Kill-Matrix Pruning Finds Golden Test Cases That Never Catch C++ AI Patch Bugs

·2 views

An AI patch pipeline team discovered its nightly suite of 214 golden cases missed a production regression despite a green dashboard. Instrumenting the suite showed 61 cases had never failed on any scored patch, while 33 failed on nearly every patch and added little signal. The author proposes a kill matrix: a set of hand-seeded or tool-generated known-bad C++ mutants run against each golden case to measure which cases actually detect defects. Cases are then classified as discriminating, redundant, never-firing, or always-firing, with each category suggesting a specific action. A greedy set-cover algorithm can finally reduce the suite to a minimal set that still kills the same mutants.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Why building for imagined future needs backfires: the YAGNI principle explained

A freelance developer in Paris spent a weekend adding multi-currency support with live exchange rates, four languages, and exports to three accounting tools. None of these features was ever used, while the exchange-rate API's changed terms eventually broke the invoices page. The author compares this to packing three suitcases for a beach weekend, only to spend it in a swimsuit. The YAGNI principle from Extreme Programming advises building only what is needed now, since future needs rarely match what was predicted. The article illustrates the approach with code examples, including an unused currencies table and a rate-conversion service.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Regulation D filings offer early public signal of private company fundraising

Companies raising money privately under Regulation D must file a Form D with the SEC within 15 days of their first sale, often before any press release announces the raise. These filings are structured XML that disclose details such as the company's address, phone number, industry, and offering and sold amounts, along with executive and director names. The article explains how to query SEC EDGAR full-text search, identifying the requester with a User-Agent header and staying under the 10-requests-per-second limit. It also notes data quirks, including fund filers, non-numeric offering amounts, zero-sold fresh filings, amendments, and inconsistent phone formats that require cleanup.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

DashUtils swaps FFmpeg.wasm for WebCodecs, cutting video compression from 97 s to 5.6 s

The developer of DashUtils, a site with 392 free browser-based file tools, replaced FFmpeg.wasm with the browser's native WebCodecs API for video and audio work. The FFmpeg engine had a large download, single-threaded processing, and no hardware acceleration, which made some jobs very slow. Mediabunny handles container reading and writing, with FFmpeg kept as a fallback for formats browsers cannot read. An initial 97-second compression time was traced to Chrome on Linux lacking AAC encoding, which the developer fixed by adding a small WebAssembly AAC encoder, bringing the time to 5.6 seconds. The article also describes throttling progress updates to reduce UI lag and adjusting resolution to meet file-size targets.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Weekly Reviews Work Better as Explanations Than Grades

Reflection is most useful when it explains what happened during a week rather than scoring it, since a list of completed tasks can hide postponed priorities and a sparse week may still have involved important progress. A useful review looks at the conditions behind progress, such as vague next steps, fragmented time, or urgent interruptions, because each calls for a different fix. Writing down concrete events makes this clearer than a mood-based summary, and vague labels like lack of discipline rarely point to a change. Reflection should not become endless analysis, so a few honest observations leading to one narrow experiment for the following week is usually enough. The strongest reviews treat the week as a feedback loop of describing events, identifying a plausible cause, and making one adjustment, while acknowledging when circumstances were beyond one's control.

0 comments Read more at DEV Community