---
id: wiki-2026-0508-prisons-and-self-correction
title: Prisons and Self-Correction
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Penitentiary System, Carceral Reform]
duplicate_of: none
source_trust_level: A
confidence_score: 0.85
verification_status: applied
tags: [criminology, justice, history, sociology]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: n/a
  framework: n/a
---

# Prisons and Self-Correction

## 매 한 줄
> **"매 1790s Quaker 의 penitence-as-cure 의 invention"**. 매 Eastern State Penitentiary (1829) 의 solitary-confinement model 매 "self-correction through silent reflection" → 매 Foucault (1975 Discipline & Punish) 의 critique → 매 2026 의 evidence-based recidivism reduction debate.

## 매 핵심

### 매 historical arc
1. Pre-1790: corporal/capital punishment, public execution.
2. 1790-1830: Quaker penitentiary (Pennsylvania system) — isolation + silence.
3. 1830-1900: Auburn system — silent congregate labor.
4. 1900s: rehabilitation ideal, parole.
5. 1970s-: "tough on crime" backlash, mass incarceration (esp. US).
6. 2010s-: evidence-based reform, Norway model (Halden), restorative justice.

### 매 modern data
- US incarceration rate: 매 ~600/100k (2024) — 매 highest among OECD.
- Norway recidivism: 매 ~20% within 2y; US: 매 ~67%.
- RAND meta-analysis: education programs 매 reduce recidivism 매 ~43%.

### 매 응용
1. Policy design (recidivism reduction, sentencing reform).
2. Software (case management, predictive risk — see fairness debate).
3. Restorative-justice programs.

## 💻 패턴

### Recidivism modeling (responsible)
```python
import pandas as pd
from sklearn.linear_model import LogisticRegression
from fairlearn.metrics import demographic_parity_difference

df = pd.read_csv('release_cohort.csv')
X = df[['age_at_release', 'prior_arrests', 'program_completed', 'employment_post']]
y = df['rearrest_within_3y']

model = LogisticRegression().fit(X, y)
pred = model.predict(X)

# fairness audit
print('DP diff (race):', demographic_parity_difference(y, pred, sensitive_features=df['race']))
```

### Risk-tool transparency (COMPAS critique)
```
ProPublica 2016 audit:
- Black defendants: 45% false-positive (predicted re-offend, didn't)
- White defendants: 23% false-positive
→ disparate-impact even when "race-blind"
```

### Halden Prison design principles (Norway)
```
1. Normalize: cell ≈ dorm room, common kitchens.
2. Education + work as default activity.
3. Short sentences (max 21y for most crimes).
4. Officer-inmate ratio high; relational, not custodial.
5. Pre-release housing transition.
```

### Restorative-justice circle script
```
1. Storytelling: harmed party speaks first.
2. Acknowledgment: harm-doer reflects.
3. Community impact discussion.
4. Repair plan: agreed actions, timeline.
5. Follow-up at 30/90 days.
```

### Education program ROI (RAND 2013)
```
Cost per inmate education: $1,400-1,744 / year
Reduced recidivism savings: ~$5/$1 invested
3-year recidivism: 43% reduction
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Reform policy design | Norway/Halden + restorative |
| Recidivism prediction | Avoid black-box; favor interpretable + fairness audits |
| Drug offenses | Treatment courts, not incarceration |

**기본값**: Education + employment + housing transition + restorative practices.

## 🔗 Graph
- 변형: [[Restorative Justice]]

## 🤖 LLM 활용
**언제**: policy analysis, criminology discussion, fairness-aware ML in criminal justice.
**언제 X**: 매 software-only topic (this is policy/sociology).

## ❌ 안티패턴
- **Black-box risk-assessment**: 매 unaudited disparate impact.
- **Solitary as default**: 매 mental-health damage 의 evidence.
- **Long sentences as deterrent**: 매 evidence weak; certainty > severity.

## 🧪 검증 / 중복
- Verified (Foucault — Discipline & Punish; Norway corrections white papers; RAND 2013 education meta-analysis; ProPublica COMPAS).
- 신뢰도 A-.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — Prisons & Self-Correction FULL content |