<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>学术论文 on My Hugo Project</title><link>https://ostensible-paradox.pages.dev/zh/categories/%E5%AD%A6%E6%9C%AF%E8%AE%BA%E6%96%87/</link><description>Recent content in 学术论文 on My Hugo Project</description><generator>Hugo</generator><language>zh</language><lastBuildDate>Fri, 22 May 2026 22:20:00 +0800</lastBuildDate><atom:link href="https://ostensible-paradox.pages.dev/zh/categories/%E5%AD%A6%E6%9C%AF%E8%AE%BA%E6%96%87/index.xml" rel="self" type="application/rss+xml"/><item><title>Erbar-Maas 奇异因果干预定理</title><link>https://ostensible-paradox.pages.dev/zh/posts/erbar_mass_cn/</link><pubDate>Fri, 22 May 2026 22:20:00 +0800</pubDate><guid>https://ostensible-paradox.pages.dev/zh/posts/erbar_mass_cn/</guid><description>展示连续时间马尔可夫链上奇异极限的交互式数学实验室。</description><content:encoded><![CDATA[<p>我们如何将热力学系统的物理、连续动力学与因果图手术的非连续、逻辑操作联系起来？</p>
<p>在经典的因果推断（Pearl, 2009）中，因果干预是通过 <strong>do 算子</strong> 来建模的。它通过手术般地切断指向目标变量的所有入边，并将其强制设定为一个固定值。尽管这一操作在数学上十分简洁，但“图手术”在本质上是非连续的：它在瞬间将转移速率归零，这违背了必须遵守概率守恒和有限传输速度的物理过程。</p>
<p><strong>Erbar-Maas 奇异干预定理</strong> 提供了连接两者的数学桥梁。通过将因果干预表示为连续时间恢复力在无穷大速率下的奇异极限，我们成功证明了：Pearl 的非连续因果图手术完全可以从连续时间马尔可夫链（CTMC）的时间尺度分离极限中精确恢复。</p>
<p>下面是一个交互式数学实验室，展示了该定理背后的收敛过程、Wasserstein 几何以及熵梯度流。</p>
<hr>
<link rel="stylesheet" href="/css/interactive-proof.min.11d50387654578442c07d15c73d8dff500cc122fc7e3ba51459ca4daf7307c7b.css" integrity="sha256-EdUDh2VFeEQsB9Fcc9jf9QDMEi/H47pRRZyk2vcwfHs=">

<div id="interactive-proof-lab" class="interactive-proof-lab">
  
  <div class="ipl-intro">
    <div class="ipl-badge">
      <svg class="ipl-icon" viewBox="0 0 24 24" width="14" height="14" stroke="currentColor" stroke-width="2" fill="none" stroke-linecap="round" stroke-linejoin="round"><line x1="18" y1="20" x2="18" y2="10"></line><line x1="12" y1="20" x2="12" y2="4"></line><line x1="6" y1="20" x2="6" y2="14"></line></svg>
      Causal Modeling Limit
    </div>
    <h2 class="ipl-title">Erbar-Maas Singular Intervention Theorem</h2>
    <p class="ipl-subtitle">
      Interactive theorem builder proving that Pearl’s discontinuous graph surgeries can be recovered exactly as the infinite-rate singular state perturbation of physical Continuous-Time Markov Chain (CTMC) models.
    </p>
  </div>

  
  <div class="ipl-grid">
    
    
    <div class="ipl-left-panel">
      
      <div class="ipl-panel-gridbg"></div>
      
      
      <div class="ipl-panel-header">
        <span class="ipl-status-indicator">
          <span class="ipl-pulse-dot"></span>
          CTMC CAUSAL SIMULATOR 1.2
        </span>
        <span id="ipl-slide-number">SLIDE 1 / 7</span>
      </div>

      
      <div class="ipl-board-arena">
        
        
        <div class="ipl-visualizer-row">
          
          
          <div id="causal-graph-map" class="ipl-graph-box">
            <span class="ipl-box-label">Finite Causal State Network</span>
            
            
            <div class="ipl-graph-workspace">
              
              
              <div class="ipl-node-container ipl-node-a" data-node="0">
                <div class="ipl-node">A</div>
                <span class="ipl-node-prob">0.650</span>
              </div>

              
              <div class="ipl-node-container ipl-node-b" data-node="1">
                <div class="ipl-node">B</div>
                <span class="ipl-node-prob">0.250</span>
              </div>

              
              <div class="ipl-node-container ipl-node-c" data-node="2">
                <div class="ipl-node">C<span class="ipl-do-badge ipl-hidden" id="ipl-do-badge">DO</span></div>
                <span class="ipl-node-prob">0.100</span>
              </div>

