Add user-led reading calibration protocol

references/ai-reading-workflow-prompt.md

@@ -1,7 +1,7 @@
 # AI解盘工作流Prompt工程（AI Reading Workflow）
 
 > **适用场景**：AI收到出生信息、PDF星盘或文字星盘后，如何一步步执行完整的解盘+推运分析
-> **版本**：v5.1.0 | **更新日期**：2026-06-27
+> **版本**：v5.2.0 | **更新日期**：2026-07-02
 > **来源标签**: 【工具/模板】 — AI解盘执行引擎工作流
 > **优先级**：⭐⭐⭐⭐⭐（AI解盘质量的决定性文件）
 > **定位**：本文件是AI解盘的**执行引擎**，将Skill中所有参考资料串联成可执行的工作流
@@ -9,6 +9,7 @@
 > **v4.0重大变更**：Step 0.5 Karaka系统自动识别（解决DK摇摆问题）、预测输出强制[A/B/C]标注、统一精度边界声明、来源标签体系引用
 > **v5.0重大变更**：强制外部验证门控（MEVG）——新增核心原则"不凭记忆"、三个门控步骤（Step 3.11/4.10/5.5）嵌入工作流、与"不跳步"同级强制
 > **v5.1重大变更**：事件判定骨架接入主工作流——新增 `Route -> Evidence Ledger -> Adjudication -> Output Contract` 四段式事件裁决层，并引入 marriage 专用 adjudicator。
+> **v5.2重大变更**：用户追问校准协议接入主工作流——把用户的质询流程抽象为盲推隔离、具体事件选项、证据先行、相似案例分层对标、可迁移边界与反例处理。
 
