Semantic systems / AI research

SemTrace

A leakage-aware test of whether graph-derived protocol context adds signal to DeFi incident severity classification.

Corrected dataset

1,608

normalized incidents from 2011–2025

Chronological holdout AUC

0.823 / 0.817

baseline / graph-enriched

Temporal CV AUC

0.775 / 0.787

baseline / graph-enriched mean; uncertainty remains material

System view

Semantic relationship schema
Semantic system graphA restrained system relationship graph with labeled entities and a metadata field.

Problem

The research asks whether protocol context known before an incident improves the ranking of severe financial-loss events. The central constraint is temporal validity: graph facts cannot be used unless their availability before the incident can be demonstrated.

Engineering approach

  1. Use the same deterministic LightGBM configuration and baseline inputs for both models.
  2. Admit a graph feature only when its availability timestamp is no later than the incident.
  3. Evaluate on a chronological holdout and five expanding-window validation folds.
  4. Keep conference reproduction and corrected portfolio evaluation as separate profiles.

What the evidence supports

Graph enrichment did not improve the corrected chronological holdout. Temporal cross-validation was slightly higher on average, but the result was not consistent enough to support a strong predictive claim.