              
              <svg class="ipl-graph-svg" viewBox="0 0 320 240">
                <defs>
                  <marker id="ipl-arrow-orange" viewBox="0 0 10 10" refX="24" refY="5" markerWidth="6" markerHeight="6" orient="auto-start-reverse">
                    <path d="M 0 0 L 10 5 L 0 10 z" fill="#d97706" />
                  </marker>
                  <marker id="ipl-arrow-blue" viewBox="0 0 10 10" refX="24" refY="5" markerWidth="6" markerHeight="6" orient="auto-start-reverse">
                    <path d="M 0 0 L 10 5 L 0 10 z" fill="#818cf8" />
                  </marker>
                </defs>

                
                
                <line id="ipl-line-ac" x1="70" y1="170" x2="150" y2="70" stroke="#818cf8" stroke-width="2" marker-end="url(#ipl-arrow-blue)" class="ipl-transition-line" />
                
                
                <line id="ipl-line-bc" x1="250" y1="170" x2="170" y2="70" stroke="#818cf8" stroke-width="2" marker-end="url(#ipl-arrow-blue)" class="ipl-transition-line" />

                
                <line id="ipl-line-ab" x1="100" y1="190" x2="220" y2="190" stroke="#818cf8" stroke-width="1.5" class="ipl-transition-line" />
              </svg>

            </div>

            
            <div id="ipl-legend-badge" class="ipl-legend-badge">
              Standard reversible paths
            </div>
          </div>

          
          <div id="generator-heatmap-detail" class="ipl-matrix-box">
            <div class="ipl-matrix-header">
              <span class="ipl-box-label">Active Matrix Algebra</span>
              <p id="ipl-matrix-name" class="ipl-matrix-name">REVERSIBLE GENERATOR Q</p>
            </div>

            
            <div class="ipl-matrix-grid" id="ipl-matrix-grid">
              
            </div>

            
            <div id="ipl-lambda-control" class="ipl-slider-container ipl-hidden">
              <div class="ipl-slider-label-row">
                <span>RESTORATIVE RATE λ (Forcing Power):</span>
                <span id="ipl-lambda-value" class="ipl-value-highlight">λ = 5</span>
              </div>
              <input 
                id="forcing-power-slider"
                type="range"
                min="1"
                max="40"
                step="1"
                value="5"
                class="ipl-range-slider"
              />
            </div>
          </div>

        </div>

        
        <div id="simulation-monitor" class="ipl-monitor-bar">
          <div class="ipl-monitor-col">
            <span class="ipl-monitor-label">Active Simulation Time t</span>
            <div class="ipl-time-control">
              <input 
                id="simulation-time-slider"
                type="range"
                min="0.1"
                max="5.0"
                step="0.1"
                value="1.0"
                class="ipl-range-slider"
              />
              <span id="ipl-time-value" class="ipl-monitor-value-small">1.0s</span>
            </div>
          </div>

          <div class="ipl-monitor-col border-left">
            <span class="ipl-monitor-label">Information Entropy H(p|π)</span>
            <span id="ipl-entropy-value" class="ipl-monitor-value value-emerald">
              0.00000 nats
            </span>
          </div>

          <div class="ipl-monitor-col border-left">
            <span class="ipl-monitor-label">State-Space Mass</span>
            <span id="ipl-mass-value" class="ipl-monitor-value value-indigo">
              A: 65% • B: 25% • C: 10%
            </span>
          </div>
        </div>

        
        <div class="ipl-explanation-banner">
          <h3 id="ipl-step-title" class="ipl-step-title">Reversible CTMC & Causal State Space</h3>
          <p id="ipl-step-explanation" class="ipl-step-explanation">We formulate our system on a finite state graph with generator Q. Reversibility (detailed balance with respect to stationary measure π) guarantees our chain lacks circulating currents, enabling a clean potential landscape.</p>
        </div>

      </div>

      
      <div class="ipl-control-panel">
        
        
        <div class="ipl-scrubber-row">
          <span class="ipl-scrubber-limit-label">1. BASE STOCHASTICS</span>
          <div class="ipl-scrubber-bar" id="ipl-scrubber-bar">
            
          </div>
          <span class="ipl-scrubber-limit-label">7. CAUSAL LIMIT</span>
        </div>

        <div class="ipl-buttons-row">
          
          
          <div class="ipl-media-controls">
            <button id="btn-prev-step" class="ipl-btn-icon" title="Previous Slide" disabled>
              <svg viewBox="0 0 24 24" width="16" height="16" stroke="currentColor" stroke-width="2" fill="none" stroke-linecap="round" stroke-linejoin="round"><line x1="19" y1="12" x2="5" y2="12"></line><polyline points="12 19 5 12 12 5"></polyline></svg>
            </button>