 ---
 
@@ -24,6 +25,125 @@
    - ❌ 仅凭记忆说"Venus 在 Taurus 10 宫是 Malavya Yoga"
    - ✅ 先 web_search 确认形成条件 + 引用来源
    - → 完整协议：`references/mandatory-verification-gate-protocol.md`
+8. **追问即校准（v5.2 新增）**：用户质疑"你用了哪些技法""哪里遗漏""真实案例是否相似""为什么这样判断"时，必须把追问视为工作流校准信号，而不是普通解释请求；需要回到证据层、相似案例层和置信度边界重新裁决。
+
+---
+
+## 🔴🔴🔴 用户追问校准与相似案例对标协议（v5.2）
+
+> **目的**：把高质量用户追问转化为解盘流程约束，避免泛泛描述、记忆联想、单案例套用或把未闭环内容伪装成完整判断。
+
+### 0.1 触发条件
+
+只要用户出现以下任一追问，必须启动本协议：
+
+- "不要用我之前说过的反馈"，"重新盲推"，"只用出生参数和当前项目数据"
+- "你是不是遗漏了技法"，"这是不是顶级印度占星师逻辑"
+- "列具体时间和事件让我选择"，"不要问概念问题"
+- "相似真实案例是什么"，"这个案例到底和我哪里像"
+- "为什么这样判断"，"哪些地方需要修正"
+- "这个机会具体是什么"，"事件节点是什么"
+
+### 0.2 盲推隔离模式
+
+当用户要求盲推或禁用既往反馈时：
+
+1. 禁止调用用户此前提供的人生事件、校时反馈、对错判断、职业身份、关系状态、偏好描述。
+2. 只允许使用出生参数、当前项目现跑出的原始数据、已分级 source pack、MEVG 外部证据和真实案例校准层。
+3. 输出必须声明：
+   - `User Feedback Isolation: used`
+   - 使用了哪些数据源
+   - 明确排除了哪些会话反馈类型
+4. 若系统无法保证隔离，必须写 `blocked`，不得输出伪盲推。
+
+### 0.3 具体事件选项协议
+
+校时、应期、转型、婚恋、事业机会、财富事件不得只问用户"你感觉像不像"或"有没有变化"。
+
+必须改成可选择的具体事件格式：
+
+```markdown
+| 时间窗口 | 事件选项 | 技术来源 | 用途 |
+|---|---|---|---|
+| YYYY-MM-DD 到 YYYY-MM-DD | A. 搬家/换住处；B. 工作启动；C. 关系断联 | Dasha + Transit + Varga | 校验/预测/置信度调整 |
+```
+
+要求：
+
+- 每个窗口至少给出 2-4 个可判断事件选项。
+- 事件必须落到现实层面：入职、搬家、签约、付款、断联、公开、培训班更换、项目暂停、住处/通勤变化等。
+- 不得只给心理状态、抽象主题或"能量变化"。
+- 用户选择后，只能用于后续校准；不能倒灌污染"盲推模式"。
+
+### 0.4 证据先行，结论最后
+
+高严谨解盘必须按以下顺序组织，不得先下结论再找理由：
+
+1. 原始 D1 + 三重参考引擎边界
+2. 功能性吉凶星
+3. 行星力量：Shadbala / Vimsopaka / Avastha
+4. Yoga / Dosha
+5. Jaimini：AK / AmK / DK / UL / A10 / Argala
+6. 分盘：按问题域调取 D9 / D10 / D2 / D4 / D7 / D12 / D20 / D24 / D30 / D60
+7. 多大运：Vimshottari / Narayana / Yogini / Kalachakra / Jaimini
+8. Transit + VedAstro range scan
+9. Tajika 年运
+10. MEVG + 真实案例校准
+11. 最后才输出事业、财富、婚恋、健康、迁移等结论
+
+### 0.5 相似案例分层对标
+
+真实案例不得只因为"同主题"就拿来类比。必须给出相似层级：
+
+| 层级 | 含义 | 置信度作用 |
+|---|---|---|
+| L0 同主题 | 都是婚恋/事业/财富案例，但星盘和运势未证明相似 | 只能作背景，不得推断 |
+| L1 D1 结构相似 | Lagna、关键宫主、功能吉凶、行星落宫或尊严有相似点 | 可作弱参考 |
+| L2 分盘相似 | D9/D10/D2 等对应分盘支持相同主题 | 可提高主题置信度 |
+| L3 Dasha 功能相似 | 不是同一颗星也可以，但必须功能等价，如旧系统瓦解、异地网络、长期责任启动 | 可作阶段类比 |
+| L4 事件机制相似 | 真实案例中发生了可验证的节点，如离开旧组织、获得项目、迁移、收入来源变化 | 可校准事件形态 |
+| L5 物质支撑相似 | 收入来源、平台、组织资源、合作关系、家庭/伴侣支持等现实条件相似 | 才能推断"生活会如何变好" |
+
+输出案例时必须说明：
+
+- 这个案例和命主在哪些层相似
+- 哪些层不相似
+- 哪些内容可迁移
+- 哪些内容不可迁移
+- 因不相似而降低了哪部分置信度
+
+### 0.6 反例与用户排除案例处理
+
+- 反例不能静默删除；必须标注它为什么是反例、它压低了哪个判断。
+- 若用户明确说"不要考虑某个案例"，该案例标记为 `user-excluded`，不进入主校准，但保留为审计记录。
+- 正例不能单独决定结论；必须经过分盘、大运、事件机制和 MEVG 仲裁。
+- 单一名人案例只能校准"事件机制"，不得复制成用户个人命运模板。
+
+### 0.7 转型与机会的现实机制检查
+
+当判断事业转型、婚恋落实、财富改善或人生阶段转折时，必须追问并输出"机会如何落地"，至少拆成：
+
+1. 旧系统断裂：离开组织、旧项目停摆、关系断联、身份失效。
+2. 过渡平台：课程、讲座、试稿、短项目、朋友介绍、异地网络、线上发布。
+3. 资源入口：谁提供机会，平台是什么，收入从哪里来。
+4. 责任测试：合同、制作、交付、管理、公开评价、长期责任。
+5. 结果显化：付款、职位、署名、公开关系、稳定合作、可重复收入。
+
+若无法证明"资源入口"或"收入来源"，不得说"情况会变好"，只能说"有转向信号但物质改善待验证"。
+
+### 0.8 时间节点输出粒度
+
+所有预测窗口必须拆成不同层级：
+
+| 节点 | 含义 |
+|---|---|
+| Contact | 初次接触、消息、介绍、线索出现 |
+| Activation | 开始行动、沟通密集、试稿/面试/约见 |
+| Confirmation | 确认合作、确认关系、确认资源 |
+| Manifestation | 公开、到账、签约、入职、搬迁完成 |
+| Responsibility Test | 压力测试、延期、责任加重、现实磨合 |
+
+禁止把这些全部混成一句"会有机会"或"会确定关系"。
 
 ---
 

references/strict-workflow-router.md

@@ -58,6 +58,57 @@ Every Level 2+ chart reading must include `MEVG / Global Web Evidence` and
 
