Benchmarks

Six sealed benchmarks.
Publicly reproducible.

Every benchmark ships with a sealed corpus, honest baselines, and step-by-step reproduction instructions. Engine implementations are proprietary — the scoring harnesses and baseline code are not.

MUNINN

Memory Validation

The first benchmark for memory validation pipelines — contradiction detection, importance ranking, and honest refusal.

F1 0.847 · Recall 0.921

SIGNAL

Pharmacovigilance NLP

Adverse drug event extraction from clinical narratives and pharmacovigilance reports.

F1 0.639 · 24.3mo median detection

CITADEL

Corporate Hierarchy

Entity resolution and hierarchical reasoning over corporate ownership structures.

F1 0.616

SENTINEL

SOC Triage

Security operations center alert triage and prioritization.

94% held-out accuracy

ORACLE

Factual Verification

Cross-domain factual verification with honest refusal on contested claims.

51% vs 31% / 25% baselines

LENS

Semantic Code Search

Deterministic semantic search over codebases without embedding models.

P@5 0.250 deterministic

Six sealed benchmarks.Publicly reproducible.

Six sealed benchmarks.
Publicly reproducible.