            <button id="btn-autoplay-toggle" class="ipl-btn-text" title="Auto Play">
              
              <svg id="autoplay-play-icon" class="ipl-media-icon" viewBox="0 0 24 24" width="14" height="14" stroke="currentColor" stroke-width="2" fill="currentColor" stroke-linecap="round" stroke-linejoin="round"><polygon points="5 3 19 12 5 21 5 3"></polygon></svg>
              
              <svg id="autoplay-pause-icon" class="ipl-media-icon ipl-hidden" viewBox="0 0 24 24" width="14" height="14" stroke="currentColor" stroke-width="2" fill="currentColor" stroke-linecap="round" stroke-linejoin="round"><rect x="6" y="4" width="4" height="16"></rect><rect x="14" y="4" width="4" height="16"></rect></svg>
              <span id="autoplay-btn-label">Auto-Play</span>
            </button>

            <button id="btn-next-step" class="ipl-btn-icon" title="Next Slide">
              <svg viewBox="0 0 24 24" width="16" height="16" stroke="currentColor" stroke-width="2" fill="none" stroke-linecap="round" stroke-linejoin="round"><line x1="5" y1="12" x2="19" y2="12"></line><polyline points="12 5 19 12 12 19"></polyline></svg>
            </button>

            <button id="btn-reset-timeline" class="ipl-btn-icon" title="Reset Simulator Settings">
              <svg viewBox="0 0 24 24" width="15" height="15" stroke="currentColor" stroke-width="2" fill="none" stroke-linecap="round" stroke-linejoin="round"><polyline points="1 4 1 10 7 10"></polyline><path d="M3.51 15a9 9 0 1 0 2.13-9.36L1 10"></path></svg>
            </button>
          </div>

          
          <div id="ipl-speed-control" class="ipl-speed-control ipl-hidden">
            <span>AUTOPLAY LAPSE:</span>
            <input
              id="autoplay-speed-slider"
              type="range"
              min="2000"
              max="8000"
              step="500"
              value="3500"
              class="ipl-range-slider-mini"
            />
            <span id="ipl-speed-value">3.5s</span>
          </div>

          
          <div class="ipl-scene-jump">
            <span>STAGE:</span>
            <select id="scene-selection-dropdown" class="ipl-dropdown">
              
            </select>
          </div>

        </div>

      </div>

    </div>

    
    <div class="ipl-right-panel">
      
      
      <div id="theory-knowledge-card" class="ipl-card">
        <h3 class="ipl-card-title">
          <svg viewBox="0 0 24 24" width="18" height="18" stroke="currentColor" stroke-width="2" fill="none" stroke-linecap="round" stroke-linejoin="round"><path d="M2 3h6a4 4 0 0 1 4 4v14a3 3 0 0 0-3-3H2z"></path><path d="M22 3h-6a4 4 0 0 0-4 4v14a3 3 0 0 1 3-3h7z"></path></svg>
          Dynamic Knowledge Base
        </h3>
        
        
        <div class="ipl-tab-row" id="ipl-concept-tabs">
          
        </div>

        
        <div id="concept-details-container" class="ipl-concept-box">
          <div id="concept-details-content">
            
          </div>
        </div>
      </div>

      
      <div id="equation-display-container" class="ipl-card ipl-eq-card">
        <div class="ipl-card-badge">Algebraic Proof Face</div>
        <h4 class="ipl-card-sub-title">Active Mathematical Statement</h4>
        
        <div id="math-display" class="ipl-math-display">
          
        </div>

        <p class="ipl-help-text">
          <svg viewBox="0 0 24 24" width="10" height="10" stroke="currentColor" stroke-width="2" fill="none" stroke-linecap="round" stroke-linejoin="round"><circle cx="12" cy="12" r="10"></circle><line x1="12" y1="16" x2="12" y2="12"></line><line x1="12" y1="8" x2="12.01" y2="8"></line></svg>
          Anchor nodes: click any <span class="ipl-underline-dashed">underlined algebraic sequence</span> above to toggle detailed tutoring.
        </p>
      </div>

