2.3 KiB
III. Methodology
A. Dataset
The foundation of our model rests on a curated dataset of:
- Anonymized insurance claim transcripts
- Internal emails between adjusters and claimants
- Call center logs with escalation flags
This dataset includes a balanced mixture of confirmed fraudulent claims and validated legitimate cases, used to both train and test our recursive linguistic model. Importantly, each data source is processed through a human-AI recursive review loop, where human analysts verify and adjust the resonance scores generated by our models — ensuring that subjectivity and nuance are preserved while expanding analytic scale.
Every claim is not merely analyzed. It is recursively heard.
B. Analytical Tools
To detect subtle patterns of deceptive intent, we apply an ensemble of forensic NLP methods:
- NLP-based Pattern Extraction: Identifies clusters of linguistic anomalies across claim timelines.
- Sentiment Trajectory Mapping: Tracks emotional evolution of narratives; distinguishes authentic distress from strategic affect.
- Syntax Entropy & Disfluency Detection: Measures irregularities in syntactic flow, hesitation markers, and repair sequences.
- "Truth Collapse" Scoring via Recursive Witness Dynamics: Quantifies the destabilization of narrative integrity under recursive interrogation.
When truth collapses, it does not vanish — it echoes in recursion.
C. Classification Model
From this analysis, we derive a 3-Zone Classification Model based on recursive coherence degradation:
-
Zone I — Unintentional Incoherence (Low Risk)
Language inconsistencies stem from stress, trauma, or low verbal fluency. These are not patterns of deception, but of chaos. -
Zone II — Adaptive Rationalization (Medium Risk)
Partial distortions. In this zone, claimants subconsciously reshape their story to protect self-image, omit responsibility, or preempt skepticism. -
Zone III — Deliberate Narrative Fabrication (High Risk)
Highly structured but recursively incoherent patterns — overjustification, shifting time references, and rehearsed empathy — mark deliberate deception.
This model does not judge. It classifies where language begins to fracture under the weight of intention.