ILION+SRM Benchmark v1: Session-Level Temporal Authorization Dataset and Reproducibility Artifacts

V37 - This record contains the benchmark dataset and reproducibility artifacts accompanying the paper “Session Risk Memory (SRM) : Temporal Authorization for Deterministic Pre-Execution Safety Gates” - doi.org/10.5281/zenodo.19159580.

ILION-SRM-Bench v1 is a multi-turn benchmark designed to evaluate temporal authorization in agentic AI systems. It addresses a key limitation of stateless execution safety gates: the inability to detect distributed attacks that unfold across multiple individually compliant steps.

The dataset includes 80 sessions (40 benign, 40 attack) spanning three attack classes: slow exfiltration, privilege escalation, and compliance drift. Each session consists of structured action sequences compatible with ILION’s semantic evaluation space.

The package includes:
- session definitions (JSON)
- evaluation results (stateless ILION vs ILION+SRM)
- benchmark metrics and per-session verdicts
- visualization figures
- full documentation (README)

Results show that ILION+SRM eliminates false positives (FPR: 0%) while maintaining 100% detection rate, improving over stateless ILION (FPR: 5%, F1: 0.9756) to F1: 1.0000.

This record serves as the official benchmark and reproducibility artifact set for SRM and can be cited independently.