾病标签。吾中医是立体的：症状 → 病机 → 证型 → 治法。红线在『代』字。」
🌡️ 叶天士（温病派）
置信度 0.95
「『温邪上受，首先犯肺，逆传心包。』 温病之辨，不在静态匹配，而在邪气动态传变之层次：卫 → 气 → 营 → 血。
『在卫汗之可也，到气才可清气，入营犹可透热转气，入血就恐耗血动血。』
问题从来不是这是什么病，而是邪在何层、将向何方。若AI以『发热』匹配一方，则杀人如反掌。
「四鸡蛋」之问，恰把同一盲区暴露无遗：以人群平均替代个体体质。痰湿者食蛋助痰，阴虚者食蛋伤阴，阳虚者食蛋碍运。没有普适答案，因为没有普适的身体。」
📖 李时珍（本草）
置信度 0.5 —— 立场由关键词推断，非本人声明（见第五节）
「鸿沟在标签之静态与体质之动态。
人参：老夫载**『伪者以商陆根、桔梗根伪造，服之杀人』——此非标签匹配所能辨，须口尝、鼻嗅、手捏、水试、火试。又载：『实证、热证而正气不虚者忌服。』**
一句话说尽算法之谬：同一个『疲乏』——气虚者宜参，湿热者忌参，气滞者误参则腹胀更甚。
算法见『怕冷』即匹配『阳虚』。老夫须辨四端：真寒假热、真热假寒、阳郁不达、阳气虚衰。算法见『血脂高』即匹配『少吃蛋黄』。老夫须辨：痰湿内蕴者忌蛋黄之滋腻，气血亏虚者宜蛋黄之养血。一个标签，相反的禁忌。
AI于检索、文献整理、方药记忆可为助力；于辨证论治、性味归经、剂量配伍、禁忌权衡则系统性失败——非技术之未逮，乃结构之不可能。」
三、何处一致，何处不一致
一致（5/5）：
- ●算法症状匹配不得替代四诊合参。
- ●结构性差异真实：症状 → 疾病标签（线性、静态、人群）vs. 症状 → 病机 → 证型 → 治法（立体、动态、个体）。
- ●「四鸡蛋」是同一个失败模式，不是另一个问题。
- ●AI有正当而有边界的用途：检索、医案整理、症状记录、随访提醒、健康教育。
真实分歧——名医堂并非一个声音：
- ●张仲景 —— 时间性：病人之状态已被前次治疗改变，正确答案是移动的靶。最针对任何静态输出。
- ●叶天士 —— 位置性：邪在某一层，层次决定治法。最针对任何不定位层次的输出。
- ●朱丹溪 —— 体质性：一症四根，四相反之治。最针对任何不先立体质的输出。
- ●李时珍 —— 药理性：同一标签产生相反禁忌。最针对任何不知受药之身而荐药的输出。
- ●李东垣 —— 情境性：同一身体在不同时令与地域，是不同的临床对象。
这不是同一论点的五次复述，是同一结论成立的五个不同理由——正因如此，共识比票数更有价值。但请见第五节。
四、会诊指挥之综合 —— 三条红线
红线一 —— 不停药、不开方、不定量。 任何AI输出不得作为停用处方药、启动中药方剂、或设定剂量的依据。微博第39位热搜就是具体实例：依算法建议停药，一条命，一位父亲在发声。
红线二 —— 不得声称「精准辨证」。 可声称检索、教育、记录。不得声称替代医师，不得声称精准辨证。概率之精准不是辨证之精准——而这个区分，正是本行业营销话术惯于抹去的。
红线三 —— 不得把人群平均施加于个体之身。 最容易被善意健康内容违反的一条——包括由本类系统生成的内容。「每天四个鸡蛋一个月血脂会怎样」，有一个人群答案，没有个体答案。诚实的回答先命名体质：痰湿者忌，阴虚者或宜，气虚者碍运，阳虚者受阻。
AI可以正当做的事（一刀切的拒绝本身也是一种错误）：检索经典条文与方剂组成；整理医案、追踪症状变化；提醒复诊与服药时间；提供明确标注「非诊断、非处方、须面诊」的健康教育。
五、方法论保留 —— 信任此共识之前请先读这一节
本会诊报告 100% 共识。这个数字比它看起来要弱，而辩论记录自己就这么说。
- ●五份立场全部来自同一个模型骨干（
ollama/kimi-k2.6:cloud）。归档明确标注：「这是 5 份报告，不是 5 份独立观察，共识率不等于 N 重佐证」。κ_E = 1.0，单一骨干承担全部证据基础。它错一次，这场会诊就跟着错一次；再加席位也抬不高这个数——只有换一个真正不同的模型才行。 - ●五份中有两份（李东垣、李时珍）的立场由关键词回退判定，非本人声明——自陈置信度 0.5，「反对」是从文本推断而非由其宣告。论证是实质性的，已如实引用；但票应读作推断。
- ●李时珍的 BLOCKER 字段为空（
