feat: +1 article — Classifier-Based Permissions (22 total) (#63)

Based on Anthropic Engineering 'Claude Code auto mode' (Mar 25, 2026).

- Two-layer defense architecture (input probe + transcript classifier)
- Four threat models (overeager / honest mistakes / prompt injection / misaligned)
- Three-tier decision flow, reasoning-blind design
- 347 lines EN + ZH translation
- Updated README, changelog, guide-data.ts

First production run of harness-guide-pipeline detected this as uncovered high-signal source.

Co-authored-by: joey <joey@joeydeMacBook-Air.local>

Author: joeylee12629-star

README.md

@@ -60,6 +60,7 @@ This guide covers harness engineering from first principles to production patter
 | [Long-Running Harness Design](guide/long-running-harness.md) | Context anxiety, self-evaluation bias, context reset vs compaction, GAN-inspired generator-evaluator architecture. |
 | [Managed Agents Architecture](guide/managed-agents-architecture.md) | Brain/hands/session decoupling, pets vs cattle, credential isolation, TTFT improvements. |
 | [Eval Infrastructure Noise](guide/eval-infrastructure.md) | Resource config swings benchmark scores by 6pp. Floor+ceiling enforcement strategy. |
+| [Classifier-Based Permissions](guide/classifier-permissions.md) | Replace approval fatigue with model-based classifiers. Two-layer defense, four threat models, reasoning-blind design. |
 
 ## Reference
 

README.zh-CN.md

@@ -60,6 +60,7 @@
 | [长时运行 Harness 设计](zh-guide/long-running-harness.md) | Context 焦虑、自评估偏差、Reset vs Compaction、GAN 启发的生成器-评估器架构。 |
 | [Managed Agents 架构](zh-guide/managed-agents-architecture.md) | Brain/Hands/Session 三层解耦、Pets vs Cattle、凭证隔离、TTFT 改进。 |
 | [评测基础设施噪声](zh-guide/eval-infrastructure.md) | 资源配置导致 Benchmark 得分波动 6 个百分点。Floor+Ceiling 执行策略。 |
+| [基于分类器的权限审批](zh-guide/classifier-permissions.md) | 用模型分类器替代 approval fatigue。双层防御、四种威胁模型、reasoning-blind 设计。 |
 
 ## 参考
 

changelog/2026-04-19.md

@@ -0,0 +1,15 @@
+---
+date: "2026-04-19"
+title: "+1 Article — Classifier-Based Permissions (22 Total)"
+---
+
+# 2026-04-19 — +1 New · 22 Total
+
+## Practice (+1 from Anthropic Engineering, 11 total)
+
+- [Classifier-Based Permissions](/guide/classifier-permissions) — Replace approval fatigue with model-based classifiers. Two-layer defense (input probe + output transcript classifier), four threat models (overeager / honest mistakes / prompt injection / misaligned), reasoning-blind design, three-tier decision flow, and real incident examples from Anthropic's internal log.
+
+## Pipeline
+
+- First production run of `harness-guide-pipeline` skill detected Anthropic's Mar 25, 2026 "Claude Code auto mode" post as uncovered high-signal source.
+- Confirmed `anthropic.com/engineering` as Tier 0 remains the highest-yield source for this site.

guide/classifier-permissions.md

@@ -43,6 +43,7 @@ export const guideSections: GuideSection[] = [
       { slug: "long-running-harness", title: "Long-Running Harness Design", zhTitle: "长时运行 Agent Harness 设计" },
       { slug: "managed-agents-architecture", title: "Managed Agents Architecture", zhTitle: "Managed Agents：Brain/Hands/Session 解耦" },
       { slug: "eval-infrastructure", title: "Eval Infrastructure Noise", zhTitle: "评测基础设施噪音" },
+      { slug: "classifier-permissions", title: "Classifier-Based Permissions", zhTitle: "基于分类器的权限审批" },
     ],
   },
   {

site/lib/guide-data.ts

@@ -0,0 +1,15 @@
+---
+date: "2026-04-19"
+title: "新增 1 篇：基于分类器的权限审批（累计 22 篇）"
+---
+
+# 2026-04-19 — 新增 1 篇 · 累计 22 篇
+
+## 实战（来自 Anthropic Engineering，新增 1，累计 11）
+
+- [基于分类器的权限审批](/zh/guide/classifier-permissions) — 用模型分类器替代 approval fatigue。双层防御（输入层 prompt-injection probe + 输出层 transcript classifier）、4 种威胁模型（overeager / honest mistakes / prompt injection / misaligned）、reasoning-blind 设计、三 Tier 决策流程，附 Anthropic 内部 incident log 的真实案例。
+
+## 管线
+
+- `harness-guide-pipeline` skill 首次生产运行，发现 Anthropic 3/25 的 "Claude Code auto mode" 是网站未覆盖的高价值内容源。
+- 确认 `anthropic.com/engineering` 作为 Tier 0 信息源产出最高。