rsn-learning-outcomes（rsn 学习成果）

Learning

Reads

Writes

Systematic improvement from experience. Convert outcomes into better future performance.

Process

1. [K] Detect learning trigger — Gap detected, experience completed, belief needs testing, or predictions off
2. [K] Select learning mode — Choose single-loop, double-loop, reflection, experimentation, or calibration
3. [S] Execute mode process — Follow mode-specific workflow systematically
4. [K] Extract insights — Identify transferable patterns and updated beliefs
5. [R] Validate learning — Confirm insights are actionable and conditions-bounded

Evaluation methods for [R] steps:

• Pattern validation — Require 3+ instances before generalizing (single/double-loop)
• Insight transferability — Verify conditions when insight applies (reflection)
• Experimental rigor — Check falsifiability and success criteria (experimentation)
• Calibration accuracy — Require 30+ predictions for meaningful adjustment (calibration)

Boundaries

In scope:

• Correcting actions (single-loop)
• Questioning frames (double-loop)
• Extracting insights from experience (reflection)
• Testing beliefs through experiments
• Adjusting prediction confidence (calibration)
• Creating learning artifacts (heuristics, playbooks, checklists)

Out of scope:

• Executing corrected actions (use rsn-reasoning-problems.causal)
• Gathering information to inform learning (use rsn-perceiving-information)
• Creative problem-solving for new situations (use rsn-creating-ideas)
• Deep reasoning with new frames (use rsn-reasoning-problems)

Core Principle

Learning is not automatic. Experience without reflection is just repetition. Learning requires deliberate extraction of insight and updating of beliefs and behaviors.

Experience → Extract → Update → Apply → Better Outcomes

Mode Selection

Decision Tree

Is there a gap between expected and actual?
  YES → Is this a pattern (3+ similar failures)?
    YES → Double-loop (question the frame)
    NO  → Single-loop (fix the action)
  NO  ↓
Has an experience completed?
  YES → Reflection (extract insights)
  NO  ↓
Do you have a belief that needs validation before commitment?
  YES → Experimentation (test the belief)
  NO  ↓
Have predictions been consistently off?
  YES → Calibration (adjust confidence)
  NO  → No learning mode needed

Mode Summaries

Single-Loop

Purpose: Correct action within existing frame.

Mental model: Thermostat — detect deviation, adjust action, return to target. The goal is not questioned.

Process: Gap detected → Diagnose cause → Identify correction → Verify fix → Prevent recurrence

Key rules:

• Fix the proximate cause
• Don't question the goal (yet)
• Add prevention to avoid repeat
• Check: is this a pattern? If yes → double-loop

Output: Corrected action with prevention

→ references/single-loop.md

Double-Loop

Purpose: Question and update the frame itself.

Mental model: Not just adjusting thermostat, but asking: "Is heating the right goal?"

Process: Pattern detected → Examine current frame → Challenge assumptions → Construct new frame → Validate change

Key rules:

• Requires 3+ single-loop failures (pattern)
• Articulate current frame (goals, assumptions, constraints)
• Challenge each element with evidence
• Test new frame before full commitment

Output: Updated frame with validation plan

→ references/double-loop.md

Reflection

Purpose: Extract transferable insight from experience.

Mental model: Mine the experience for reusable gold.

Process: Capture experience → Analyze what worked/didn't → Extract insights → Update beliefs → Create artifacts → Disseminate

Key rules:

• Reflection is scheduled, not accidental
• Analyze both successes and failures
• Specify conditions when insight applies
• Create persistent artifacts (heuristics, playbooks, checklists)

Output: Insights and artifacts for future use

→ references/reflection.md

Experimentation

Purpose: Test belief through deliberate action before commitment.

Mental model: Scientific method applied to operational decisions.

Process: Formulate hypothesis → Design experiment → Execute → Analyze results → Conclude → Act

Key rules:

• Hypothesis must be falsifiable
• Define success criteria before testing
• Control variables where possible
• Don't peek at results early

Output: Validated or invalidated belief with next steps

→ references/experimentation.md

Calibration

Purpose: Adjust prediction confidence based on track record.

Mental model: Weather forecaster — when I say 80% confident, it should be right 80% of the time.

Process: Assemble track record → Stratify by confidence level → Calculate calibration error → Identify patterns → Define adjustment rules

Key rules:

• Need 30+ predictions for meaningful calibration
• Stratify by domain (calibration varies)
• Adjust gradually, not dramatically
• Monitor ongoing calibration

Output: Calibration adjustment rules

→ references/calibration.md

Output Format

Every learning output includes:

## [Mode]: [Topic]

**Trigger:** [What triggered this learning mode]

**Analysis:**
[Mode-specific analysis]

**Conclusion:**
[What was learned/changed]

**Artifacts:**
- [Any persistent outputs: rules, checklists, playbooks]

**Next:**
- [Actions to take]
- [What to monitor]

Mode Transitions

Learning → Other Skills Handoff

Anti-Patterns

References