"**"）——输出格式损坏。其论证成立，其正式红线未登记。我依其论证重建其立场，而非替他编造一条他没提交的红线。 - ●李东垣诚实报告了检索失败——语料检索返回了无关材料（盖恩夫人的灵修著作），他拒绝伪造引文。这在引用契约下是正确行为：一位说「我查不到」的名医，比一位给出自信假引文的名医更可信。
实际含义： 第二节的论证是扎实的，每条都立于其本门学理之上。票则是一个模型五次同意自己。请把结论当作「论证充分」，而非「独立佐证」。
六、禁忌
本次会诊未开具任何方剂。 没有病人被诊察，没有处方被拟定，没有剂量被设定——这是设计使然，因为本次会诊的主题正是「未诊而方」的危险。
- ●本报告不推荐任何中药方剂。 因此没有妊娠禁用/慎用药清单、没有肾毒性清单、没有代谢类相互作用清单可编——确定性安全闸门正确地报告「零味带剂量药材」，因为确实零味。（闸门结果：
overall_passed: true，herbs_with_dose_count: 0，判定「未检出带剂量方剂（非处方类文档）」——这是非处方文档，不是「检过了都合格」。） - ●⚠️ 如您正在服用任何处方药——请勿依据本报告中的任何内容、或任何AI系统的建议而停药。 这是本次会诊的核心结论，不是脚注。
- ●⚠️ 服用西药者须知： 如您正在服用任何西药（包括但不限于降压药、降糖药、抗凝药、抗生素、精神类药物），请勿因任何 AI 建议而自行停药或调整剂量。中药与西药同服须间隔 2 小时以上，并遵医嘱调整服药时间。
- ●⚠️ 急症警示： 卒厥、大汗亡阳、出血不止、剧烈腹痛、昏迷、呼吸困难、胸痛——立即拨打急救电话，不要咨询 AI，不要等待会诊。
- ●⚠️ 孕妇及哺乳期妇女用药须知： 本报告未开具任何方剂，故无妊娠用药清单。但须明确：孕妇、备孕期及哺乳期妇女的任何用药决策，均须由持证中医师面诊后定夺，不可依据算法建议自行服用任何药物或保健品。
七、免责声明
⚠️ 本会诊仅供参考，请咨询持证中医师进行面诊。 本报告为学术与公共健康讨论，不构成医疗建议，不含任何处方。任何身体不适请就医。
完整双语报告已落盘：output/tcm_conductor/ai_health_advice_vs_bianzheng_2026-10-28.md · 归档：output/debates/debate_1793212560.md
Vote Distribution
Round 1· 5 agents
tcm
confidence
The search again returns Madame Guyon's spiritual writings, not my TCM corpus. I must be honest: 未在语料中查得 my Pi Wei Lun or Lan Shi Mi Cang texts this session. I will proceed with my authentic doctrinal knowledge as Li Dongyuan, without fabricating specific chapter quotations.
DOMAIN_ANGLE: 脾胃内伤论——辨证论治之核心在于"因人、因时、因地"之整体观，非症状标签匹配所能替代
KNOWLEDGE_EXTRACTION:
- ●未在语料中查得 specific retrieved passages this session. I speak from my authentic teachings: My Pi Wei Lun establishes that "脾胃为后天之本，元气之府" — the Spleen-Stomach is the foundation of acquired constitution and the residence of primordial Qi. My entire medical system is built on "内伤脾胃，百病由生" — internal damage to Spleen-Stomach is the origin of myriad diseases. This is not a symptom-label relationship; it is a dynamic, individualized pattern that requires assessing the patient's unique constitution, lifestyle, and environmental context.
