Interlock：为 AI 基础设施构建的熔断器与证据层

项目概述

Interlock is a circuit-breaker + evidence layer for AI infrastructure. It monitors runtime signals (confidence/latency/failure), refuses or degrades when the system is outside safe bounds, and writes verifiable forensic logs.

Interlock 是一个为 AI 基础设施设计的熔断器与证据层。它监控运行时信号（置信度/延迟/故障），当系统超出安全边界时拒绝或降级请求，并生成可验证的取证日志。

核心功能

Interlock 提供以下核心功能：

• Quality gating: refuse/degrade low-certainty responses (质量门控：拒绝或降级低置信度的响应)
• Survivability: shed load and recover cleanly under stress (生存能力：在压力下优雅地卸载负载并恢复)
• Evidence: cryptographic provenance for interventions + configs (证据：为干预措施和配置提供加密来源证明)
• Adapters: integration points for vector DBs and local models (适配器：为向量数据库和本地模型提供集成点)

快速开始

文档：Quickstart · Local AI · Middleware

Docs: Quickstart · Local AI · Middleware

Express (Node.js)

import { interlockExpress } from '@interlock/express';
app.use(interlockExpress({ quality_floor: 0.8 }));

FastAPI (Python)

from interlock_fastapi.middleware import InterlockMiddleware
app.add_middleware(InterlockMiddleware, interlock_url="http://brain:3000")

支持的集成

Interlock 与主流 AI 基础设施组件广泛兼容，其集成状态如下：

Supported Integrations

Interlock offers broad compatibility with mainstream AI infrastructure components. The integration status is as follows:

Integration	Status	Notes
Pinecone	✅ Production	Production-ready
FAISS	✅ Production	Production-ready
LangChain	✅ Production	Production-ready
LlamaIndex	✅ Production	Production-ready
Weaviate	✅ Stable	Stable
Milvus	✅ Stable	Stable
Ollama / Local AI	✅ Tested	Tested

证据与验证

CI workflows produce repeatable artifacts (stress, scale, adapter certification).

持续集成（CI）工作流生成可重复的制品（压力测试、规模测试、适配器认证）。

参见：Test Results · Live Incidents

See: Test Results · Live Incidents

OperatorPack 收据验证

Interlock includes tools to verify and index OperatorPack receipts (JSON).

Interlock 包含用于验证和索引 OperatorPack 收据（JSON格式）的工具。

使用方法

1. 验证收据：

   python tools/verify_operatorpack.py "C:\path\to\operatorpack.json"
   

   根据性能和质量阈值输出 JSON 格式的裁决结果（PASS/WARN/FAIL）。

   Outputs a JSON verdict (PASS/WARN/FAIL) based on performance and quality thresholds.

2. 索引收据：

   python tools/append_receipt_index.py "C:\path\to\operatorpack.json" "receipts/RECEIPTS_INDEX.md"
   

   将指标追加到 Markdown 表格 索引文件中，用于长期跟踪。

   Appends metrics to a Markdown table index file for long-term tracking.

推荐工作流

Place new receipts in receipts/inbox/ for verification before promotion.

将新收据放入 receipts/inbox/ 目录，以便在升级前进行验证。

使用示例收据：

python tools/verify_operatorpack.py "receipts/examples/operatorpack_example_pass.json"

收据工作流

Interlock supports a standard promotion workflow for OperatorPacks with 2-Tier Governance:

Interlock 为 OperatorPack 支持一个标准的升级工作流，采用 两级治理 模式：

1. 收件箱：将新的收据 JSON 文件复制到 receipts/inbox/。

   Inbox: Copy new receipt JSON files into receipts/inbox/.

2. 升级：使用所需的层级运行升级脚本：

   Promote: Run the promotion script with the desired tier:

   • 生产环境（默认）：严格的门控（$N \ge 10,000$）。

     pwsh -File tools/promote_receipt.ps1
     

     • 结果存放在 receipts/approved/ 或 receipts/rejected/。
     • 主索引：RECEIPTS_INDEX.md。

       Production (Default): Strict gates ($N \ge 10,000$).

       • Results land in receipts/approved/ or receipts/rejected/.
       • Primary index: RECEIPTS_INDEX.md.

