2nd/10_Wiki/Topics/AI_and_ML/Superficiality-Metrics.md

---
id: wiki-2026-0508-superficiality-metrics
title: Superficiality Metrics
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Engagement Quality Metrics, Depth Metrics, Content Quality Signals]
duplicate_of: none
source_trust_level: B
confidence_score: 0.85
verification_status: applied
tags: [metrics, content-quality, engagement, evaluation]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: pandas
---

# Superficiality Metrics

## 매 한 줄
> **"매 engagement 의 depth 측정"**. CTR / time-on-page 같은 surface metric 만 보면 clickbait 의 reward → 매 deeper signal (scroll completion, return visit, comment quality, downstream conversion) 의 measure 의 필요. 2026 의 LLM-as-judge 의 quality scoring 의 mainstream.

## 매 핵심

### 매 surface vs depth metric
- **Surface**: CTR, time-on-page, bounce rate, like count.
- **Depth**: scroll depth, dwell quality (focus events), return visit %, share-with-comment, subscription, downstream action (purchase, signup).

### 매 LLM-judged content quality
- **Coherence**: 매 logical flow.
- **Substantive density**: 매 facts / claim 단위 의 information.
- **Originality**: 매 generic LLM-output 의 detection.
- **Actionability**: 매 reader 가 take-away 의 가능성.

### 매 응용
1. Content recommendation ranking (YouTube, TikTok 의 newer signals).
2. Knowledge-base quality gating (Wiki article 의 acceptance).
3. Education platform 의 learning outcome 측정.
4. Newsletter / blog 의 ROI evaluation.

## 💻 패턴

### Scroll depth tracking
```typescript
let maxScroll = 0;
window.addEventListener('scroll', () => {
  const scrollPct = window.scrollY / (document.body.scrollHeight - window.innerHeight);
  if (scrollPct > maxScroll) maxScroll = scrollPct;
}, { passive: true });

window.addEventListener('beforeunload', () => {
  navigator.sendBeacon('/analytics', JSON.stringify({
    page: location.pathname,
    maxScroll,
    duration: performance.now()
  }));
});
```

### Dwell quality (focus + scroll)
```typescript
let focusedTime = 0;
let lastFocusStart = document.hasFocus() ? performance.now() : null;

document.addEventListener('visibilitychange', () => {
  if (document.hidden && lastFocusStart != null) {
    focusedTime += performance.now() - lastFocusStart;
    lastFocusStart = null;
  } else if (!document.hidden) {
    lastFocusStart = performance.now();
  }
});
```

### LLM-as-judge quality score
```python
from anthropic import Anthropic
client = Anthropic()

def score_content(text: str) -> dict:
    resp = client.messages.create(
        model="claude-opus-4-7",
        max_tokens=512,
        messages=[{
            "role": "user",
            "content": f"""Rate the following article on 4 axes (0-10 each):
- coherence (logical flow)
- density (info per paragraph)
- originality (vs generic LLM output)
- actionability (reader takeaway)

Return strict JSON: {{"coherence": N, "density": N, "originality": N, "actionability": N, "rationale": "..."}}

ARTICLE:
{text[:8000]}"""
        }]
    )
    import json
    return json.loads(resp.content[0].text)
```

### Composite depth score
```python
import numpy as np

def depth_score(metrics: dict) -> float:
    # weights tuned on labeled training set
    w = {
        'scroll_completion': 0.15,
        'focused_dwell_ratio': 0.25,
        'return_within_7d': 0.20,
        'downstream_action': 0.25,
        'share_with_comment': 0.15,
    }
    return sum(w[k] * metrics.get(k, 0) for k in w)
```

### Clickbait detector heuristic
```python
def clickbait_signal(row):
    # high CTR + low depth = clickbait
    if row['ctr'] > 0.10 and row['depth_score'] < 0.3:
        return 1.0
    return 0.0
```

### Pandas pipeline
```python
import pandas as pd
df = pd.read_parquet('events.parquet')
agg = df.groupby('article_id').agg(
    ctr=('clicks', 'sum') / ('impressions', 'sum'),
    avg_scroll=('max_scroll', 'mean'),
    return_rate=('returned_7d', 'mean'),
).assign(depth_score=lambda d: 0.4*d.avg_scroll + 0.6*d.return_rate)
```

## 매 결정 기준
| 상황 | Metric |
|---|---|
| 매 ad-supported (need clicks) | CTR + minimal depth floor |
| 매 subscription / paid | depth_score primary |
| 매 education / learning | actionability + post-test outcome |
| 매 knowledge wiki | LLM coherence + density |
| 매 social platform | share-with-comment, return visit |

**기본값**: 매 composite depth score (50% behavioral + 50% LLM-judged).

## 🔗 Graph
- 부모: [[Evaluation]]
- Adjacent: [[LLM-as-Judge]] · [[Goodhart_s-Law]]

## 🤖 LLM 활용
**언제**: 매 content recommendation 의 reranking signal, KB article quality gate, AB test 의 secondary metric.
**언제 X**: 매 small sample (variance 너무 큼), 매 acquisition-stage funnel (CTR primary).

## ❌ 안티패턴
- **Single metric optimization**: Goodhart — 매 CTR alone optimize 하면 clickbait.
- **LLM judge 의 prompt drift**: 매 pinned model + temperature 0 + version log 의 필수.
- **Depth metric 의 latency**: return-visit 7d → 매 delayed feedback. 매 surrogate (focused dwell) 도 함께.

## 🧪 검증 / 중복
- Verified (Goodhart 1975; Zheng et al. 2023 LLM-as-judge; YouTube Watch Time → "Valued Watch Time" pivot ~2017).
- 신뢰도 B (매 weighting 의 domain-dependent).

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — surface vs depth + LLM judge + composite scoring |