{
  "added_at": "2026-10-19",
  "additional_evidence_urls": [
    "https://github.com/2akouwu/reverify/blob/f32ea84ccd6bb2aff90eca711ccfdbf04e6b8923/BENCHMARK.md",
    "https://github.com/2akouwu/reverify/blob/f32ea84ccd6bb2aff90eca711ccfdbf04e6b8923/EXAMPLE.md",
    "https://github.com/2akouwu/reverify/blob/f32ea84ccd6bb2aff90eca711ccfdbf04e6b8923/pyproject.toml",
    "https://github.com/2akouwu/reverify/blob/f32ea84ccd6bb2aff90eca711ccfdbf04e6b8923/reverify/mcp_server.py",
    "https://github.com/2akouwu/reverify/blob/f32ea84ccd6bb2aff90eca711ccfdbf04e6b8923/reverify/agent.py",
    "https://github.com/2akouwu/reverify/blob/f32ea84ccd6bb2aff90eca711ccfdbf04e6b8923/reverify/ledger.py",
    "https://github.com/2akouwu/reverify/blob/f32ea84ccd6bb2aff90eca711ccfdbf04e6b8923/.github/workflows/ci.yml",
    "https://github.com/2akouwu/reverify/blob/f32ea84ccd6bb2aff90eca711ccfdbf04e6b8923/SECURITY.md",
    "https://github.com/2akouwu/reverify/blob/f32ea84ccd6bb2aff90eca711ccfdbf04e6b8923/LICENSE",
    "https://github.com/2akouwu/reverify/releases/tag/v0.11.0",
    "https://pypi.org/project/reverify/0.10.0/"
  ],
  "category": "agent-workflows",
  "description_en": "Python CLI and stdio MCP server that checks a model’s structured claims about a binary, such as bytes or typed values at an offset, strings, imports and exports, instruction mnemonics, emulation results and tested or Z3-proved equivalence, and returns VERIFIED, REFUTED or INCONCLUSIVE with the evidence the tools observed. Grounded and refuted results are kept in a per-binary ledger on disk, so a reset or later session resumes from them instead of from a summary, and a reconstruct loop has a model propose claims until an informative set verifies. A separate rollover command replaces context compaction in Claude Code, Codex, Gemini CLI or OpenCode with file hand-offs and fresh sessions.",
  "description_zh": "Python 命令行工具和 stdio MCP 服务器，用于检查模型对二进制文件提出的结构化断言，例如某偏移处的字节或类型化数值、字符串、导入与导出、指令助记符、模拟执行结果，以及经测试或经 Z3 证明的等价性，并连同工具实际观察到的证据返回 VERIFIED、REFUTED 或 INCONCLUSIVE。已确证和已驳斥的结果保存在磁盘上按二进制文件划分的台账中，重置后或之后的会话可据此继续，而不依赖摘要；reconstruct 循环则让模型不断提出断言，直至一组有信息量的断言通过验证。另有 rollover 命令，用文件交接和新会话取代 Claude Code、Codex、Gemini CLI 或 OpenCode 的上下文压缩。",
  "entry_url": "https://awesomerea.org/resources/reverify-claim-verifier/",
  "entry_url_zh": "https://awesomerea.org/zh/resources/reverify-claim-verifier/",
  "evidence_url": "https://github.com/2akouwu/reverify/blob/f32ea84ccd6bb2aff90eca711ccfdbf04e6b8923/README.md",
  "group": "agents",
  "id": "reverify-claim-verifier",
  "json_url": "https://awesomerea.org/resources/reverify-claim-verifier/record.json",
  "kind": "integration",
  "markdown_url": "https://awesomerea.org/resources/reverify-claim-verifier/index.md",
  "markdown_url_zh": "https://awesomerea.org/zh/resources/reverify-claim-verifier/index.md",
  "name": "Reverify",
  "name_zh": "Reverify：用确定性工具核验模型二进制分析断言",
  "record_type": "resource",
  "relationship": "independent-agent-integration",
  "requirements": [
    "The reviewed main commit is 25 commits after the v0.11.0 GitHub release of 2026-09-04; PyPI’s newest version is 0.10.0, and the README’s status section still describes v0.9.0. The core is pure Python with no dependencies, and pyproject.toml declares Python 3.8+. The full extra adds capstone, unicorn, lief and z3-solver, and the angr extra adds the function, call and cross-reference claims, which return INCONCLUSIVE without an engine. The toolkit also parses PE, ELF and Mach-O files, scans byte patterns, decodes Protobuf and TLV data and generates Frida hook scripts.",
