OpenClawClaude
CursorGeminiGitHub CopilotPerplexityMistral AIDeepSeekOllamaQwenTHE PROCESS WE’RE BUILDING
Keep the agent’s work connected to what you asked for. Here’s the workflow we’re working toward.
Write down what the agent should build and what matters most. The intended workflow starts with requirements you approve, so there’s a shared definition of done.
We’re developing a way to turn those requirements into checks, challenge the checks, and compare the agent’s work with the plan. The goal is to make missed requirements easier to see.
A useful result should explain what matched the plan and what still needs attention. We want the evidence to guide the next fix, so the agent has a clearer place to start.
What we’re designing for.
Build with the coding tools you already use.
Give every check a written requirement.
Compare expected and observed responses.
Trace a decision back to its requirement.
Inspect the method behind the product.
The method, reported benchmark results, and limits of the first verification protocol.
Read the working paper 1 October 2026 · Not peer reviewedOpenClawClaudeCursorGeminiGitHub CopilotPerplexityMistral AIDeepSeekOllamaQwenHave questions? We’re here to help.
No. Your agent still writes the code. Demaloa checks the finished work against the requirements you approved.
Demaloa turns written requirements into executable checks, then tests those checks against deliberately broken code before using them as a gate.
The release pauses. You see the requirement, the observed behavior, and the evidence needed to reproduce the failure.
Yes. The public verification study includes the method, results, and limitations. It is linked from the Study page.
The goal is fewer retries and more focused fixes. Savings depend on your tasks and workflow; we do not promise a fixed reduction.
Tell us what your agents build and share a requirement that keeps getting missed. We’ll start with that task.
Tell us what you’re building. We’ll help you find a place to start.
Get in touch