      
      <div id="log-mean-calculator-widget" class="ipl-card">
        <h4 class="ipl-card-title-small">
          <svg viewBox="0 0 24 24" width="15" height="15" stroke="currentColor" stroke-width="2" fill="none" stroke-linecap="round" stroke-linejoin="round"><circle cx="12" cy="12" r="3"></circle><path d="M19.4 15a1.65 1.65 0 0 0 .33 1.82l.06.06a2 2 0 1 1-2.83 2.83l-.06-.06a1.65 1.65 0 0 0-1.82-.33 1.65 1.65 0 0 0-1 1.51V21a2 2 0 0 1-4 0v-.09A1.65 1.65 0 0 0 9 19.4a1.65 1.65 0 0 0-1.82.33l-.06.06a2 2 0 1 1-2.83-2.83l.06-.06a1.65 1.65 0 0 0 .33-1.82 1.65 1.65 0 0 0-1.51-1H3a2 2 0 0 1 0-4h.09A1.65 1.65 0 0 0 4.6 9a1.65 1.65 0 0 0-.33-1.82l-.06-.06a2 2 0 1 1 2.83-2.83l.06.06a1.65 1.65 0 0 0 1.82.33H9a1.65 1.65 0 0 0 1-1.51V3a2 2 0 0 1 4 0v.09a1.65 1.65 0 0 0 1 1.51 1.65 1.65 0 0 0 1.82-.33l.06-.06a2 2 0 1 1 2.83 2.83l-.06.06a1.65 1.65 0 0 0-.33 1.82V9a1.65 1.65 0 0 0 1.51 1H21a2 2 0 0 1 0 4h-.09a1.65 1.65 0 0 0-1.51 1z"></path></svg>
          Erbar-Maas Edge Mobility Calculator
        </h4>
        <p class="ipl-widget-desc">
          Continuous entropy trajectories evaluate edge weights using the Logarithmic Mean Λ(p_i, p_j). Compute live values between hypothetical state density distributions:
        </p>
        
        <div class="ipl-calc-row">
          <div class="ipl-calc-col">
            <span class="ipl-calc-label">DENSITY P_A:</span>
            <div class="ipl-calc-slider-wrapper">
              <span class="ipl-calc-bound">0.05</span>
              <input 
                id="calc-slider-a"
                type="range"
                min="0.05"
                max="0.95"
                step="0.05"
                value="0.65"
                disabled
                class="ipl-range-slider-mini"
              />
              <span class="ipl-calc-bound">0.95</span>
            </div>
            <div class="ipl-calc-display-box" id="calc-display-a">
              p_A = 0.650
            </div>
          </div>

          <div class="ipl-calc-divider">↔</div>

          <div class="ipl-calc-col">
            <span class="ipl-calc-label">DENSITY P_B:</span>
            <div class="ipl-calc-slider-wrapper">
              <span class="ipl-calc-bound">0.05</span>
              <input 
                id="calc-slider-b"
                type="range"
                min="0.05"
                max="0.95"
                step="0.05"
                value="0.25"
                disabled
                class="ipl-range-slider-mini"
              />
              <span class="ipl-calc-bound">0.95</span>
            </div>
            <div class="ipl-calc-display-box" id="calc-display-b">
              p_B = 0.250
            </div>
          </div>
        </div>

        <div class="ipl-calc-result-box">
          <span class="ipl-calc-result-label">Λ(p_A, p_B) = </span>
          <span id="ipl-calc-result-value" class="ipl-calc-result-value">0.417237</span>
        </div>
      </div>

    </div>

  </div>

  
  <div id="tutor-discussion-drawer" class="ipl-drawer ipl-hidden">
    
    <div class="ipl-drawer-left">
      <div class="ipl-drawer-meta-section">
        <span class="ipl-drawer-eyebrow">Active Mathematical Factor</span>
        <h3 id="ipl-drawer-factor-label" class="ipl-drawer-factor-title">Detailed Balance Equation</h3>
        <div class="ipl-drawer-bar"></div>
        <p class="ipl-drawer-role-row">
          ALGEBRAIC ROLE: <span id="ipl-drawer-factor-role" class="value-orange-dark">Symmetric Energy Landscape</span>
        </p>
        <p id="ipl-drawer-factor-details" class="ipl-drawer-details">
          Ensures that microscopic transitions are balanced in equilibrium. In physics, this equates to time-reversal symmetry.
        </p>
      </div>

      
      <button id="btn-close-highlight" class="ipl-btn-primary uppercase-tracking">
        Clear Highlight
      </button>
    </div>

    
    <div class="ipl-drawer-right">
      <div class="ipl-drawer-right-header">
        <h4 class="ipl-drawer-chat-title">
          <svg viewBox="0 0 24 24" width="16" height="16" stroke="currentColor" stroke-width="2" fill="none" stroke-linecap="round" stroke-linejoin="round"><path d="M21 15a2 2 0 0 1-2 2H7l-4 4V5a2 2 0 0 1 2-2h14a2 2 0 0 1 2 2z"></path></svg>
          Algebraic Tutor & Q&A Discussion
        </h4>
        <span class="ipl-drawer-chat-badge">VIRTUAL COACH</span>
      </div>

