因果推断

一个用于预测行动结果的轻量级因果层，其方法并非基于模式匹配的相关性分析，而是通过对干预措施和反事实情境进行建模。

核心不变原则

每个行动都必须能够表示为对因果模型的明确干预，并包含预测效果、不确定性以及可验证的审计轨迹。

计划必须在因果上有效，而不仅仅是看似合理。

触发时机

在任何高层次行动时触发此技能，包括但不限于：

同样在以下情况触发：

• 复盘结果— "那封邮件得到回复了吗？" → 记录结果，更新预估
• 调试失败— "为什么这个没起作用？" → 追踪因果图
• 回填历史— "分析我过去的邮件/日历" → 解析日志，重建操作
• 规划— "我是该现在发送还是稍后发送？" → 查询因果模型

回填：基于历史数据引导启动

不要从零开始。解析现有日志以重建过去的操作和结果。

邮件回填

# Extract sent emails with reply status
gog gmail list --sent --after 2024-01-01 --format json > /tmp/sent_emails.json

# For each sent email, check if reply exists
python3 scripts/backfill_email.py /tmp/sent_emails.json

日历回填

# Extract past events with attendance
gog calendar list --after 2024-01-01 --format json > /tmp/events.json

# Reconstruct: did meeting happen? was it moved? attendee count?
python3 scripts/backfill_calendar.py /tmp/events.json

消息回填（WhatsApp/Discord/Slack）

# Parse message history for send/reply patterns
wacli search --after 2024-01-01 --from me --format json > /tmp/wa_sent.json
python3 scripts/backfill_messages.py /tmp/wa_sent.json

通用回填模式

# For any historical data source:
for record in historical_data:
    action_event = {
        "action": infer_action_type(record),
        "context": extract_context(record),
        "time": record["timestamp"],
        "pre_state": reconstruct_pre_state(record),
        "post_state": extract_post_state(record),
        "outcome": determine_outcome(record),
        "backfilled": True  # Mark as reconstructed
    }
    append_to_log(action_event)

架构

A. 行动日志（必需）

每一个执行的动作都会发出一个结构化事件：

{
  "action": "send_followup",
  "domain": "email",
  "context": {"recipient_type": "warm_lead", "prior_touches": 2},
  "time": "2025-01-26T10:00:00Z",
  "pre_state": {"days_since_last_contact": 7},
  "post_state": {"reply_received": true, "reply_delay_hours": 4},
  "outcome": "positive_reply",
  "outcome_observed_at": "2025-01-26T14:00:00Z",
  "backfilled": false
}

存储在memory/causal/action_log.jsonl文件中。

B. 因果图（按领域）

每个领域从10-30个可观测变量开始。

邮件领域：

send_time → reply_prob
subject_style → open_rate
recipient_type → reply_prob
followup_count → reply_prob (diminishing)
time_since_last → reply_prob

日历领域：

meeting_time → attendance_rate
attendee_count → slip_risk
conflict_degree → reschedule_prob
buffer_time → focus_quality

消息领域：

response_delay → conversation_continuation
message_length → response_length
time_of_day → response_prob
platform → response_delay

任务领域：

due_date_proximity → completion_prob
priority_level → completion_speed
task_size → deferral_risk
context_switches → error_rate

将图定义存储在memory/causal/graphs/目录下。

C. 估计

对于每个"旋钮"（干预变量），估计其处理效应：

# Pseudo: effect of morning vs evening sends
effect = mean(reply_prob | send_time=morning) - mean(reply_prob | send_time=evening)
uncertainty = std_error(effect)

首先使用简单回归或倾向得分匹配。当图结构明确且需要识别时，逐步过渡到do-演算。

D. 决策策略

在执行动作之前：

1. 识别干预变量
2. 查询因果模型以获取预期结果分布
3. 计算期望效用 + 不确定性边界
4. 如果不确定性 > 阈值 或 预期危害 > 阈值 → 拒绝或上报给用户
5. 记录预测以供后续验证

工作流程

针对每个行动

BEFORE executing:
1. Log pre_state
2. If enough historical data: query model for expected outcome
3. If high uncertainty or risk: confirm with user

AFTER executing:
1. Log action + context + time
2. Set reminder to check outcome (if not immediate)

WHEN outcome observed:
1. Update action log with post_state + outcome
2. Re-estimate treatment effects if enough new data

规划一个行动

1. User request → identify candidate actions
2. For each action:
   a. Map to intervention(s) on causal graph
   b. Predict P(outcome | do(action))
   c. Estimate uncertainty
   d. Compute expected utility
3. Rank by expected utility, filter by safety
4. Execute best action, log prediction
5. Observe outcome, update model

调试一次失败

1. Identify failed outcome
2. Trace back through causal graph
3. For each upstream node:
   a. Was the value as expected?
   b. Did the causal link hold?
4. Identify broken link(s)
5. Compute minimal intervention set that would have prevented failure
6. Log counterfactual for learning

快速入门：今日启动

# 1. Create the infrastructure
mkdir -p memory/causal/graphs memory/causal/estimates

# 2. Initialize config
cat > memory/causal/config.yaml << 'EOF'
domains:
  - email
  - calendar
  - messaging
  - tasks

thresholds:
  max_uncertainty: 0.3
  min_expected_utility: 0.1

protected_actions:
  - delete_email
  - cancel_meeting
  - send_to_new_contact
  - financial_transaction
EOF

# 3. Backfill one domain (start with email)
python3 scripts/backfill_email.py

# 4. Estimate initial effects
python3 scripts/estimate_effect.py --treatment send_time --outcome reply_received --values morning,evening

安全约束

定义需要用户明确批准的"受保护变量"：

protected:
  - delete_email
  - cancel_meeting
  - send_to_new_contact
  - financial_transaction

thresholds:
  max_uncertainty: 0.3  # don't act if P(outcome) uncertainty > 30%
  min_expected_utility: 0.1  # don't act if expected gain < 10%

文件

• memory/causal/action_log.jsonl— 所有已记录的行动及其结果
• memory/causal/graphs/— 特定领域的因果图定义
• memory/causal/estimates/— 学习到的处理效应
• memory/causal/config.yaml— 安全阈值和受保护变量

参考资料

• 请参阅references/do-calculus.md以了解形式化干预语义
• 请参阅references/estimation.md以了解治疗效果估计方法