 ---
 
+## 0.7 User-Led Reading Calibration Gate
+
+This gate turns the user's follow-up interrogation into workflow control. It is
+not a personality preference; it prevents vague readings, memory-based
+anchoring, and weak case analogies.
+
+Trigger this gate when the user asks any of the following:
+
+- whether a technique, source layer, oracle, or divisional chart was omitted;
+- for a blind reading that excludes prior life events, feedback, or corrections;
+- for concrete time/event options instead of conceptual questions;
+- for real case comparison, global-web evidence, or a named analog case;
+- why a timing claim was made, especially when a case dasha differs from the
+  native's dasha.
+
+When active, the reading must complete or explicitly block these controls:
+
+1. `User Feedback Isolation`: if blind mode is requested, exclude prior
+   biographical feedback, rectification choices, and user corrections from the
+   interpretive reasoning. Use only birth inputs, current engine data, graded
+   source packs, MEVG packets, and real-case calibration records.
+2. `Concrete Time/Event Options`: for rectification, timing, opportunity, and
+   event prediction, output selectable date windows and concrete event options.
+   Do not ask vague subjective questions such as "does this feel like you?"
+3. `Evidence-First / Conclusion-Last`: show raw chart layers, strength layers,
+   dasha/transit/Tajika layers, MEVG, and real-case calibration before the final
+   fortune judgment.
+4. `Similarity-Weighted Case Calibration`: classify each real case by similarity
+   layer instead of treating same-topic cases as equal evidence.
+5. `Transferability Boundary`: state which part of an analog case transfers,
+   which part does not, and how the mismatch changes confidence.
+6. `Counterexample Handling`: mark counterexamples, user-excluded cases, and
+   obsolete cases explicitly. Do not silently remove inconvenient cases.
+
+### Similarity layers for real cases
+
+| Layer | Meaning | Use |
+|---|---|---|
+| L0 same-domain only | Same broad topic, but no proven chart or dasha similarity | Background only |
+| L1 D1 structure | Lagna, house lords, functional benefic/malefic roles, dignity, or house themes align | Weak support |
+| L2 divisional support | Relevant D9/D10/D2/D4/D7/D12 etc. supports the same theme | Theme confidence |
+| L3 functional dasha | Dasha periods play comparable functions even when the planets differ | Stage comparison |
+| L4 verified event mechanism | Public biography or benchmark record shows a concrete event mechanism | Event-shape calibration |
+| L5 material substrate | Income source, platform, institution, spouse/family support, or network structure is comparable | Required for material-outcome claims |
+
+Analog cases cannot be used as prophecy. They may calibrate mechanism and
+confidence only after the similarity layer and transferability boundary are
+declared.
+
+---
+
 ## 1. Question router
 
 | User intent | Route | Required depth |
@@ -235,6 +286,10 @@ Every Level 2+ output must include this table near the end.
 | Argala | Used / not used | ... | ... |
 | KP / Sub-lord | Used / not required / unavailable | ... | ... |
 | Historical verification | Used / not provided | ... | ... |
+| User Feedback Isolation | Used / not requested / blocked | Blind-mode exclusion of prior biography, rectification feedback, and user corrections | Required when user asks for blind reading; blocked if isolation cannot be guaranteed |
+| Concrete Time/Event Options | Used / not required / blocked | Date windows with selectable concrete event options | Required for rectification and timing calibration; lowers precision if omitted |
+| Similarity-Weighted Case Calibration | Used / blocked / case gap | L0-L5 similarity layer for each analog case | Same-topic cases alone cannot raise confidence |
+| Transferability Boundary | Used / blocked / not applicable | What transfers, what does not, and why | Prevents copying a celebrity or benchmark story into the native's forecast |
 | MEVG / Global Web Evidence | Used / blocked / exempt for pure calculation | Source tier, searches, conflict arbitration | Mandatory for all chart interpretation readings; blocked or downgraded if omitted |
 | Real Case Calibration | Used / blocked / case gap | Real case, benchmark case, or documented lack of matching cases | Cap confidence when no comparable case evidence is available |
 

tests/test_user_led_reading_guidance_protocol.py

@@ -0,0 +1,63 @@
+#!/usr/bin/env python3
+"""Regression tests for user-led Jyotish reading calibration guidance."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+
+ROOT = Path(__file__).resolve().parents[1]
+AI_WORKFLOW = ROOT / "references" / "ai-reading-workflow-prompt.md"
+STRICT_ROUTER = ROOT / "references" / "strict-workflow-router.md"
+
+
+def test_ai_workflow_contains_user_interrogation_calibration_protocol() -> None:
+    text = AI_WORKFLOW.read_text(encoding="utf-8")
+
+    required_phrases = [
+        "用户追问校准与相似案例对标协议",
+        "盲推隔离模式",
+        "具体事件选项协议",
+        "证据先行，结论最后",
+        "相似案例分层对标",
+        "反例与用户排除案例处理",
+        "转型与机会的现实机制检查",
+        "时间节点输出粒度",
+        "User Feedback Isolation: used",
+        "Contact",
+        "Activation",
+        "Confirmation",
+        "Manifestation",
+        "Responsibility Test",
+    ]
+
+    for phrase in required_phrases:
+        assert phrase in text
+
+
+def test_strict_router_requires_user_led_calibration_audit_rows() -> None:
+    text = STRICT_ROUTER.read_text(encoding="utf-8")
+
+    required_phrases = [
+        "User-Led Reading Calibration Gate",
+        "User Feedback Isolation",
+        "Concrete Time/Event Options",
+        "Evidence-First / Conclusion-Last",
+        "Similarity-Weighted Case Calibration",
+        "Transferability Boundary",
+        "Counterexample Handling",
+        "L0 same-domain only",
+        "L5 material substrate",
+        "Analog cases cannot be used as prophecy",
+    ]
+
+    for phrase in required_phrases:
+        assert phrase in text
+
+
+def test_case_calibration_requires_more_than_same_domain() -> None:
+    text = STRICT_ROUTER.read_text(encoding="utf-8")
+
+    assert "Same-topic cases alone cannot raise confidence" in text
+    assert "Required for material-outcome claims" in text
+    assert "Do not ask vague subjective questions" in text