causal-inference（因果推理）

Causal Inference

A lightweight causal layer for predicting action outcomes, not by pattern-matching correlations, but by modeling interventions and counterfactuals.

Core Invariant

Every action must be representable as an explicit intervention on a causal model, with predicted effects + uncertainty + a falsifiable audit trail.

Plans must be *causally valid*, not just plausible.

When to Trigger

Trigger this skill on ANY high-level action, including but not limited to:

Also trigger when:

• Reviewing outcomes — "Did that email get a reply?" → log outcome, update estimates
• Debugging failures — "Why didn't this work?" → trace causal graph
• Backfilling history — "Analyze my past emails/calendar" → parse logs, reconstruct actions
• Planning — "Should I send now or later?" → query causal model

Backfill: Bootstrap from Historical Data

Don't start from zero. Parse existing logs to reconstruct past actions + outcomes.

Email Backfill

# 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

Calendar Backfill

# 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

Message Backfill (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

Generic Backfill Pattern

# 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)

Architecture

A. Action Log (required)

Every executed action emits a structured event:

{
  "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
}

Store in memory/causal/action_log.jsonl.

B. Causal Graphs (per domain)

Start with 10-30 observable variables per domain.

Email domain:

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

Calendar domain:

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

Messaging domain:

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

Task domain:

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

Store graph definitions in memory/causal/graphs/.

C. Estimation

For each "knob" (intervention variable), estimate treatment effects:

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

Use simple regression or propensity matching first. Graduate to do-calculus when graphs are explicit and identification is needed.

D. Decision Policy

Before executing actions:

1. Identify intervention variable(s)
2. Query causal model for expected outcome distribution
3. Compute expected utility + uncertainty bounds
4. If uncertainty > threshold OR expected harm > threshold → refuse or escalate to user
5. Log prediction for later validation

Workflow

On Every Action

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

Planning an Action

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

Debugging a Failure

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

Quick Start: Bootstrap Today

# 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

Safety Constraints

Define "protected variables" that require explicit user approval:

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%

Files

• memory/causal/action_log.jsonl — all logged actions with outcomes
• memory/causal/graphs/ — domain-specific causal graph definitions
• memory/causal/estimates/ — learned treatment effects
• memory/causal/config.yaml — safety thresholds and protected variables

References

• See references/do-calculus.md for formal intervention semantics
• See references/estimation.md for treatment effect estimation methods