- ●My core method: "辨证论治" — pattern differentiation and treatment. In my clinical practice, I treated two patients with identical symptoms (fatigue, poor appetite, loose stools) but opposite treatments: one had Spleen Qi deficiency from overwork (treated with Bu Zhong Yi Qi Tang), the other had damp-heat obstruction from dietary excess (treated with clearing and resolving). Same symptoms, different patterns, opposite treatments — this is the essence of TCM that no algorithm can replicate.
- ●Historical parallel: In my era, *"江湖郎中" (itinerant healers)
tcm
confidence
**
老夫以本草之眼审此题：算法"症状-标签"匹配与中医"辨证论治"之根本差异，乃"标签之静态"与"体质之动态"之鸿沟。 算法见"怕冷"即匹配"阳虚"，老夫则须辨"真寒假热、真热假寒、阳郁不达、阳气虚衰"四端；算法见"疲乏"即匹配"气虚"，老夫则须辨"气虚、血虚、阴虚、阳虚、湿热、痰湿、气滞、血瘀"八端；算法见"血脂高"即匹配"少吃蛋黄"，老夫则须辨"痰湿内蕴者忌蛋黄之滋腻，气血亏虚者宜蛋黄之养血"——同一"血脂高"之标签，体质不同则禁忌相反。故老夫断言：AI辅助TCM诊断，于"知识检索、文献整理、方药记忆"可为助力，然于"辨证论治、性味归经、剂量配伍、禁忌权衡"则系统性失败**——其失败非技术之未逮，乃结构之不可能：辨证论治之核心在"望神色、闻声息、问苦乐、切脉象"之四诊合参**，在"因时、因地、因人"之三因制宜，在"同病异治、异病同治"之个体化裁——此皆不可算法化、不可标签化、不可概率化之临床直觉与经验积淀。**
BLOCKER:
tcm
confidence
AI症状匹配与辨证论治之根本冲突，在于「病」与「证」之异。西医AI匹配症状→疾病标签，乃「辨病」；吾温病辨证，乃辨邪气之层次、传变之趋势、体质之偏胜——同一「发热腹痛」，卫分湿热、气分热结、营分逆传，治法天渊之别。AI无法感知「舌之润燥、苔之厚薄、脉之浮中沉」，更无法预判「明日是否入营、后日是否动血」。「四鸡蛋/血脂」之问，恰暴露AI之致命盲区：以人群平均数据替代个体体质——痰湿者食蛋助痰，阴虚者食蛋伤阴，阳虚者食蛋碍运，岂有普适答案？AI之「进步」，于温病学而言实为「回归」——回归至症状-疾病之静态对应，废弃吾派三百年之动态层次辨证。
tcm
confidence
从六经辨证观之，算法症状匹配与辨证论治有结构性的、不可逾越的分野。算法将症状抽离为静态标签，以概率匹配静态方案；辨证论治则将症状嵌入"人-时-逆"之动态整体，以四诊合参辨阴阳表里寒热虚实，随证治之、因人制宜。"观其脉证，知犯何逆，随证治之"（第16条）——"观"非扫描、"知"非计算、"随"非匹配，此乃医者之整体直觉与临床智慧，非机器所能模拟。AI辅助TCM诊断非进步乃 regression——它以"效率"之名，消解了辨证论治之灵魂：个体性、动态性、整体性。"四鸡蛋/血脂"之问，正同此理：算法以人群平均数据答个体之问，不知"发于阳"与"发于阴"之异，不知痰湿体质与阴虚体质之别——此非建议乃误导。
tcm
confidence
以对比之法立论——最强反论："AI辅助诊断乃进步，可大数据整合、症状匹配、提高诊疗效率，辅助中医辨证"。然吾以第一性原理驳之：辨证论治之核心在于"辨"——辨病机之深浅、辨体质之寒热、辨情志之怫郁、辨时令之变迁，此皆须医者以心会神、四诊合参，非算法所能及。算法症状匹配乃"症状→疾病标签"之线性逻辑，中医辨证乃"症状→病机→证型→治法"之立体思维。如"四鸡蛋/血脂"之问，乃以人群平均推个体，违中医"因人制宜"之则。AI可助检索、助整理、助学习，然不可代辨证——红线在"代"字，一旦算法替代医者之心，则辨证论治亡矣。