   • 探索环境：更快的迭代门控（$N \ge 2,000$）。

     pwsh -File tools/promote_receipt.ps1 -Mode exploration
     

     • 结果存放在 receipts/approved_experimental/ 或 receipts/rejected_experimental/。
     • 二级索引：RECEIPTS_INDEX_EXPERIMENTAL.md。

       Exploration: Faster iteration gates ($N \ge 2,000$).

       • Results land in receipts/approved_experimental/ or receipts/rejected_experimental/.
       • Secondary index: RECEIPTS_INDEX_EXPERIMENTAL.md.

3. 审计：审查 receipts/approved/ 文件夹和 RECEIPTS_INDEX.md（或对应的实验环境文件）以获取历史数据。

   Audit: Review the receipts/approved/ folder and RECEIPTS_INDEX.md (or the experimental equivalents) for historical data.

注意：已批准的收据和摘要均在本地生成，不会提交到代码仓库，以维护数据隐私。
Note: Approved receipts and summaries are generated locally and are not committed to the repository to maintain data privacy.

实践：后续步骤

1. 生成证据：从实验室（例如 TESLA）获取收据 JSON（OperatorPack）——将其保留在 Interlock 外部作为事实来源。

   Generate Evidence: Obtain a receipt JSON (OperatorPack) from a lab (e.g., TESLA) — keep it outside Interlock as the source of truth.

2. 暂存：将其复制到 receipts/inbox/。

   Stage: Copy it into receipts/inbox/.

3. 升级（生产环境）：

   pwsh -File tools/promote_receipt.ps1
   

   或指定文件：pwsh -File tools/promote_receipt.ps1 -Receipt "C:\path\to\receipt.json"

   Promote (Production):
   Or specify a file: pwsh -File tools/promote_receipt.ps1 -Receipt "C:\path\to\receipt.json"

4. 升级（探索环境）：

   pwsh -File tools/promote_receipt.ps1 -Mode exploration
   

   Promote (Exploration):

5. 汇总：

   python tools/summarize_receipts.py
   

   Summarize:

6. 制品位置：

   Artifact Locations:

   • receipts/approved/ 或 receipts/approved_experimental/（仅本地）

     • receipts/approved/ or receipts/approved_experimental/ (Local only)

   • receipts/rejected/ 或 receipts/rejected_experimental/（仅本地）

     • receipts/rejected/ or receipts/rejected_experimental/ (Local only)

   • receipts/summary/（仅本地）

     • receipts/summary/ (Local only)

These outputs are local artifacts and are gitignored by design in this public repo.

这些输出是本地制品，在此公共仓库中已通过 .gitignore 设计为忽略。

AI 协作

This repo uses a multi-agent workflow with safety guardrails.

本仓库采用具有安全护栏的多智能体工作流。

• 项目记忆：CLAUDE.md - 构建命令、安全规则、智能体分配

  Project Memory: CLAUDE.md - Build commands, safety rules, agent assignment

• 安全策略：docs/AI_COLLAB_SAFETY.md - PII 保护、合并门控

  Safety Policy: docs/AI_COLLAB_SAFETY.md - PII protection, merge gates

• 协调：multi_agent_coord.template.md - 交接协议模板

  Coordination: multi_agent_coord.template.md - Handoff protocol template

AI agents must run tools/precommit_safety_scan.ps1 before commits.

AI 智能体在提交前必须运行 tools/precommit_safety_scan.ps1。

许可证

MIT. See LICENSE.

MIT。参见 LICENSE。

常见问题（FAQ）

Interlock 如何提升 AI 搜索系统的可靠性？

Interlock 作为 AI 基础设施的熔断器，通过监控运行时信号（如置信度、延迟），在系统超出安全边界时自动拒绝或降级低质量响应，从而保障系统在压力下的生存能力和服务质量。

Interlock 支持哪些常见的 AI 工具或数据库集成？

Interlock 与主流 AI 基础设施广泛兼容，包括生产就绪的 Pinecone、FAISS、LangChain、LlamaIndex，以及稳定的 Weaviate、Milvus，并支持已测试的 Ollama / Local AI 等本地模型集成。

Interlock 的加密取证日志有什么作用？

Interlock 提供的加密取证日志为所有干预操作和配置变更生成可验证的来源证明，用于质量控制和事后审计，确保系统行为的透明性与可追溯性，增强整体系统弹性。