Roadmap H2 2026: Validation and Harnesses

As people who have heard me talk about it know, §Blawx is a project that I spend about 3 hours a week working on, usually Saturday mornings. This morning, the work that needed to be done was to come up with a plan for what's next, and I wanted to share where my head is at.

Agentic & Symbolic AI for Rules as Code is a "Thing"

As of now, §Blawx is a rules as code tool that can be used by non-programmer human experts, and by agentic AI. The AI can help write the code that normal people can still read, can help test the code, and can help run the code to accomplish things. That is an objective that I have had in mind since leaving the Government of Canada and starting development of v2 nearly three years ago, so it's very gratifying to see that come to fruition.

This LinkedIn post by Marty Perron is a great example of how far we have come. He illustrates generating a §Blawx encoding and using it to generate a gc.ca style eligibility web app in about 15 minutes.

What Got Left Out: Validation

When designing v2 of §Blawx, I anticipated a number of ways in which a new architecture could support AI-assisted drafting and validation. But because I was limited to only a few hours a week - and because validation of code that never gets written has no value - my focus has been squarely on the portions of that design that were necessary in order to facilitate AI-assisted drafting and testing. Validation got left for later.

With how far we have come in getting AI assistance in the drafting of the code, I think "later" is now. It's time to demonstrate not only that the combination of §Blawx and agentic AI makes it easier than ever to create formal legal specifications, but that it possible to prove that the encodings are correct.

That requires some infrastructure that hasn't been created yet. We need to have users with different roles on projects. We need validation workflows where people create sets of questions and associated fact scenarios, and acceptability criteria. Acceptability criteria may even require a new language for testing the content of §Blawx explanations. We need to document when code has and has not been approved, by whom, when, based on what versions of the source material, and expose that information to users of the encoding. We need automated workflows that notify people when changes to the encoding have changed approved outputs, so that the new outputs can be re-validated. We need user interfaces designed specifically for the validation task, similar to the scenario editor in v1, so that no one needs to learn any block-based coding language in order to generate fact scenarios.

So I think the focus on development in H2 2026 is going to be validation. And because that is a larger architectural undertaking than I have been working on in H1, the frequency of feature updates is probably going to go down.

At the Same Time, Time is Short, so Harness Time.

It's trite to say, at this point, but software development really has changed monumentally in the last few years. And things are in the process of shifting again. Pre-AI, developers' output was code. Then with AI it was often instructions about how to code. Now with agentic AI developers have the opportunity to create "harnesses" that generate the instructions and provide the resources that autonomous agents use to generate the code.

Being able to use my limited time on higher-level tasks and delegate more work to more agents would be a force multiplier for §Blawx development. But it requires me to learn new skills, which means to learn to go faster, I need to go slower first. So thanks in advance for your patience while the pace of visible change slows. I'm optimistic that in the long run it will be worth it.

Happy coding, Blawxers.🥊