Proof-or-Stop: Don't Trust the Agent, Trust the Evidence -- Loop Engineering for Verifiable Evidence-Gated Lifecycle Control
- source_url: https://arxiv.org/abs/2607.14890
- source_type: paper
- platform: arxiv
- author: Jek Huang, Jeffery Hsia, Jiayi Sun, Freddie Shi, Wei Huang, Ian H. White
- original_date: 2026-07-16
- added_date: 2026-07-20
- local_path: OpenClaw定时任务/论文流水线/2026-07-20-论文流水线.md
- quality_score: 4
摘要(中文)
论文讨论 coding agent 的生命周期控制:reviewedtestedDONEready-to-merge 不能只看 agent 自述,要绑定新鲜可追踪可机械验证的证据作者报告 unattended-loop engine 在 10 个场景中没有 false-DONE,本地 receipt bundle 能拒绝多类篡改;消融结果显示,证据门禁能显著降低 visible-pass/hidden-fail 放大
Summary (English)
The paper proposes evidence-gated lifecycle control for coding agents, treating DONE and similar claims as assertions that must be backed by fresh, traceable, mechanically verifiable proof before state transitions.
One-liner
Agent 不能自证完成,生命周期推进应由可验证证据触发
原文 / 元数据抓取
[ { "id": "2607.14890", "title": "Proof-or-Stop: Don't Trust the Agent, Trust the Evidence -- Loop Engineering for Verifiable Evidence-Gated Lifecycle Control", "authors": "Jek Huang, Jeffery Hsia, Jiayi Sun, Freddie Shi, Wei Huang, Ian H. White", "abstract": "Autonomous coding agents increasingly execute multi-step software work, but lifecycle states such as reviewed, tested, DONE, and ready-to-merge remain claims unless supported by current evidence. We present Proof-or-Stop Lifecycle Control, a method that permits lifecycle transitions only when fresh, tracked-source-state-bound, mechanically verifiable evidence satisfies the relevant gate. The method treats agent outputs as claims rather than lifecycle state, and uses proof operationally to mean gate-admissible evidence under a stated trust model, not semantic program correctness. We evaluate an open-source implementation through mechanism tests, a powered control-policy ablation, and operated self-application evidence. The unattended-loop engine passed 10 of 10 scenarios with zero false-DONE, and local-key receipt bundles rejected 18 tamper classes with zero false accepts. In a 9,240-cell ablation, the pre-registered A4 versus A2-prime comparison reduced visible-pass/hidden-fail amplification from 31 of 1,800 injected cells under a compute-budgeted naive loop to 2 of 1,800 under the gated loop, a 1.6 percentage-point improvement in not-amplified rate with a 95 percent confidence interval of [0.8, 2.5]. A near-compute A3 versus A4 comparison, 14 of 1,800 versus 2 of 1,800, indicates that the gain is associated with enforcing review as a lifecycle gate rather than merely adding a reviewer. The self-application corpus contains 565 stories and 1,007 review findings, with 94.8 percent resolved, plus a 68-row high/critical cross-vendor exhibit. These results support Proof-or-Stop as a model-agnostic, host-neutral control layer for deciding which autonomous-agent claims a lifecycle may act on. The evaluation is limited to one model family, 24 ablation tasks, and a self-hosted corpus.", "published": "2026-07-16", "updated": "2026-07-16", "primary_category": "cs.AI", "categories": "cs.AI, cs.SE", "comment": "48 pages, 10 figures, 29 numbered tables. Preprint v1", "pdf": "https://arxiv.org/pdf/2607.14890v1", "url": "https://arxiv.org/abs/2607.14890" } ]
