what is pawl fr?

Pawl is a verification framework meant to keep AI coding agents honest across long or repeated sessions. The problem it addresses is that a coding agent starting a new chat has no memory of what was already tried, what already failed, and what was actually finished versus just claimed as done. Pawl solves this by turning every meaningful change into a claim with success criteria that get written down before any code is implemented, then tracking the outcome of that claim in a CSV registry file rather than in chat history or planning documents. The core idea is borrowed from a ratchet mechanism, the small part called a pawl that lets a wheel turn forward but never slip backward. In this project, that means agent work can only move ahead on verified progress, and once something is confirmed to have failed, that outcome, called an honest negative, is recorded permanently so future sessions do not waste time retrying the same dead end. When a claim fails, any other claim that depended on it gets automatically flagged for review. Pawl ships as a set of plain Python scripts with no external dependencies, plus templates and documentation describing the workflow. To adopt it, a project copies the pawl folder in and runs a script to set up a registry file and configuration. From there, a developer or agent registers a claim describing a measurable outcome, creates a folder for that claim's work, implements the feature, and then runs an official checker script that writes back either Verified or Failed to the registry. A separate memory tool lets an agent search past claims and plans using simple text search. The project comes out of Agentic Commerce Lab, an initiative from the company Shopware focused on AI agent tooling for commerce, and was first tested on an internal project that ran nearly six hundred claims through it over about two and a half weeks using several parallel agent sessions.