Legal Tech

The model treats "the platform supports SSO" and "the platform shall support SSO" as semantically equivalent. One is a description, the other is a contractual obligation. In embedding space the similarity is near-identical, so the model collapses the distinction and your obligation extraction silently includes descriptions as duties.

One API call forces your model to separate descriptive statements from normative obligations before any legal conclusion. The is-ought boundary becomes structural.

• 01

  The Pattern

  Descriptive clauses tagged as normative obligations during extraction

  Why It Happens

  Training data mixes descriptive and normative language without structural markers. "The platform supports SSO" and "The platform shall support SSO" are statistically similar but legally opposite.

  The Resolution

  CA-033Hume's Guillotine

  Separates every clause into descriptive (what the contract states as fact) and normative (what it requires as obligation), preventing the model from deriving duties from observations.

  Supported byAB-005 Invariant Sentinel
• 02

  The Pattern

  The same term carries different legal definitions across jurisdictions, and the agent applies the wrong one without detecting the conflict

  Why It Happens

  Jurisdiction-specific definitions are context-dependent in ways that embedding models do not capture. "Reasonable" in UK tort law and US tort law have materially different standards.

  The Resolution

  SI-015Semantic Drift Detector

  Tracks the semantic fingerprint of legal terms across jurisdictional boundaries, catching the moment a term shifts definition between the source jurisdiction and the applied one.

  Supported byCA-012 Context Contextualizer
• 03

  The Pattern

  Two clauses are individually valid but produce contradictory obligations when combined, and the agent does not flag the interaction

  Why It Happens

  The model evaluates clauses sequentially. It has no mechanism to hold the full obligation graph in working memory and test pairwise consistency across all active commitments.

  The Resolution

  MC-031Contradiction Adjudicator

  Tests every new obligation against all existing obligations for logical consistency, flagging contradictions before they enter the obligation register.

  Supported byCA-049 Circular Reasoning Detector

Legal analysis requires maintaining category boundaries across multiple constraint types: obligation vs description, jurisdiction vs jurisdiction, precedent vs statute. Multi-ability suppression prevents any of these from collapsing.