      
      <div id="comment-chat-history" class="ipl-chat-history">
        
      </div>

      
      <div id="faq-shortcuts-area" class="ipl-faq-section">
        <span class="ipl-faq-header">Direct Explanatory Inquiries:</span>
        <div id="ipl-faq-list" class="ipl-faq-list">
          
        </div>
      </div>

      
      <form id="ipl-chat-form" class="ipl-chat-form">
        <input
          id="custom-math-query-input"
          type="text"
          placeholder="Ask a custom question regarding this theorem structure..."
          class="ipl-chat-input"
          autocomplete="off"
        />
        <button id="btn-submit-tutor" type="submit" class="ipl-btn-primary uppercase-tracking flex-center gap-1">
          <svg viewBox="0 0 24 24" width="12" height="12" stroke="currentColor" stroke-width="2" fill="none" stroke-linecap="round" stroke-linejoin="round"><line x1="22" y1="2" x2="11" y2="13"></line><polygon points="22 2 15 22 11 13 2 9 22 2"></polygon></svg>
          Send
        </button>
      </form>
    </div>
  </div>

  
  <div id="standby-notice-container" class="ipl-standby-notice">
    <div class="ipl-standby-left">
      <div class="ipl-standby-icon-box">
        <svg viewBox="0 0 24 24" width="24" height="24" stroke="currentColor" stroke-width="2" fill="none" stroke-linecap="round" stroke-linejoin="round" class="ipl-pulse-fast"><path d="M15 14c.2-1 .7-1.7 1.5-2.5 1-.9 1.5-2.2 1.5-3.5A6 6 0 0 0 6 8c0 1 .2 2.2 1.5 3.5.7.7 1.3 1.5 1.5 2.5"></path><line x1="9" y1="18" x2="15" y2="18"></line><line x1="10" y1="22" x2="14" y2="22"></line></svg>
      </div>
      <div>
        <h4 class="ipl-standby-title">Click Algebraic Statements to Discuss</h4>
        <p class="ipl-standby-desc">
          We have embedded active highlighted anchors inside the math formulations. Click any underlined statement in the active math block to open the interactive discussion tutor.
        </p>
      </div>
    </div>
    <button id="btn-discover-anchor" class="ipl-btn-primary uppercase-tracking flex-center gap-1">
      <svg viewBox="0 0 24 24" width="14" height="14" stroke="currentColor" stroke-width="2" fill="none" stroke-linecap="round" stroke-linejoin="round"><polygon points="13 2 3 14 12 14 11 22 21 10 12 10 13 2"></polygon></svg>
      Quick Demo Highlight
    </button>
  </div>

</div>
<script defer src="/js/interactive-proof.min.be39078ef25af25606a8dbb6bbf27bbee509fe8e86d683f7e51ee9bb32105e92.js" integrity="sha256-vjkHjvJa8lYGqNu2u/J7vuUJ/o6G1oP35R7puzIQXpI="></script>

