Here's some promising results: under the hood a GPT-powered agent chose a strategy of 1. Collecting an ontology from a Blawx endpoint, 2. Encoding facts in that ontology, 3. Sending those facts to the "interview" endpoint, and 4. Generating a natural language statement of the answer. The interface still looks like ChatGPT, but by giving it the ability to use Blawx as a tool to solve problems, there is no risk of hallucination, and the reasons are explainable. #RulesAsCode #LawFedi #LegalTech
Jason Morris
- 184 Followers
- 125 Following
- 559 Posts
Rules as Code guy. Lawyer, coder, Blawx dev, He/Him
Today, I have a version of lawMirror that can generate all the same XML elements as CLEAN, including arbitrarily nested spans with custom ids, headings, and sandwich text. ProseMirror doesn't "do" nested inline elements, so the next question to answer is whether we can use some CSS magic to make the XML look right without breaking how navigation works. Then it's just a matter of cleaning up the UI, which will be a lot of work, but there are no roadblocks, there. #LegalTech #RulesAsCode #LawFedi
So I'm experimenting with a new feature for Blawx "tests" called an "agenda." It's in a separate coding environment, with a different semantics, but it re-uses the categories and attributes and relationships from your Blawx code.
The implementation-independent features I can think of involve questions, loops, and questions with sub-loops, plus phases. Here's what the current rough prototype looks like:
5/
Prototype WYSIWYM web editor for #AkomaNtoso (aka #LegalDocML): https://github.com/Lexpedite/lawmirror
think "google docs for laws"
Please boost, and feedback welcome.
pinging @verbman
There you go. Trustworthy legal AI. Blawx makes sure that the answer is correct, and we use ChatGPT to make sure the answer is legible.