what is evidence-first-agent-workflow fr?

Evidence-First Agent Workflow is a governance system for AI coding agents, built for situations where verified facts matter more than guesses: small models, closed networks, long lived projects, and stacks where correctness counts more than speed. The project is currently in an alpha stage, and the author is inviting public feedback rather than claiming the system works perfectly on its own. The project has two parts. A generic core manages memory, routes context to the right place, and forces every task through a five stage process: Analysis, Task, Todo and Micro-Verify, Checklist, and Verification. If a step fails, the agent has to go back and redo it rather than move forward. The second part is a stack pack, a set of facts and rules specific to whatever tech stack is being used, for example React with ASP.NET Core, since the generic core cannot invent things like which SDK version is supported or what the internal conventions are. The repository tracks project state through a few files: a current state, a history of past changes, and a temporary worklog for work still in progress. This keeps an AI agent's context accurate across long sessions instead of relying on it remembering everything on its own. A core idea in the project is that a model saying a task passed is not proof by itself. Instead, the workflow relies on deterministic checks: validating the stack setup, checking that code changes match documentation, scanning build logs, and comparing rendered screens against reference images for visual correctness. It also checks color contrast and runs browser based end to end tests using Playwright. The tool works alongside several existing AI coding assistants, including Codex, Roo Code, Zoo Code, Cline, and Claude Code, connecting them to the same shared instructions file rather than requiring separate setup for each. The rules are enforced through automated checks at commit time and in continuous integration, not by simply asking the agent to follow instructions. It is released under the MIT License, so it can be used freely, including for commercial and closed source projects, as long as the copyright notice stays intact.