testing

replayd is an open source Python tool that captures failed AI agent runs and replays them as regression tests to prevent regressions from returning after changes.

Discusses how to benchmark and grade production builds, focusing on key performance indicators like context-drift, hallucinations, and governance.

Blue Origin's New Glenn rocket exploded during a hotfire test at Cape Canaveral, marking a significant setback. All personnel are safe, and an investigation is underway.

The Bot Company, a $2 billion startup founded by Tesla and Cruise alums, is accused of secretly testing household robots in Airbnbs, causing extensive damage; a host is suing for $12,383.50.

A web page that measures keyboard latency via reaction time and tap duration tests, allowing users to submit results for comparison.

This paper introduces LGMT, a framework that uses first-order logic to generate semantically invariant test cases for evaluating LLM reasoning reliability. Experiments on six LLMs show that LGMT exposes hidden defects missed by static benchmarks, suggesting evaluation should focus on robustness under logical invariance.

This paper introduces CAFD, a learning-based approach for DNN fault detection that integrates model-based, distance-based, and a novel concept-based feature called Concept Failure Ratio (CFR) derived from Vision-Language Models. CAFD consistently outperforms state-of-the-art baselines in fault detection rate across multiple datasets and budgets.

Anubis OSS, an Apple Silicon Mac app for benchmarking local LLMs, now supports direct model downloads from the UI via a 'Browse Models' button that pulls from ollama.com library. The developer is seeking testers to confirm installation and functionality.

Despite Anthropic stating that its powerful Mythos model would remain restricted, users have repeatedly spotted 'Mythos 1' in the Claude Code interface and in source code, suggesting an unannounced staged rollout or quiet production testing.

datasette-fixtures 0.1a0 是一个新插件，利用 Datasette 1.0a30 新增的 fixture 数据库 API，方便插件测试。可通过 uvx 快速试用，内置示例数据。

CadX Studio teases an AI CAD model for 2026, demonstrating a workflow from typing to building to sectioning, with a model release planned for Monday.

A comparison of 8 AI voice agents for dental clinic workflows, highlighting performance in latency, interruption handling, and integration.

Matt Pocock proposes documenting agreed test seams in apps, arguing that AI agents cannot be trusted to make good testing decisions, often leading to fragile tests that break on implementation changes.

A developer created LLM Canary, an open-source quiz program that sends randomized tasks to multiple LLMs to track performance over time. After a week of hourly testing across seven models, the results show all models fluctuate throughout the day with no consistent pattern, and no clear evidence of degradation was found.

A developer built a zero-code visual MCP client within AgentSwarms that allows testing remote MCP servers directly in the browser, demonstrated with Cloudflare's free MCP server for documentation.

ARK is an open-source Go runtime that governs AI agent decisions, compiles and tests generated code before delivery, featuring a 6-phase verification pipeline and cost-efficient model routing.

dari-docs is a CLI tool that tests documentation quality by simulating AI agents performing tasks, identifying where agents get stuck, and optionally generating proposed edits to improve doc clarity.

Two skills for AI coding agents that design and run claim-driven tests for distributed and stateful systems, producing structured test plans and findings reports with 9-state verdicts and blame classification.

SpaceX has moved the Starship and Super Heavy V3 to the launch pad at Starbase for final testing and launch preparations.

Two Sigma has open-sourced four tools—Flint, BeakerX, Marbles, and Cook—used by major tech companies like Twitter, Apple, and Indeed, covering time-series analysis, multi-language notebooks, readable test failures, and batch job scheduling.