Research Papers 论文研究 7h ago Updated 3h ago 更新于 3小时前 45

ClaimReceipt: Verifying Evidence Sufficiency and Coverage in Agent Evaluations ClaimReceipt:验证智能体评估中的证据充分性与覆盖范围

ClaimReceipt introduces a claim-relative receipt specification and selective verifier that addresses two critical gaps in agent evaluation: evidence sufficiency (whether claims are recomputable from retained evidence) and coverage (whether retained records encompass the full committed experiment set) The system binds typed transaction evidence to a signed experiment manifest, returning one of three verdicts per claim: PASS, INVALID, or INCONCLUSIVE On 1,392 historical buyer-seller records, the C 提出ClaimReceipt框架,解决Agent评估中证据充分性(claim是否可从保留证据重新计算)和覆盖度(保留记录是否覆盖承诺实验集)两大核心问题 引入claim-relative收据规范和选择性验证器,将类型化交易证据绑定到签名实验清单,返回PASS/INVALID/INCONCLUSIVE三类裁决 在1,392条历史买家-卖家记录上验证:CR-2验证器重现全部5个手动审计裁决,精确重放600个确定性记录和792个后生成记录,11/11语义错误零误报 CR-3前瞻性实验实现:30个分配在推理前承诺,终端收据签名链接,私有证据加密供审计,完整证据通过验证,缺失部分证据时返回INCONCL

58
Hot 热度
72
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • ClaimReceipt introduces a claim-relative receipt specification and selective verifier that addresses two critical gaps in agent evaluation: evidence sufficiency (whether claims are recomputable from retained evidence) and coverage (whether retained records encompass the full committed experiment set)
  • The system binds typed transaction evidence to a signed experiment manifest, returning one of three verdicts per claim: PASS, INVALID, or INCONCLUSIVE
  • On 1,392 historical buyer-seller records, the CR-2 verifier reproduced all five manually labeled audit verdicts, exactly replayed 600 deterministic and 792 post-generation records, and achieved 11/11 semantic fault detection with zero false positives
  • A prospective CR-3 epoch demonstrated that withholding a single terminal receipt yields INCONCLUSIVE_COVERAGE, while withholding private evidence openings preserves coverage but renders economic claims inconclusive—exactly matching preregistered predictions
  • Receipt instrumentation imposes minimal overhead: 0.021% of model-inference time and 9.9 KB per transaction, though the frozen specification was found not yet unambiguous to independent readers

Why It Matters

Agent evaluation has lacked rigorous evidentiary standards, with generic logs and hash-linked transcripts failing to reliably verify whether claims are genuinely supported by retained evidence. ClaimReceipt provides a formal framework that enables auditors, researchers, and practitioners to cryptographically verify both the sufficiency and completeness of evaluation evidence, addressing a critical trust gap in multi-agent system benchmarking.

Technical Details

  • Claim-relative receipt specification: Binds typed transaction evidence to a signed experiment manifest, enabling per-claim verification with three possible outputs: PASS, INVALID, or INCONCLUSIVE
  • Specification freezing: The specification was cryptographically hashed (SHA-256) before implementation to prevent post-hoc modifications, establishing a reproducible baseline
  • Empirical validation on 1,392 records: CR-2 verifier achieved perfect reproduction of manually labeled audit verdicts, exact replay of all deterministic and post-generation records, non-redundancy of all 13 declared field groups under ablations, and 100% semantic fault detection with no false positives
  • Prospective CR-3 epoch: 30 assignments were committed before inference, terminal receipts were signed and chained, and private evidence was encrypted for auditor access—demonstrating real-world protocol behavior under selective evidence withholding
  • Overhead metrics: Receipt instrumentation adds only 0.021% of model-inference time and 9.9 KB per transaction, making it practically deployable

Industry Insight

  • The minimal overhead (0.021% inference time, 9.9 KB per transaction) suggests ClaimReceipt-style verification can be integrated into existing agent evaluation pipelines without significant performance penalties, making it viable for production benchmarking systems
  • The finding that the frozen specification is not yet unambiguous to independent readers highlights an urgent need for improved specification legibility and standardization in cryptographic evaluation frameworks—researchers should prioritize human-readable, independently verifiable protocols
  • The distinction between coverage verification and economic claim verification (where withholding private openings preserves protocol verification but renders economic claims inconclusive) suggests that agent evaluation frameworks should adopt layered verification strategies, with separate guarantees for structural integrity versus substantive claim validity

TL;DR

  • 提出ClaimReceipt框架,解决Agent评估中证据充分性(claim是否可从保留证据重新计算)和覆盖度(保留记录是否覆盖承诺实验集)两大核心问题
  • 引入claim-relative收据规范和选择性验证器,将类型化交易证据绑定到签名实验清单,返回PASS/INVALID/INCONCLUSIVE三类裁决
  • 在1,392条历史买家-卖家记录上验证:CR-2验证器重现全部5个手动审计裁决,精确重放600个确定性记录和792个后生成记录,11/11语义错误零误报
  • CR-3前瞻性实验实现:30个分配在推理前承诺,终端收据签名链接,私有证据加密供审计,完整证据通过验证,缺失部分证据时返回INCONCLUSIVE_COVERAGE
  • 开销极低:收据仪器仅增加0.021%模型推理时间和每笔交易9.9 KB存储

为什么值得看

Agent评估的可验证性是当前AI安全研究的关键瓶颈,ClaimReceipt首次系统性地解决了"证据是否充分"和"记录是否完整"两个独立但互补的验证问题。该框架为Agent评估提供了可审计、可复现的证据链机制,对构建可信AI系统具有重要参考价值。

技术解析

  • 双问题框架:将Agent评估的证据验证拆分为充分性(sufficiency)和覆盖度(coverage)两个独立维度,通用日志和哈希链接转录无法可靠回答这两个问题
  • ClaimReceipt规范:采用claim-relative收据规范,将类型化交易证据绑定到签名实验清单,支持选择性验证,返回三类裁决(PASS/INVALID/INCONCLUSIVE)
  • 规范冻结机制:在实现前冻结规范(SHA-256 18d109...b81),确保评估标准不可篡改
  • CR-2验证器实验:在1,392条历史买家-卖家记录上验证,重现全部5个手动审计裁决,精确重放600个确定性记录和792个后生成记录,13个声明字段组在消融实验中均非冗余,11/11语义错误零误报
  • CR-3前瞻性实验:30个分配在推理前承诺,终端收据签名链接,私有证据加密供审计员使用;完整证据通过覆盖度和会计验证,缺失一个终端收据返回INCONCLUSIVE_COVERAGE,缺失所有私有密钥时覆盖度验证通过但经济声明无法判定,与预注册预测完全一致
  • 开销分析:收据仪器仅增加0.021%模型推理时间和每笔交易9.9 KB存储,具备实际部署可行性

行业启示

  • Agent评估的可验证性将成为AI安全基础设施的关键组成部分,ClaimReceipt提供的证据链机制为构建可信Agent系统提供了可落地的技术方案
  • 评估框架需同时考虑充分性和覆盖度两个维度,单一维度的验证无法保证评估结果的可靠性
  • 低开销设计(0.021%推理时间、9.9 KB/交易)证明可验证性机制可以在不影响Agent性能的前提下部署,为大规模Agent评估提供了实践路径

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Agent Agent Evaluation 评测 Research 科学研究