<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>Measurement-Contract on 「开源之道」</title><link>https://www.opensourceway.blog/tags/measurement-contract/</link><description>Recent content in Measurement-Contract on 「开源之道」</description><generator>Hugo</generator><language>zh-CN</language><copyright>Copyright (c) 2016 - 2026, 「开源之道」·适兕; all rights reserved.</copyright><lastBuildDate>Wed, 21 Oct 2026 04:31:13 +0800</lastBuildDate><atom:link href="https://www.opensourceway.blog/tags/measurement-contract/index.xml" rel="self" type="application/rss+xml"/><item><title>2026-10-21 「开源之道」·论文略读：Agent 优化了什么，又交代了什么？——1,130 对 1,130 的度量契约审计</title><link>https://www.opensourceway.blog/posts/osbook-book-recommendation/agent-perf-validation-reporting-2026-10-21/</link><pubDate>Wed, 21 Oct 2026 04:31:13 +0800</pubDate><guid>https://www.opensourceway.blog/posts/osbook-book-recommendation/agent-perf-validation-reporting-2026-10-21/</guid><description>Peng、Calvo、Kalu 与 Davis（Purdue Duality Lab, arXiv 2610.03969）用 1,130 份 agent 性能 PR 对 1,130 份人类性能 PR 做分层配平对照：合并率 54.4% vs 73.3%，但验证报告率 82.3% vs 80.6% 无统计差异。真正的差距在后面——有验证证据的 PR 里约一半根本没有量化指标，缓存类 PR 只有 17.7% 交代了权衡。这不是 agent 稀释质量，是开源从未为「性能主张」建立度量契约，而 agent 把这处空位照了出来。</description></item></channel></rss>