Contract
JUDAI

How it works

How a case is analysed

Every figure is read from the corpus as this page loads.

What happens when you ask

  1. 01

    Ask

    255citable units of legislation
  2. 02

    Search the corpus

    256,053passages indexed
  3. 03

    Read the judgments

    3,055judgments in full text42,462 citations read from them
  4. 04

    Verify every quote

    138quoted sentences checked2 answers struck by that check
  5. 05

    Answer, or refuse

    68%of published analyses declined17 / 25
  1. 01

    The corpus

    deterministic

    Judgments are pulled in full text from HUDOC (European Court of Human Rights) and CourtListener (US federal courts). Not summaries, not snippets. Each judgment is split on its own paragraph numbering, so a quote can be traced back to the paragraph it came from.

  2. 02

    Retrieval

    deterministic

    Two searches run over the corpus: keyword and meaning. Their results are merged by position rather than score, so neither can dominate the other by being differently scaled. The case being analysed is excluded, or it would retrieve itself.

  3. 03

    Reading the passages

    model

    A model reads the retrieved passages and scores how far each one actually answers the question, as opposed to merely mentioning the same words. Most of what search returns is discarded here.

  4. 04

    The reasoning

    model

    The issue is identified, the governing rule is stated from the sources, and the rule is applied to these facts. A precedent only counts when the analysis names the concrete fact that makes it apply or not apply.

  5. 05

    The challenge

    model

    A model from a different family attacks the draft: unsupported claim, precedent that does not fit, fact that was ignored, rule stated too broadly. Its objections are published whether or not they were resolved.

  6. 06

    Verification and the decision to answer

    deterministic

    Every quoted sentence is matched literally against the source text. It is the one check that does not depend on any model. Then the evidence is weighed: how deep, how many independent sources, how binding, how current. Below the threshold JUDAI declines rather than guessing.

Why it is allowed to say nothing

A system that always answers is easy to build and impossible to trust. Studies of general-purpose models put hallucination on federal case questions between 58% and 88%; purpose-built legal tools with retrieval still land between 17% and 33%. The abstention rate is not a defect being hidden. It is the mechanism that makes the answers worth reading, and it is published on the same page as the accuracy.

What JUDAI is built on

The law is not one instrument. Neither is an answer.

The corpus
  • European Court of Human Rights2,318 judgments in full text
  • The Convention and its Protocols116 Articles
  • The Rules of Court139 Rules
  • CourtListener737 judgments in full text
  • RECAP378 pleadings
Runs on
  • anthropic/claude-sonnet-5writes the answer
  • openai/text-embedding-3-large1,536 dimensions, the meaning half of search
  • openrouter.aimodel provider
  • PostgreSQL 16.15 · pgvector 0.8.6256,053 passages indexed
  • Next.js 15.5.23 · Node 20.20.2application
Chain
BNB Chainbuilt on

Two searches run over one index: pgvector finds a passage by meaning, Postgres full-text finds it by wording, and the two are merged by rank rather than score, because their scales are not comparable and normalising one to the other would take a constant nobody can justify.

Every quoted sentence is matched character-for-character against the stored source text. It is the one check in the system with no model in it: either the sentence is in the judgment, or it is not.