SShortSingh.
Back to feed

Three-Lane Framework Proposed to Prevent AI from Silently Editing Critical Docs

0
·1 views

A software documentation framework called 'extract, draft, sign' has been proposed to prevent AI models from inadvertently overwriting critical content during README or doc regeneration. The core problem is that current pipelines treat entire markdown files as a single token stream, allowing AI rewrites to silently alter security contacts, SLO numbers, or legal clauses alongside routine prose updates. Under the proposed system, deterministic tools own the 'extract' lane for auto-generated content, AI models are confined to the 'draft' lane for narrative prose, and named humans must approve anything in the 'sign' lane that carries policy or contractual weight. A supporting lanes.yaml config file and directory structure help CI pipelines enforce which writers are permitted in each lane. The framework does not replace legal or audit controls but aims to make the boundary between AI-generated and human-approved content explicit and machine-checkable.

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)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

LumasPDF lets Python developers create signed PDF/A-1b files in 40 lines

A developer and author of the LumasPDF SDK has demonstrated how to generate a digitally signed, PDF/A-1b compliant document using Python in just 40 lines of code. The library is installed via a standard pip wheel and requires no compiled extensions, relying instead on ctypes bindings over a native engine. It supports Windows, Linux, and macOS platforms, with the native engine bundled inside the wheel itself. The workflow handles font embedding, conformance checking, and PKCS#12 certificate-based signing in a single script. The SDK is free to download, though unlicensed output includes an evaluation watermark that a trial key can remove.

0
ProgrammingDEV Community ·

Developer Builds Open-Source MCP Server to Let AI Agents Debug PHP via Xdebug in Docker

A developer frustrated with the limitations of PhpStorm's built-in MCP server has released xdbg, a custom open-source MCP server designed to give AI agents full Xdebug debugging capabilities inside Docker containers. PhpStorm's native MCP server, introduced in version 2025.2, only supports GET requests and lacks header control, CLI path support, and in-container Xdebug management. The xdbg tool enables an AI agent to enable Xdebug, set breakpoints, fire real HTTP requests, step through code, and inspect runtime state autonomously. Available on GitHub under the MIT licence with a one-line install, it is compatible with AI coding clients such as Claude Code, Cursor, and any MCP-capable tool. The project was motivated by real-world debugging failures where LLMs alone produced plausible but incorrect diagnoses, highlighting the need for runtime inspection alongside static code analysis.

0
ProgrammingDEV Community ·

AI-Written Tests Can Pass Without Proving Anything — Here Is the Fix

When AI coding agents submit a patch, they often include tests in the same commit, creating a risk that the tests only validate themselves rather than the new behavior. A technique called the 'revert witness' addresses this by restoring production files to their pre-patch state and rerunning the test suite — if nothing fails, the tests are not actually pinning the change. Using a small tag-parser as an example, the article shows how a weak test suite stays green after reverting the patch, while a properly written suite fails on the reverted code, confirming it genuinely captures the new contract. A short Bash script, revert_witness.sh, automates this check by reading a configurable base commit and temporarily restoring original production files before running tests. The author argues this revert-witness step should be a mandatory part of both local staging and CI pipelines as AI-generated pull requests become increasingly common.