    "The verifier and benchmarks need no model. reconstruct calls an OpenAI-compatible endpoint configured by OPENAI_API_KEY, OPENAI_BASE_URL and OPENAI_MODEL, orchestrate can drive Claude through the Claude Agent SDK on a Claude Code login, and the MCP server gives any MCP client parsing, disassembly, pattern-scan, semantic, claim, ledger and checkpoint tools. Ledgers and checkpoints are written under .reverify in the working directory. Users supply the binaries to analyse; the repository holds benchmark sources and result records, not target binaries.",
    "reverify rollover install edits the user settings of each supported CLI it finds to add hooks and switch off native compaction, keeping backups and offering uninstall. The README warns that a Claude Code or Codex session not started through its launcher can then grow without limit, and reports one session that reached 909k tokens.",
    "The headline benchmark applies a fixed textbook push rbp; mov rbp, rsp prologue guess, not a live model, to the entry points of system binaries. BENCHMARK.md reports that guess refuted on 69 of 71 Windows 11 files and never falsely verified, and 0 false VERIFIED among 475 known-false claims on 50 Windows binaries; CI repeats both checks on Linux, Windows and macOS and fails on any false VERIFIED. The author notes that 0 of N bounds the error rate rather than proving it zero. These are project-reported results.",
    "MIT. SECURITY.md limits use to authorised work such as malware analysis, CTF, interoperability research and software the user owns or may analyse, and excludes circumventing licensing or copy protection on software the user does not own or license."
  ],
  "requirements_zh": [
    "所审阅的 main 提交位于 2026-09-04 发布的 GitHub v0.11.0 之后 25 个提交；PyPI 上最新版本为 0.10.0，README 的状态部分仍在介绍 v0.9.0。核心为纯 Python 实现、无依赖，pyproject.toml 声明支持 Python 3.8+。full 附加组件加入 capstone、unicorn、lief 和 z3-solver，angr 附加组件提供函数、调用和交叉引用类断言，未安装相应引擎时这些断言返回 INCONCLUSIVE。工具包还能解析 PE、ELF 和 Mach-O 文件，扫描字节模式，解码 Protobuf 与 TLV 数据，并生成 Frida 钩子脚本。",
    "验证器与基准测试无需模型。reconstruct 调用由 OPENAI_API_KEY、OPENAI_BASE_URL 和 OPENAI_MODEL 配置的 OpenAI 兼容端点，orchestrate 可借助 Claude Agent SDK 使用 Claude Code 登录驱动 Claude，MCP 服务器则向任意 MCP 客户端提供解析、反汇编、模式扫描、语义分析、断言核验、台账与检查点工具。台账与检查点写入工作目录下的 .reverify。待分析的二进制文件由使用者自备；仓库包含基准测试源码与结果记录，不含目标二进制文件。",
    "reverify rollover install 会修改所找到的每个受支持命令行工具的用户设置，加入钩子并关闭其原生压缩功能，同时保留备份并提供卸载命令。README 提醒，未通过其启动器启动的 Claude Code 或 Codex 会话随后可能无限增长，并报告有一个会话达到 909k token。",
    "主要基准测试把固定的教科书式 push rbp; mov rbp, rsp 序言猜测（而非实时模型）应用于系统二进制文件的入口点。BENCHMARK.md 报告该猜测在 71 个 Windows 11 文件中有 69 个被驳斥，且从未被误判为 VERIFIED；在 50 个 Windows 二进制文件上的 475 条已知错误断言中误判为 VERIFIED 的为 0。CI 在 Linux、Windows 和 macOS 上重复这两项检查，只要出现一次误判 VERIFIED 即失败。作者指出“N 次中 0 次”只给出错误率的上界，并不证明其为零。这些均为项目报告的结果。",
    "采用 MIT 许可。SECURITY.md 将用途限定于获授权的工作，如恶意软件分析、CTF、互操作研究以及使用者拥有或获准分析的软件，并禁止规避使用者未拥有或未获许可软件的授权或复制保护。"
  ],
  "targets": [
    "binary"
  ],
  "url": "https://github.com/2akouwu/reverify",
  "verification": "source-reviewed",
  "verified_at": "2026-10-19"
}