two much labs / tenthman

Lab project 008 · research · not legal advice

Preserve the rule. Preserve the exception.

A legal answer can quote the right document and still miss the exception that changes the outcome. Tenthman is research into fast models that turn legal text into explicit structure, so an AI's answer can be checked against it while it is being written.

Where it stands Talk to us

An invented ruleA permit is required unless the work is emergency repair. The exemption does not apply to protected buildings.

  1. RuleA permit is required
  2. Exception · unlessThe work is emergency repair
  3. Limit on the exceptionNot for protected buildings

Finding the sentence is the easy part. The answer is only right if the exception, and the limit on the exception, stay attached to the rule they change.

/what

Structure, not just phrases

Recognising individual phrases is not enough. Tenthman reads a provision into the parts that decide the outcome, and keeps track of how they connect.

01

Rules

What the provision requires, permits or forbids.

02

Conditions

When it applies, including nested AND and OR branches.

03

Exceptions

What switches it off, and the exceptions to those exceptions.

04

Consequences

The effect once the conditions are met.

05

References

Links to other provisions the meaning depends on.

/status

Where it stands

An early research prototype. We publish what works and what doesn't, and failed approaches stay on the record.

  • Annotation converter

    Converts published legal annotations into full structure without losing branches, grouping or effects. Mechanically checked.

    working
  • Small learned parser

    A small fine-tuned model that reads a provision and predicts its structure with source spans. Fast, but not yet accurate enough to rely on.

    accuracy not there yet
  • Evaluation harness

    Compares structures against annotations and general-purpose models, inspects exception failures and measures speed.

    working
  • Constrained generation

    Using verified structure to restrict an answer while it is being generated.

    research goal
  • Independent legal review

    Whether the structure preserves legal meaning, judged by qualified reviewers.

    still needed
What we are not claiming. Tenthman does not yet beat general-purpose models at reconstructing full legal structure, we have not shown zero hallucinations, and we claim no novel algorithm. We'll publish numbers once there is a fresh, independently reviewed test to run them on.

Honest limits

What it doesn't do yet

  • It is a research project, not a legal advice product.
  • Full relationships between clauses are still the main failure, especially on longer provisions.
  • Our conversions and corrections have not been certified by legal experts.
  • Cross-references that need context outside the provision are not handled.

/next

What comes next

  1. Build a small, independently reviewed set of complete structures with interacting exceptions.
  2. Fix missing and duplicated clauses before making the model any bigger.
  3. Test on new, untouched provisions, with experts adjudicating alternative readings.
  4. Only then connect the parser to generation, and measure correctness, abstention, speed and cost together.

Get involved

Read a provision with us

The most useful help right now is legal expertise: reviewing whether a structure keeps the meaning of the text it came from.

Offer a review