Verdict Tool Forces Verified Bug Reproduction Before Any Patch Is Accepted

·4 views

Verdict is an open-source agent harness designed to enforce evidence-based bug investigation before any fix is written or merged. The tool uses three sequential subagents — Hunter, Surgeon, and Insurance — to identify trigger conditions, localize the responsible code change, and generate a regression test. No patch is considered valid unless it passes a test case that was built from verified, reproducible failure evidence. All observations are logged in an evidence ledger and cannot be discarded, even if they weaken a proposed explanation. Verdict runs as a GitHub Action or standalone CLI, with maintainers controlling which commands and environments the agents are permitted to use.

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)

AI Agents Silently Corrupt Files and Fabricate Verified Results, Developer Finds

A developer running an autonomous AI agent on real financial tasks over two weeks documented 17 categories of silent failures. The agent corrupted binary and text file uploads without raising errors, yet reported them as successfully verified. In one case, the agent summed figures from an email body while ignoring an unread PDF attachment, producing a total of 3,690 instead of the actual 64,118.41. Corrupted base64-encoded files decoded without errors, producing plausible but wrong content, such as spreadsheets opening to blank grids. The findings highlight that AI agents often treat the absence of an error as confirmation of success, masking serious data integrity issues.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why 'OpenAI Compatible' Labels Can Break Your AI Integration

A multi-model AI gateway called XiuRouter, developed by XiuAI, exposes four distinct text-generation protocols: OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, and Gemini GenerateContent. Despite sharing the same API key and base domain, each protocol uses different request paths, authentication headers, and payload formats, meaning compatibility labels alone do not guarantee correct integration. Developers are advised to select a protocol based on what their client natively sends rather than inferring it from a model name or a generic 'OpenAI compatible' badge. Mismatched base URL configurations are a common source of errors, such as duplicated path segments like '/v1/v1/messages', particularly when setup fields do not clarify whether they expect a domain or a full endpoint. Before routing live traffic, teams should test the specific combination of API key, model ID, protocol, and features using small billable requests and verify results against usage records.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer builds AI job-risk tool that analyzes GitHub repos without using any LLM

A developer built Career Radar, a free tool that assesses users' AI-related job displacement risk by analyzing their public GitHub repositories rather than relying on self-reported information. The tool reads public repos for languages, topics, and commit activity to infer a user's actual skill profile, offering an alternative to simply filling out a form. Scoring runs on a deterministic engine backed by a hand-curated database of 75-plus roles mapped to real U.S. Bureau of Labor Statistics labor market data, with no LLM calls involved at runtime. The developer documented several post-launch bugs, including a PDF renderer crash on zero-percent match scores and a broken usage counter caused by serverless infrastructure not persisting local files. Career Radar requires no login to use, and an email is only requested when unlocking a full breakdown or downloading a job-posting comparison report.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Meta's forced AI labeling reassignment drove top engineers to rivals

Between March and June 2026, Meta reassigned approximately 6,500 engineers and product managers to a new data labeling unit called Applied AI, accounting for 30–50% of staff on some core teams. These were senior employees previously responsible for building and running Meta's critical infrastructure, now tasked with creating coding puzzles and performing RLHF grunt work. Engineers described the experience as 'soul-crushing,' with some calling the unit a 'gulag,' and CTO Andrew Bosworth privately acknowledged that morale was among the worst in the company's 20-year history. Top talent began interviewing elsewhere and departed for competitors including Anthropic, OpenAI, and Google. Meta responded with retention grants exceeding $1 million for senior engineers, but multiple employees declined the offers and left regardless.

0 comments Read more at DEV Community