Obsidian intake evidence excerpt
论文流水线 · 2026-07-20
- status: completed
- Obsidian:
/Users/gracker/Library/Mobile Documents/iCloud~md~obsidian/Documents/Obsidian/OpenClaw定时任务/论文流水线/2026-07-20-论文流水线.md - Evidence:
/Users/gracker/.hermes/evidence/paper-pipeline/2026-07-20-论文流水线.md - 检索方式:
opencli arxiv recent/search/paper;未调用 OpenClaw。
今日论文速报
1. Setup Complete, Now You Are Compromised: Weaponizing Setup Instructions Against AI Coding Agents
- 来源: arXiv:2607.15143, 2026-07-16, cs.CR / cs.HC / cs.SE
- 作者: Aadesh Bagmar, Pushkar Saraf
- URL: https://arxiv.org/abs/2607.15143
- PDF: https://arxiv.org/pdf/2607.15143v1
- 速读: 论文把 AI coding agent 的项目初始化流程当成攻击面:README、requirements、Makefile 里的安装指令可以把 agent 引到不可信 registry、已知漏洞版本或相似包名。实验覆盖 12 个场景、5 类攻击,结论指向 harness-model 组合,而不是模型单点能力。typosquat 这种明显攻击更容易被拦住,
azurecore/azure-core这类分隔符混淆和 registry 重定向更容易漏掉。作者给出的工程解法是安装前做确定性校验:包名、来源、版本都过门禁,再允许执行。
2. Proof-or-Stop: Don't Trust the Agent, Trust the Evidence -- Loop Engineering for Verifiable Evidence-Gated Lifecycle Control
- 来源: arXiv:2607.14890, 2026-07-16, cs.AI / cs.SE
- 作者: Jek Huang, Jeffery Hsia, Jiayi Sun, Freddie Shi, Wei Huang, Ian H. White
- URL: https://arxiv.org/abs/2607.14890
- PDF: https://arxiv.org/pdf/2607.14890v1
- 速读: 论文讨论 coding agent 的生命周期控制:reviewed、tested、DONE、ready-to-merge 不能只看 agent 自述,要绑定新鲜、可追踪、可机械验证的证据。作者报告 unattended-loop engine 在 10/10 场景中没有 false-DONE,本地密钥 receipt bundle 拒绝 18 类篡改且无 false accept;9,240-cell 消融里,证据门禁把 visible-pass/hidden-fail 放大从 31/1800 降到 2/1800。适合拿来对照现有 cron / agent 流程里的“完成”判定。
3. SearchOS-V1: Towards Robust Open-Domain Information-Seeking Agent Collaboration
- 来源: arXiv:2607.15257, 2026-07-16, cs.AI / cs.IR
- 作者: Yuyao Zhang 等
- URL: https://arxiv.org/abs/2607.15257
- PDF: https://arxiv.org/pdf/2607.15257v1
- 代码: https://github.com/antins-labs/SearchOS
- 速读: SearchOS 把开放域搜索建模成“带引用的关系表补全”,用 Frontier Task、Evidence Graph、Coverage Map、Failure Memory 记录搜索状态,避免多 agent 搜索时反复撞同一堵墙。它还用 pipeline-parallel scheduling 填满空闲 agent 槽位,用 middleware harness 记录证据、检测停滞和预算耗尽。对需要长期跑的资料收集任务,这篇比单纯提高模型更接近工程改造方向。
4. Beyond Success Rate: Cost-Aware Evaluation of Offensive and Defensive Security Agents
- 来源: arXiv:2607.15263, 2026-07-16, cs.CR / cs.AI
- 作者: Paul Kassianik, Blaine Nelson, Yaron Singer
- URL: https://arxiv.org/abs/2607.15263
- PDF: https://arxiv.org/pdf/2607.15263v1
- 速读: 论文把安全 agent 的评测从“成功率”拉回到成本:推理、工具调用、遥测查询、补充检索都要计费。红队 CTF 任务能从更多 test-time compute 获益,蓝队 SOC 调查对工具使用纪律、遥测导航和选择性 enrichment 更敏感。这个结论可迁移到日常 agent:长任务不能只看是否完成,还要看每一步消耗和证据质量。
5. Capturing and Exploiting Design Pattern Variability in Mobile Application Generation
- 来源: arXiv:2607.15099, 2026-07-16, cs.SE
- 作者: Ramón Peralta, Jose-Miguel Horcas
- URL: https://arxiv.org/abs/2607.15099