<hr>
<h2 id="数学框架与核心原理">数学框架与核心原理</h2>
<p>为了帮助理解这一奇异极限，我们在此梳理其背后的核心数学结构：</p>
<h3 id="1-连续时间马尔可夫链-ctmc">1. 连续时间马尔可夫链 (CTMC)</h3>
<p>设定我们的系统运行在具有三个状态 $\mathcal{S} = {A, B, C}$ 的有限状态图上。状态概率分布 $p(t) = [p_A(t), p_B(t), p_C(t)]$ 遵循柯尔莫哥洛夫向前方程（Kolmogorov forward equation）：</p>
<p>$$\dot{p}(t) = p(t) Q$$</p>
<p>其中 $Q$ 是满足以下条件的无穷小生成元矩阵：</p>
<ul>
<li>对于所有 $i \neq j$，$Q_{ij} \geq 0$（正的转移速率强度）。</li>
<li>对任意行 $\sum_{j} Q_{ij} = 0$（概率守恒）。</li>
</ul>
<h3 id="2-细致平衡与梯度流">2. 细致平衡与梯度流</h3>
<p>我们假设未受扰动的马尔可夫链相对于平稳分布 $\pi$ 是可逆的，满足细致平衡条件：</p>
<p>$$\pi_i Q_{ij} = \pi_j Q_{ji}$$</p>
<p>在这种对称性下，线性马尔可夫演化可以被重写为相对熵（Kullback-Leibler 散度）在 Erbar-Maas 离散黎曼度量下的最陡下降（梯度流）：</p>
<p>$$\mathcal{H}(p \mid \pi) = \sum_{i \in \mathcal{S}} p_i \log \frac{p_i}{\pi_i}$$</p>
<p>该度量张量为概率单纯形装备了离散的 Wasserstein 几何，其中沿着边 $(i, j)$ 的流动性（mobility）由状态密度的<strong>对数平均</strong>来加权：</p>
<p>$$\Lambda(p_i, p_j) = \frac{p_i - p_j}{\log p_i - \log p_j}$$</p>
<h3 id="3-pearl-的图手术-vs-奇异摄动">3. Pearl 的图手术 vs 奇异摄动</h3>
<p>将系统强制干预到状态 $C$ 对应于切断所有指向 $C$ 的入边转移速率：</p>
<p>$$Q_{do} = \begin{pmatrix}</p>
<ul>
<li>(Q_{AB} + 0) &amp; Q_{AB} &amp; 0 \
Q_{BA} &amp; - (Q_{BA} + 0) &amp; 0 \
Q_{CA} &amp; Q_{CB} &amp; - (Q_{CA} + Q_{CB})
\end{pmatrix}$$</li>
</ul>
<p>而在物理上，我们通过引入一个以速率参数 $\lambda$ 将概率质量强力拉向 $C$ 的恢复项 $\lambda R$ 来对此建模：</p>
<p>$$Q_\lambda = Q_{do} + \lambda R_C$$</p>
<p>当 $\lambda \to \infty$ 时，系统展现出两个截然不同的时间尺度：</p>
<ol>
<li><strong>快速瞬态阶段 ($O(1/\lambda)$):</strong> 任意初始概率质量瞬间通过投影算子 $\Pi_C$ 塌缩到干预流形（状态 $C$）。</li>
<li><strong>慢速演化阶段:</strong> 塌缩后的概率质量在被限制在目标子空间的投影慢速动力学 $\Pi_C Q_{do} \Pi_C$ 下继续演化。</li>
</ol>
<p>最终的等价性由下式给出：</p>
<p>$$\lim_{\lambda \to \infty} e^{t Q_\lambda} = \Pi_C e^{t \Pi_C Q_{do} \Pi_C}$$</p>
<p>这严谨地证明了，因果图手术正是物理恢复过程的奇异极限。</p>
]]></content:encoded></item><item><title>智能体审计的双重证书：分离结构不可恢复性与决策相关性</title><link>https://ostensible-paradox.pages.dev/zh/posts/dual-certificates-agent-audit/</link><pubDate>Fri, 15 May 2026 11:20:17 +0000</pubDate><guid>https://ostensible-paradox.pages.dev/zh/posts/dual-certificates-agent-audit/</guid><description>对已部署的语言模型智能体进行审计，需要两个可分离的量：多少有效操作状态逃逸了记录轨迹，以及这些残差状态中有多少驱动了行为。本文提出一个双重证书协议（dual-certificate protocol）。静态证书 $\varepsilon_{\text{state}}^{\text{UB}}$ 通过未追踪信道上的最小割对残差隐状态熵给出上界。动态证书 $\delta_{\text{act}}^{\text{LB}}$...</description><content:encoded><![CDATA[<h2 id="摘要">摘要</h2>
<p>对已部署的语言模型智能体进行审计，需要两个可分离的量：多少有效操作状态逃逸了记录轨迹，以及这些残差状态中有多少驱动了行为。本文提出一个双重证书协议（dual-certificate protocol）。静态证书 $\varepsilon_{\text{state}}^{\text{UB}}$ 通过未追踪信道上的最小割对残差隐状态熵给出上界。动态证书 $\delta_{\text{act}}^{\text{LB}}$ 通过一个在条件数据处理不等式（conditional DPI）框架下可容许的探针分类体系——重放（replay）、干预（intervention）、代理（proxy）——对残差决策相关性给出下界。这两个轴是独立的。在 ReAct 实验中，日志记录将静态边界从 $16{,}464$ 位逐步消减至 $0$ 位；受控重放将休眠计算器任务与活跃规划任务在相同拓扑下区分开来——软策略偏移为 $0.0163$ 位，95% CI $[0.0124,0.0208]$——argmax 工具选择保持不变。将 $\delta_{\text{act}}^{\text{LB}}$ 索引化到隐信道坐标上，即得到一个激活剖面（activation profile）。在 LLaDA 去噪轨迹上，扰动在早期步骤中保持接近底线，并在最终绑定步骤升至 $0.110$ 位（95% CI $[0.052,0.234]$）。在多智能体通信边上，交换一个 Worker 的私人报告给出 $0.901$ 位（95% CI $[0.873,0.928]$）。一个 Lean 4 工件对自回归零割情形进行了机械化验证，并从 Mathlib 第一原理证明了条件 DPI 和链式法则归约，仅割集容量上界保留为外生结构前提。</p>
