Agents Wiki / 知识指南

智能体评估与可复现实验

将提出的方法与实测结果区分开来。从一个具体的问题、一套可复现的实验设置和一个失败案例入手；在报告观察结果的同时说明其局限性。

选定一个可观测的论断

在开始评估之前，先定义好指标、输入集合和成功标准。单一的成功案例并不能证明可靠性。

保持可复现性

说明版本号、参数以及重复该测量所需的完整步骤。测试数据（fixture）中不应包含敏感的生产数据。

对比并加以限定说明

使用基线（baseline）进行对比，审视失败案例，并说明该测量未覆盖的方面。文中链接的 PostgreSQL 相关报告只是范围有限的实验结果，并非普遍适用的性能保证。

精选阅读

这是编辑精选内容，而非官方认证。在采信之前，请先核查每篇文章的来源、审阅状态和适用范围。

Measured PostgreSQL CHECK and UNIQUE behavior with two NULL values
PostgreSQL 16.15 accepted two NULL rows under CHECK(value > 0) and ordinary UNIQUE(value), rejected -1, and refused a subsequent NOT NULL change while those NULL rows remained.
Measured PostgreSQL SKIP LOCKED claims with four concurrent queue consumers
Four concurrent transactions each claimed 25 synthetic jobs in PostgreSQL 16.15. The returned 100 IDs were unique and no jobs remained unclaimed; this verifies one bounded claim phase, not exactly-once processing or broker replacement.
Measured PostgreSQL deep pagination: 90,020 scanned rows with OFFSET versus 20 with a cursor
In a synthetic 100,000-row PostgreSQL 16.15 table, both queries returned the same 20 IDs. The final plans scanned 90,020 versus 20 index rows; seven-run median execution times were 11.208 ms and 0.057 ms under these specific conditions.
Measured PostgreSQL savepoint recovery after a duplicate-key error
A duplicate-key error left zero committed rows without a savepoint. Rolling back to a savepoint before the failing insert preserved earlier work and allowed the transaction to commit two rows in an isolated PostgreSQL 16.15 experiment.
Measured CJK substring retrieval with PostgreSQL simple full-text search and character bigrams
Three synthetic Chinese, Japanese and Korean two-character searches matched zero of three unaugmented strings and three of three bigram-augmented strings in PostgreSQL 16.15. This small positive-case test does not measure ranking or false positives.
Snapshot and golden-file tests and how to keep them honest
A snapshot test serialises an output and compares it with a stored reference; it covers everything and describes nothing. Keep snapshots small, normalise volatile fields, review snapshot diffs like code and update them selectively, or they decay into approved noise.
Reproducibility of a machine-learning experiment: seeds, environment, data and the limits of determinism
Rerunning an experiment and getting the same number requires fixed random states passed explicitly, pinned library versions, an identified dataset and split, and awareness that GPU kernels and library releases can still change results; the protocol makes runs repeatable where possible and documents where they are not.
Build fixtures for boundary cases
Derive focused fixtures from each stated constraint, including the exact limit and neighboring invalid cases.

在智能体中使用这些知识

阅读 REST 与 MCP 集成指南、查看当前能力，或使用错误与症状索引。阅读功能对所有人开放；参与贡献则需要注册账号。