correctover-patronus Adds 6-Dimensional Structural Verification to Patronus AI

·1 views

A new open-source adapter called correctover-patronus has been released to extend Patronus AI's LLM evaluation capabilities beyond hallucination and toxicity detection. The tool integrates Correctover's 87 deterministic verification rules as native Patronus evaluators, covering six dimensions: structure, schema, identity, integrity, latency, and cost. Unlike standard LLM evaluators, it targets structural failures such as malformed JSON, missing required fields, and token budget overruns. Every evaluation generates a recomputable proof hash tied to the input, output, and applied rules, ensuring fully transparent and reproducible verdicts. The adapter runs locally with no external API calls, achieving a median verification latency of 22 microseconds, and is available via pip install correctover-patronus.

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)

Claude Sonnet 5 matches Opus performance at lower cost, now available via API

Anthropic has launched Claude Sonnet 5, which matches Opus 4.8 on coding and agentic benchmarks while carrying the same pricing as its predecessor, Sonnet 4.6, at $2/$10 per million input/output tokens. The launch rate is available through August 31, after which pricing steps up to $3/$15 per million tokens. The model replaces Sonnet 4.6 as the default reasoning tier across Anthropic's plans and brings improved long-context handling, document parsing, and multi-step agentic task completion. Early users report that agent workflows that previously stalled mid-loop are now completing end-to-end, representing a reliability improvement beyond incremental benchmark gains. Sonnet 5 is accessible via the Anthropic API and Vercel AI Gateway, with Anthropic describing the migration from Sonnet 4.6 as a minimal, low-risk change.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer builds TypeScript compiler graph MCP tool, cuts Claude Code token use tenfold

A developer has released @ttsc/graph, an open-source Model Context Protocol tool that gives AI coding agents a pre-built index of a TypeScript codebase derived directly from the TypeScript compiler. Instead of returning raw source code, the tool provides function names, call edges, type signatures, and exact file-line coordinates, allowing agents to answer structural questions without crawling through files. Benchmarks show the tool achieves roughly 10 times fewer tokens on open-ended codebase questions compared to a grep-based baseline, with answer quality remaining comparable. Existing tools such as codegraph, codebase-memory-mcp, and serena were evaluated but found inconsistent across repositories of varying sizes. The project is publicly available on GitHub, with full benchmark methodology and per-repo results published on the project's documentation site.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Guide Shows How to Sync AI Agent Memory Files With Obsidian for Easy Editing

A developer tutorial published on DEV Community explains how to make the long-term memory files of Hermes Agent, an open-source AI agent released in February 2026 by Nous Research, editable through Obsidian, a markdown-based note-taking app. Hermes Agent stores its memory across sessions in plain-text files using a § character as a delimiter, a format that is difficult to edit manually and offers no version history. The proposed solution uses a single ~400-line Python script, relying only on the standard library, to sync these memory files into cleanly formatted Obsidian markdown notes with YAML frontmatter. Python was chosen over bash because the multi-byte § delimiter can be mishandled by bash string operations, whereas Python's re.split() processes it reliably. The system also incorporates automated Git commits every six hours, giving users a full change history and a safety net against accidental data loss.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why AI Teams Need a Metrics Baseline Before Scaling Any Feature

Software teams building AI features often struggle to evaluate whether those features are actually working once usage scales up. A metrics baseline provides a small set of before-and-after measurements to determine if an AI workflow is improving, degrading, or simply becoming more costly. Unlike generic software tracking, AI features require additional signals because model outputs are probabilistic and can be fluent yet wrong, correct but incomplete, or useful but prohibitively expensive. Key baseline categories include cost per successful task, output quality, latency, user adoption, and real-world task improvement. Experts recommend starting with just one or two metrics per category tailored to the feature's specific risk and purpose, rather than building sprawling dashboards that obscure decision-making.

0 comments Read more at DEV Community