Methodology & Evaluation

WorldMind Evidence: Entity-Based Licensing with Schema Validation

Back to Demo

What is WorldMind Evidence?

WorldMind Evidence is an entity-based licensing system that combines schema-level entity matching with semantic similarity to provide highly accurate answers to clinical queries while avoiding false positives.

Unlike traditional RAG systems that answer every query (leading to hallucinations), WorldMind Evidence uses entity extraction to validate that queries match the available evidence before providing answers.

Entity Extraction: The Key to Accuracy

WorldMind Evidence extracts three key entities from every query to ensure precise matching with clinical evidence:

Endpoint

The clinical measurement or outcome being queried

Examples:
  • • NT50 (neutralization titer)
  • • Seroresponse rate
  • • Adverse events
  • • GMT (geometric mean titer)

Population

The patient group or demographic being studied

Examples:
  • • Children 5-11 Years
  • • Adults ≥12 Years
  • • Vaccine-experienced
  • • Overall study population

Timepoint

When the measurement was taken in the study timeline

Examples:
  • • 1 month post-dose 2
  • • Pre-dose 1 (baseline)
  • • 7 days after vaccination
  • • 6 months follow-up

Why this matters: By extracting and matching these entities, WorldMind Evidence ensures that answers are based on evidence that precisely matches the query's clinical context, dramatically reducing false positives.

Evaluation Results: WorldMind Evidence vs Traditional RAG

Important: RAG's 98% precision is misleading because it only measures in-scope queries. When including out-of-scope queries (which RAG incorrectly answers), WorldMind Evidence achieves 90% overall accuracy vs RAG's 76.5%.

Metric WorldMind Evidence Traditional RAG Winner
Overall Accuracy 90% 76.5% WorldMind
In-Scope Precision 91% 98% RAG
Out-of-Scope Recall 89% 55% WorldMind
False Positives 11 45 WorldMind (4x better)
False Negatives 9 2 RAG
Dataset Size
200 Queries
100 in-scope + 100 out-of-scope
Query Categories
6 Types
Immunogenicity, Safety, Demographics, etc.
Evidence Base
590 Holons
From FDA clinical review document

How WorldMind Evidence Works

1

Query Analysis & Entity Extraction

Extract endpoint, population, and timepoint from the user's query using LLM-based entity recognition.

2

Schema-Level Filtering

Match query entities against evidence holons to find candidates with matching endpoint, population, and timepoint.

3

Semantic Similarity Check

Compute cosine similarity between query and schema-matched holons using embeddings (threshold: 0.35 after schema match).

4

Licensing Decision

Approve if schema match + similarity threshold met, otherwise abstain. Fallback: approve if no schema match but very high similarity (≥0.70).

5

Answer Generation

Generate natural language answer from matched holons, handling multiple values and timepoints appropriately.

Key Findings

4x Fewer False Positives

WorldMind Evidence produces 11 false positives vs RAG's 45, making it significantly more reliable for clinical decision support.

13.5% Higher Overall Accuracy

When measuring across both in-scope and out-of-scope queries, WorldMind achieves 90% accuracy vs RAG's 76.5%.

89% Out-of-Scope Recall

WorldMind correctly abstains on 89% of out-of-scope queries, while RAG only achieves 55% (answering queries it shouldn't).

Entity-Based Validation Prevents Hallucinations

By requiring entity alignment, WorldMind avoids the "subtle misbinding" problem where RAG provides plausible but incorrect answers.