This document is a draft shared for review and comment. Enter the collaborator code you received to continue.
Take text describing a person's situation and return the codes from a shared legal issue taxonomy that say what kind of problem it is.
Here is where each shared asset stands for this tool. The functional agenda and the conformance standard are drafted and open for comment. The test sets are the missing piece, and they are what the cohort is forming to build.
What a classifier has to do, from accepting a three-word phrase through returning a code path with confidence, urgency, and a declared taxonomy version. Three scope tiers let a team see only the capabilities that apply to the system they are building.
Read the agenda →Eleven binary gates a system cannot fail, then scored dimensions under the five Commons gates, then a mapping from composite score to what the system is allowed to do. Every numeric threshold is blank until the first evaluation round produces a distribution.
Read the standard →The suite needs eight datasets and most of them do not exist yet. Taxonomy example prompts partly exist. Real content items, real user questions, sparse inputs, multi-issue inputs, and the safety set all need building and annotating. Two annotators per set, with agreement reported using Gwet's AC1.
Join the cohort to contribute →A working classifier built to the agenda, with the code open for another team to adopt. Ensemble approaches that vote across several models are the current best practice in this space, and the reference build should show one.
Coming as the cohort formsHow to build one: how to give a model the taxonomy without giving it all 1,124 codes at once, how to enforce path validity in code rather than in a prompt, and how to decide the depth an input actually supports.
Coming as the cohort formsCase studies from teams running classification in production. What they measured, where it failed, and what it costs to keep a classifier current as the taxonomy changes.
Coming as the cohort formsThe functional agenda defines what the system must do. The conformance standard turns those capabilities into tests. The test sets are what the tests run against, and without them the standard is a structure with nothing to measure.
You do not need all of them. Pick the resources that match your role.
A legal issue classifier takes text describing a person's situation and returns one or more codes from a shared legal issue taxonomy. The text can be a sentence someone typed into a legal help website, a paragraph a call handler wrote down, the title of an article, a row in a service directory, or a full intake narrative.
Almost every other tool in the Commons depends on this one. Referral routing needs to know the issue before it can match a service. Content tagging needs it to connect an article to the people who need it. Triage needs it before it can suggest a next step. When the classifier is wrong, everything downstream is confidently wrong in the same direction.
The reference taxonomy is LIST, maintained at taxonomy.legal. It runs five levels deep, and the current public export holds 1,124 codes.
Deciding that a problem is about housing is easy. Deciding whether it is a repair issue or a habitability issue, from an input of nine words, is hard, and a system that guesses at that depth sends people to the wrong page while reporting high confidence. The agenda makes returning a shallow code a requirement whenever the input supports nothing deeper.
Who this serves: Legal help websites tagging their content, hotlines routing callers, referral systems matching people to services, and researchers who need consistent issue codes across organizations.
The cohort is collecting this. If you are classifying issues today, whether with a model, a keyword list, or staff reading each item by hand, tell us. The systems already in use are the ones the standard needs to be tested against.
These are unresolved and marked so nobody mistakes them for settled. The full lists sit at the end of the agenda and the standard.
The Legal Issue Classifier working group is forming now. Members settle the open questions, build and annotate the test sets, and run the standard against the systems they already use.