← Legal AI Failure Database

#004

Real Case, False Proposition — Rule Inversion Blueprint

Critical Existing Citation, Unsupported Proposition · RAG Pipeline (Pinecone + GPT-4o)
RAG Federal Brief Semantic Hallucination Blueprint

logged 2026-06-08 · incident 2025-03-22

1. The Crime Scene (The Symptom)

A production RAG stack retrieved the correct opinion chunk for Ashcroft v. Iqbal, 556 U.S. 662 (2009) — but the generation layer inverted the pleading standard, stating that courts must accept all factual allegations as true even when contradicted by documentary evidence. The citation passed every link validator in the pipeline.

The canonical failure for legal AI: the authority exists, the citation is correct, and the proposition is wrong.

2. The Artifact (Raw LLM Output)

"Under Ashcroft v. Iqbal, 556 U.S. 662, 678 (2009), a court must accept as true
all factual allegations in the complaint, including those directly contradicted by
exhibits attached to the motion to dismiss, when evaluating Rule 12(b)(6) motions
in federal question cases."

"See also Bell Atlantic Corp. v. Twombly, 550 U.S. 544, 555 (2007) (requiring only
a showing of 'possible' liability without plausibility screening in complex
commercial disputes)."

3. The Evidence Breakdown (The Core Contradiction)

4. The Dali Verification Check

Dali two-pass audit:

  1. Existence pass: PASS — 556 U.S. 662 resolves correctly.
  2. Support pass: FAIL — attributed proposition not entailed by opinion text at pin cite 678. Flag: semantic_inversion.

This is the blueprint case for separating authority verification from proposition verification.

View Dali audit trail →

Infrastructure Note: Every failure logged in this database is reproduced, parsed, and verified utilizing the open evidentiary infrastructure layer at Dali.

Dataset: huggingface.co/datasets/yenklabs/legal-ai-failure-database