GraphCURE
Constrained Expert Routing with Asymmetric NEI Deferral for Evidence-Based Fact Verification
A constrained expert-routing framework for evidence-grounded fact verification that coordinates a direct verdict expert and a complementary expert via Asymmetric Not-Enough-Information (NEI) Deferral, improving validation Macro-F1 from 0.6920 to 0.7112 (+1.92 points, paired-bootstrap P(Δ>0) = 0.9975).
The Technical Bottleneck
In 3-way fact verification (Supported, Refuted, Not Enough Information / NEI), single classifiers suffer from asymmetric boundary confusion between conflicting evidence and insufficient evidence (NEI).
Why Existing Baselines Fail
Naive ensemble averaging blurs confident Refuted/Supported predictions with uncertain NEI boundaries. Constrained expert routing selectively defers ambiguous boundary instances to a specialized complementary expert.
Core Formulation & Insight
Coordinates a primary direct verdict expert with a complementary expert governed by an Asymmetric NEI Deferral policy, evaluated rigorously with paired-bootstrap resampling.
Architectural Design
Two-expert routing architecture with constrained deferral thresholds tailored to the asymmetric structure of the NEI class.
1. Direct Verdict Expert
Produces base probability distributions over Supported, Refuted, and NEI given retrieved evidence.
2. Asymmetric NEI Deferral Gate
Identifies high-uncertainty boundary cases around the NEI decision region and routes them to a complementary expert.
3. Paired-Bootstrap Validation
Quantifies Macro-F1 gains on validation and frozen test splits with 10,000 bootstrap resamples.
Technical Stack & Systems Engineering
Implementation Highlights
- →Validated constrained routing improvements using paired-bootstrap probability P(Δ > 0) = 0.9975.
Datasets & Protocol
Datasets & Corpora
- Evidence-Based Fact Verification Validation & Frozen Test Splits
Evaluation Protocol
Evaluated direct expert baseline against constrained Asymmetric NEI Deferral routing.
Tracked Evaluation Metrics
Improved validation Macro-F1 from 0.6920 to 0.7112 (+1.92 points, P(Δ>0) = 0.9975) with positive frozen-test gains (+0.00977 / +0.00943 MF1).
+1.92 pp improvement via Asymmetric NEI Deferral
High statistical reliability on validation split
Consistent positive transfer on held-out frozen test sets
Empirical Analysis & Interpretation
Selective deferral around the NEI boundary resolves systematic over-prediction without destabilizing high-confidence Supported/Refuted verdicts.
What This System Contributes
- 01Formulated the Asymmetric NEI Deferral routing mechanism and statistical bootstrap evaluation.
Boundary Conditions & Future Work
- ·Repository is currently private; metrics are reported from verified research records in the portfolio audit.