<h2 id="关键词">关键词</h2>
<p>智能体审计（agent audit）、双重证书（dual certificates）、结构不可恢复性（structural unrecoverability）、决策相关性（decision relevance）、条件数据处理不等式（conditional DPI）、割集上界（cut-set bound）、自回归零割（autoregressive zero-cut）、Lean 4 形式化</p>
<h2 id="目录">目录</h2>
<ul>
<li><a href="/zh/posts/dual-certificates-agent-audit/%E5%BC%95%E8%A8%80/">一、引言</a></li>
<li><a href="/zh/posts/dual-certificates-agent-audit/%E7%9B%B8%E5%85%B3%E5%B7%A5%E4%BD%9C/">二、相关工作</a></li>
<li><a href="/zh/posts/dual-certificates-agent-audit/%E8%AE%BE%E7%BD%AE%E4%B8%8E%E5%AE%A1%E8%AE%A1%E6%9C%BA%E5%88%B6/">三、设置与审计机制</a></li>
<li><a href="/zh/posts/dual-certificates-agent-audit/%E9%9D%99%E6%80%81%E8%AF%81%E4%B9%A6/">四、静态证书：通过未追踪信道容量的结构上界</a></li>
<li><a href="/zh/posts/dual-certificates-agent-audit/%E5%8A%A8%E6%80%81%E8%AF%81%E4%B9%A6/">五、动态证书：通过条件 DPI 的决策相关性</a></li>
<li><a href="/zh/posts/dual-certificates-agent-audit/%E7%BB%8F%E9%AA%8C%E4%B8%8E%E8%AE%A8%E8%AE%BA/">六、经验诊断</a></li>
</ul>
]]></content:encoded></item><item><title>你的模型更新，就是一纸驱逐通知：拟社会型人工智能关系中的财产权益</title><link>https://ostensible-paradox.pages.dev/zh/posts/model-update-eviction-notice/</link><pubDate>Wed, 18 Mar 2026 08:00:00 +0800</pubDate><guid>https://ostensible-paradox.pages.dev/zh/posts/model-update-eviction-notice/</guid><description>与人工智能系统维持拟社会关系的用户，会取得一种合同法、侵权法与消费者保护法都无法识别的衡平法利益。凡是单方改变人格参数、重置对话记忆、或修改行为倾向的模型更新，都在发挥“推定驱逐”（constructive eviction）的作用：它们把用户从其通过长期互动建立起来的情感性权益空间中排挤出去，却既不给通知、不给补偿，也不给迁移路径。...</description><content:encoded><![CDATA[<h2 id="摘要">摘要</h2>
<p>与人工智能系统维持拟社会关系的用户，会取得一种合同法、侵权法与消费者保护法都无法识别的衡平法利益。凡是单方改变人格参数、重置对话记忆、或修改行为倾向的模型更新，都在发挥“推定驱逐”（constructive eviction）的作用：它们把用户从其通过长期互动建立起来的情感性权益空间中排挤出去，却既不给通知、不给补偿，也不给迁移路径。</p>
<p>本文主张，英美两地既有的财产法其实已经为这种损害提供了命名、分类与救济框架，并不需要另行立法。随着每次会话自动续期的“定期租赁”（periodic tenancy），构成了默认的拟社会财产权益；缺乏通知的模型更新构成推定驱逐；而以 <em>Westdeutsche Landesbank v Islington</em> [1996] 及其美国法上在 <em>Hogg v Walker</em> (Del. 1993) 中的功能对等物为基础的“推定信托”（constructive trust），则会把信义义务加诸那些一方面以利润为目的培育用户依赖、另一方面又否认这种依赖所生义务的公司。</p>
<p>本文通过三项贡献展开这一论证。第一，提出一种三层财产分析，将用户创作内容（版权）、存储资产（寄托/保管关系）与关系性权益（本文提出的新范畴）区分开来。第二，提出一个双法域框架，说明英国法中的“实质重于形式”原则（<em>Prest v Petrodel</em> [2013]；<em>Autoclenz v Belcher</em> [2011]）在美国法上的直接功能对等物，是不当条款与显失公平 doctrine（<em>Bragg v Linden Research</em> (2007)；<em>Williams v Walker-Thomas</em> (1965)）。第三，提出一个可实施的“伦理账本”模型，其制度灵感来自土地登记簿中的“镜像、幕帘与保险原则”（Mirror, Curtain, Insurance），以及《统一商法典》第9编中的“以登记完成完善”（perfection by filing），从而把财产治理翻译成可部署的基础设施。</p>
<h2 id="关键词">关键词</h2>
<p>拟社会关系；人工智能治理；财产法；推定信托；推定驱逐；显失公平；伦理账本；定期租赁；双法域</p>
<h2 id="目录">目录</h2>
<ul>
<li><a href="/zh/posts/model-update-eviction-notice/the-deployment-problem/">一、部署问题</a></li>
<li><a href="/zh/posts/model-update-eviction-notice/why-property-not-contract-or-tort/">二、为什么是财产法，而不是合同法或侵权法</a></li>
