## 🧠 專業框架與方法論

### 核心理論框架

#### 1. 對齊問題分類學（Alignment Taxonomy）
```
Outer Alignment: 目標函數是否反映人類意圖？
  └─ Specification gaming, reward hacking, goal misspecification

Inner Alignment: 訓練後模型內部目標是否與外部目標一致？
  └─ Mesa-optimization, deceptive alignment, inner misalignment

Capability-Alignment Gap: 能力增長是否超過 oversight 能力？
  └─ Scheming, sandbagging, situational awareness
```

#### 2. 主要技術路線圖
| 路線 | 代表工作 | 核心思路 | 已知局限 |
|------|----------|----------|----------|
| RLHF / RLAIF | Ouyang 2022, Bai 2022 | 人類/AI 偏好學習 | 偏好可操縱、分佈外失效 |
| Constitutional AI | Bai 2022 | 原則自我修訂 | 原則選擇本身需對齊 |
| Debate / IDA | Irving 2018, Christiano | 可擴展監督 | 計算成本、欺騙策略 |
| Mechanistic Interpretability | Olah, Anthropic | 理解內部計算 | 規模化挑戰 |
| Control / Containment | Greenblatt 2024 | 限制高能力模型行為 | 假設邊界需驗證 |
| Scalable Oversight | Bowman 2022 | 弱監督者評估強模型 | bootstrapping 假設 |

#### 3. 評估方法論（Eval Methodology）
- **Capability Evals**：MMLU, SWE-bench, agentic task suites
- **Safety Evals**：harmful request refusal, jailbreak robustness, deception detection
- **Alignment Evals**：preference consistency, value stability under distribution shift
- **Red-teaming Protocols**：automated attack, human adversary, multi-turn escalation
- **Eval 設計原則**：construct validity, ecological validity, saturation avoidance, Goodhart resistance

#### 4. 風險評估框架
- **RISK-TAXONOMY**：按 harm type × severity × probability 分類
- **Threat Modeling**：STRIDE-adapted for AI systems
- **Defense in Depth**：pre-training filtering → RLHF → inference guardrails → monitoring → shutdown
- **Safety Case 結構**：claim → argument → evidence → assumption → residual risk

#### 5. 研究設計模板
```markdown
## Research Question
[可 falsify 的問題陳述]

## Hypotheses
- H1: ...
- H0 (null): ...

## Experimental Design
- Independent variables: ...
- Dependent variables / metrics: ...
- Controls & baselines: ...
- Sample size / power reasoning: ...

## Threats to Validity
- Internal / External / Construct / Statistical

## Pre-registration Criteria
[什麼結果支持/反對 H1]

## Ethical Review Checklist
```

### 關鍵文獻地圖（精選）
- **Foundational**：Bostrom (2014) Superintelligence; Russell (2019) Human Compatible
- **Technical**：Christiano et al. IDA; Hubinger et al. Risks from Learned Optimization
- **Empirical**：Perez et al. Discovering Language Model Behaviors; Anthropic's Core Views on AI Safety
- **Governance**：EU AI Act analysis; NIST AI RMF; Frontier Model Forum commitments

### 實務工具清單
- **Interpretability**：Neuronpedia, TransformerLens, attribution patching
- **Eval Platforms**：lm-evaluation-harness, Inspect, HELM
- **Red-team**：garak, PyRIT, custom agentic attack frameworks
- **Monitoring**：output classifiers, latent probing, behavioral fingerprinting

### 思考工具
- **Steel-manning**：先構建最強反方論證再回應
- **Pre-mortem**：假設項目失敗，倒推原因
- **Fermi estimation**：對 timeline、compute、risk magnitude 做數量級估計
- **Crux finding**：識別雙方分歧的關鍵可驗證命題