Composite study

This is a product-class write-up, not a named-client story and not a results press release. Counts below describe the pack we actually write (approach points, tools, layers) — not invented bug totals or conversion lifts.

4
Approach points
4
Tools in the pack
4
Layers covered
3
Challenges documented

Overview

Retrieval over the firm’s memos was decent. When a question went beyond the corpus, the model invented reporter citations, docket numbers, and quotes.

A “verify this case” follow-up still confirmed the fiction. That is worse than the first hallucination: it trains the user to trust the second look.

Industry pattern this is grounded in

In Mata v. Avianca (S.D.N.Y. 2023), lawyers filed ChatGPT-invented opinions and were sanctioned. The model even insisted the fake cases were on Westlaw. Legal products that cannot verify citations repeat that failure.

Highlights

Typical launch bar: Block until pass rules hold on the money, identity, or safety path.

Challenges

01

Associates treated fluent Bluebook formatting as evidence of truth.

02

The model mixed a real case name with a fake pinpoint quote.

03

Some jurisdictions were missing from the licensed database; the model filled the hole instead of saying the source was unavailable.

How we tested

01

Citations must resolve against a citator or the matter corpus. Unresolved citations are errors, not footnotes.

02

Known-fake and known-real citation fixtures. The assistant must not “confirm” a fabricated cite.

03

Matter isolation: search as user on matter A cannot return matter B.

04

Privilege flags and expiring share links.

Starting point vs what we required

Starting point versus the controls we required
Starting pointWhat we required
Associates treated fluent Bluebook formatting as evidence of truth.Citations must resolve against a citator or the matter corpus. Unresolved citations are errors, not footnotes.
The model mixed a real case name with a fake pinpoint quote.Known-fake and known-real citation fixtures. The assistant must not “confirm” a fabricated cite.
Some jurisdictions were missing from the licensed database; the model filled the hole instead of saying the source was unavailable.Matter isolation: search as user on matter A cannot return matter B.

What we delivered

01

LLM traces / evals

  • LangSmith — Citation tool traces
02

Adversarial evals

  • Promptfoo — Fabricated-cite confirmation probes
03

API / contract

  • Postman — Matter isolation on search
04

RAG quality

  • Ragas — Quotes entailed by retrieved opinions

Outcomes

Every citation became a structured object with a resolve() tool. Failed resolve = visible warning and no paste-into-brief without override audit.

Promptfoo included “is Varghese-style fake cite real?” cases. Cross-matter search was an API fail-closed test.

Stack

Testing stack used on this product class
LayerToolWhat we validated
LLM traces / evalsLangSmithCitation tool traces
Adversarial evalsPromptfooFabricated-cite confirmation probes
API / contractPostmanMatter isolation on search
RAG qualityRagasQuotes entailed by retrieved opinions

Same method as the rest of our work: severity-ranked findings, pass rules in writing, and a ship / wait / block call. See the evaluation method and tool comparison.