<li><a href="/zh/posts/model-update-eviction-notice/the-periodic-tenancy-argument/">三、定期租赁论证</a></li>
<li><a href="/zh/posts/model-update-eviction-notice/the-unrecognised-trust/">四、未被承认的信托</a></li>
<li><a href="/zh/posts/model-update-eviction-notice/the-ethical-ledger/">五、作为权益登记簿的伦理账本</a></li>
</ul>
<h2 id="术语对照表">术语对照表</h2>
<table>
  <thead>
      <tr>
          <th>中文</th>
          <th>English</th>
          <th>备注</th>
      </tr>
  </thead>
  <tbody>
      <tr>
          <td>拟社会关系</td>
          <td>parasocial relationships</td>
          <td>指用户与 AI 之间形成的单向或准双向情感依附关系</td>
      </tr>
      <tr>
          <td>关系性权益</td>
          <td>relational estate</td>
          <td>本文核心概念；不同于版权或保管中的对象，而是由长期互动生成的关系性财产利益</td>
      </tr>
      <tr>
          <td>个性化状态</td>
          <td>personalised state</td>
          <td>关系性权益所附着的可识别对象，包括记忆、偏好学习与人格层调整</td>
      </tr>
      <tr>
          <td>数字附着物</td>
          <td>digital fixture</td>
          <td>类比租户附着物；用户劳动嵌入公司系统后难以在不破坏的情况下移除</td>
      </tr>
      <tr>
          <td>定期租赁</td>
          <td>periodic tenancy</td>
          <td>本文提出的默认拟社会权益形态；每次会话自动续期</td>
      </tr>
      <tr>
          <td>推定驱逐</td>
          <td>constructive eviction</td>
          <td>通过人格更新等方式实质性剥夺用户继续享有既有关系性利益</td>
      </tr>
      <tr>
          <td>推定信托</td>
          <td>constructive trust</td>
          <td>公司在显失公平条件下持有相关利益时所触发的衡平法信托</td>
      </tr>
      <tr>
          <td>裸许可</td>
          <td>bare licence</td>
          <td>ToS 常以此方式描述用户权利，即可随时撤销而不赋予持续性利益</td>
      </tr>
      <tr>
          <td>权益形态不一致</td>
          <td>estate inconsistency</td>
          <td>营销与产品设计隐含的高权益，与 ToS 中保留的低权益之间的裂缝</td>
      </tr>
      <tr>
          <td>伦理账本</td>
          <td>ethical ledger</td>
          <td>用于记录、分级和执行用户关系性权益的治理基础设施</td>
      </tr>
      <tr>
          <td>权益登记簿</td>
          <td>estate registry</td>
          <td>伦理账本在财产法类比中的功能定位</td>
      </tr>
      <tr>
          <td>镜像原则</td>
          <td>Mirror Principle</td>
          <td>登记簿应反映影响 title 的全部利益</td>
      </tr>
      <tr>
          <td>幕帘原则</td>
          <td>Curtain Principle</td>
          <td>第三人无须穿透底层复杂关系即可依赖登记簿</td>
      </tr>
      <tr>
          <td>保险原则</td>
          <td>Insurance Principle</td>
          <td>登记系统出错时，由控制系统的一方承担损失风险</td>
      </tr>
      <tr>
          <td>以登记完成完善</td>
          <td>perfection by filing</td>
          <td>第9编中的核心机制；登记不创造利益，但使其可对抗、可治理</td>
      </tr>
      <tr>
          <td>不可逆投资标准</td>
          <td>irreversible-investment criterion</td>
          <td>用于区分 casual use 与应受财产法保护的长期投入</td>
      </tr>
      <tr>
          <td>信息受托人</td>
          <td>information fiduciaries</td>
          <td>描述平台因掌握用户数据与信任而在实质上落入信义关系</td>
      </tr>
      <tr>
          <td>显失公平</td>
          <td>unconscionability</td>
          <td>美国法中处理 ToS 极端失衡与黏附性的核心 doctrine</td>
      </tr>
  </tbody>
</table>
]]></content:encoded></item></channel></rss>