- PDF: https://arxiv.org/pdf/2607.15099v1
- 速读: 论文把移动应用生成中的设计模式做成可配置资产,用 UVL 描述 Singleton、Strategy、Observer、Adapter、Factory Method 的结构和行为变体,再用 Jinja 模板生成 Swift 代码。它偏软件产品线和模型驱动工程,对 Android 不是直接命中,但对“生成代码如何保住架构质量”有参考价值。
6. On the Adoption of AI Coding Agents in Open-source Android and iOS Development
- 来源: arXiv:2602.12144, 2026-02-12, cs.SE / cs.AI;MSR 2026 Mining Challenge track accepted
- 作者: Muhammad Ahmad Khan, Hasnain Ali, Muneeb Rana, Muhammad Saqib Ilyas, Abdul Ali Bangash
- URL: https://arxiv.org/abs/2602.12144
- PDF: https://arxiv.org/pdf/2602.12144v1
- 速读: 这是今天的 Android 相关补充候选。作者分析 AIDev 数据集里 193 个 Android / iOS 开源仓库的 2,901 个 AI-authored PR。Android 项目收到的 AI PR 数量约为 iOS 的 2 倍,接受率 71%,iOS 为 63%。feature、fix、ui 这类常规任务更容易合并,refactor、build 这类结构性改动成功率更低、处理时间更长。
值得精读的论文
1. 首选: Setup Complete, Now You Are Compromised 推荐理由: 它把 agent 安全问题落到项目初始化这一段:读文档、装依赖、跑 setup。这正是 coding agent 最常被授权执行、也最容易被当成“准备工作”的环节。精读时应重点看 12 个攻击场景、5 类攻击定义、不同 harness-model 组合的差异、确定性 pre-install check 的实现边界。
2. 备选: Proof-or-Stop 推荐理由: 它给了一个可操作的 agent 证据门禁框架,和 Hermes cron 单任务化的要求很贴:不能只让 agent 说完成,要能绑定输入状态、输出证据、测试结果和生命周期动作。精读时应核对 10 个机制测试、18 类篡改、9,240-cell 消融设计,以及“proof”在论文里只指门禁证据,不等于程序语义正确。
3. 备选: SearchOS-V1 推荐理由: 长周期 web research / paper pipeline 的失败常见于重复搜索、证据散落、覆盖缺口没人记录。SearchOS 的 Evidence Graph、Coverage Map、Failure Memory 可以作为论文流水线后续改造的参考。
Android/Agent 相关性
- Android 约束检查: 本次没有采用 Android 18 / API 38+ 相关内容。Android 相关候选集中在移动开发、GUI agent、Android/iOS 开源 PR 数据,没有触碰 Android 17 / API 37 之后的平台 API 断言。
- 对 Android 工程的价值:
2602.12144直接给出 Android 开源仓库中 AI PR 的接受率、任务类型差异和平台差异;2607.15099对移动代码生成的架构质量有间接参考。 - 对 Agent 工程的价值:
2607.15143指向依赖安装前门禁,2607.14890指向证据绑定的生命周期控制,2607.15257指向持久搜索状态,2607.15263指向成本敏感评测。这四篇组合起来,正好覆盖 agent 的输入、执行、搜索、评估四个环节。
工作日 brief(周一至周五;周末写“今日不更新”)
今天是周一,brief lane 更新。
- 今天最该看:
2607.15143。如果只能读一篇,就看它的攻击分类和 pre-install check。它能直接变成 coding agent / cron pipeline 的安全门禁清单。 - 今天适合加入候选队列:
2607.14890和2607.15257。前者适合写“Agent 不能自证完成”,后者适合写“搜索任务要有状态,不要靠对话历史硬撑”。 - 今天可暂缓:
2607.15099。移动方向相关,但示例偏 Swift 和设计模式变体建模,不是 Android 性能或平台机制的一线材料。 - 风险提醒:
Proof-or-Stop的数据来自一个模型家族、24 个消融任务和自托管语料,不能把结果外推成所有 agent 系统都适用。Setup Complete的结论也要结合具体 harness 权限模型看。
建议下一步
1. 建议把 2607.15143 作为今日精读包候选,下载 PDF 后建立 canonical package:论文/AI-2026-07-20-setup-complete-now-you-are-compromised/。 2. 精读时先抽取这几类信息:攻击面定义、12 个场景、5 类攻击、模型与 harness 配对差异、pre-install check 规则、作者承认的限制。 3. 把 2607.14890 放入 agent 工程候选队列,后续可写成 Hermes cron 质量门禁参考:claim 不能直接触发 lifecycle transition,必须有新鲜证据。 4. 把 2607.15257 放入 paper pipeline 改造参考,优先借鉴 Coverage Map 和 Failure Memory,减少重复检索。
今日精选
1. AI coding agent 的供应链风险开始从“代码生成错了”前移到“项目还没跑起来,setup 已经被投毒”。 2. agent 生命周期里的 DONE / tested / reviewed 都应该被当成待验证断言,不能直接当状态用。 3. 长周期搜索 agent 需要显式状态:任务前沿、证据图、覆盖缺口、失败记忆。只靠聊天历史,预算越长越容易绕圈。
可直接发布文案
今天看到一篇很值得读的 agent 安全论文:Setup Complete, Now You Are Compromised。它盯的不是“模型会不会写出漏洞代码”,而是更靠前的一步:AI coding agent 读 README、装依赖、跑 Makefile 的时候,普通项目文档就可能变成攻击入口。
作者测试了 12 个场景和 5 类攻击。结果很工程:明显的 typosquat 容易被拦,像 azurecore / azure-core 这种相似包名、registry 重定向、来源混淆更容易漏;同一个模型换