Skip to content
All proof of work
Risk and governance · Legal

Taking hallucination risk out of legal filings

A stoplight triage process that keeps a lawyer as the final decision-maker while cutting the volume they have to re-read.

Citations triaged per filing review
51
Risk escalation: low, moderate, high
3-tier
Final decision-maker, by design
Human

The situation

A lawyer certifies every filing under Rule 11 and state bar rules. That means citations cannot be automated away. The accountability is personal and non-delegable. But re-reading every citation in every brief is exactly the kind of work that expands to fill available time.

What we did

We built a stoplight process that scores every citation by risk and lets a paralegal or attorney focus on the passages carrying the most exposure instead of re-reading everything uniformly.

The architecture uses low-cost models to do the initial assessment and spends expensive model calls only where the risk score justifies it, escalating the riskiest citations to a deeper multi-model review tier. A panel design across multiple models produces an accuracy number instead of a single opinion.

We delivered the proof-of-concept; the client’s own development team is building the production version, which is the handoff we prefer.

Why it is the pattern

This is our template for regulated work generally. Do not remove the human from the decision. Remove the undifferentiated reading that stops them from spending attention where it counts, and make the escalation logic explicit enough to defend in a review.

Capabilities
  • Human as final decision-maker
  • Multi-LLM panel architecture
  • Risk-scored escalation to control model cost
  • AI triage of technical and regulatory content
  • Proof-of-concept to client-owned production handoff
More proof
Next step

Want the detail behind this one?

We are happy to walk through the architecture, the decisions, and what we would do differently now.