Detect neural network training pathologies before they compound into brittle or unsafe behavior. Veriscope provides early-warning detection with auditable, evidence-based gating.
Veriscope is a research framework for detecting soft collapse, mode collapse, and gradient instabilities in neural network training.
Veriscope identifies training pathologies where standard metrics look healthy but internal representations are degrading.
Monitor effective dimension and variance distribution across eigenspaces to detect representational collapse early.
Track entropy gradients and gradient noise scale to identify instabilities in training dynamics.
Compute sliced Wasserstein-2 and total variation distances to catch distribution drift during training.
Optional H0 persistence computation via Ripser for advanced structural analysis of learned representations.
Every pass/fail decision includes explicit reasons and artifacts suitable for compliance and post-hoc review.
Separate short-horizon change detection from sustained deviation relative to healthy reference regimes.
Monitoring as an auditable decision contract. Declare observable windows, calibrate healthy ranges, and enforce sensitivity budgets with full accountability.
# GateEngine Decision Flow
def evaluate_gate(metrics):
# Gain check
assert prequential_gain >= threshold
# Stability check
assert D_W(past, recent) <= ฮต - ฮต_stat
# Sensitivity budget
assert ฮบ_sens <= ฮต_sens
return Decision(
status="PASS",
reasons=reasons,
artifacts=artifacts
)Open-source under AGPL-3.0. Commercial licensing available for organizations.