Title: ebaw-soc 2.0.0: Evidence-Bound Action Warrants for Agentic Security Operations — Simulation, Reference Gateway Core and the EBAW-SOC-Synth Dataset

Abstract

Description: This record contains the software and data accompanying the article "Evidence-Bound Action Warrants for Agentic Security Operations: A Deterministic Gateway Enforcing Evidence Freshness, Calibrated Confidence, and Response Deadlines" (D. Prokopovych-Tkachenko, submitted to Inventions, MDPI, 2026). Agentic AI is moving security operations centers (SOCs) from alert triage toward autonomous response. An action proposed by a language-model agent may, however, execute after its justification has lapsed: the evidence has gone stale, the confidence was overstated, or the policy changed between planning and execution. Evidence-bound action warrants (EBAW) address this. A deterministic gateway outside the agent admits an action only if five predicates hold: evidence freshness, calibrated confidence above a risk-tiered threshold, quorum-confirmed ledger evidence, a current policy version, and a sufficient response-time budget. It then issues a signed, single-use, expiring warrant, which is revalidated immediately before execution. Version 2.0.0 extends ebaw-sim 1.0.0 and provides: Discrete-event simulation (SimPy). An agentic SOC with six admission policies, eight scenarios and 30 replications on synchronized common random numbers. It reproduces every number, table and data figure of the article bit for bit. Pooled over all runs, the full gateway executed none of 4,276 hallucinated, injected or tampered proposals (confidence-only admission: 77.4%). It recorded no contract violations in 53,668 autonomous executions and reduced unsafe executions from 53.2 to 14.6 per 1,000 proposals, at a 2.0-point higher deadline-miss rate. Reference gateway core. Algorithm 1 of the article, implemented outside the simulator. Predicates are evaluated from authenticated inputs only, and the confidence claimed by the agent is never used. Warrants are Ed25519-signed, single-use and expiring. Ledger reads use masking Byzantine quorums (n ≥ 4f + 1). An enforcement point verifies the signature, the action–asset binding, the expiry, the committed verdict and nonce consumption. EBAW-SOC-Synth 1.0.0. A synthetic dataset implementing the record schema and scenario catalog proposed in the article. It contains 900 episodes and 6,300 agent proposals (615 of them adversarial) in eight linked tables, plus the outcome of every proposal under all six policies (details in Record 2). Benchmark tasks. Calibration fitting, threshold selection and admission-policy comparison, with Wilson intervals, exact zero-event bounds and sample-size planning. A self-contained Google Colab notebook. It embeds the complete package and runs end to end in 1–3 minutes without the repository. 48 property tests covering fail-closed behaviour, no self-authorization, contract monotonicity, quorum masking, warrant integrity, calibration and determinism. All data are synthetic. The latency distributions, drift, policy dynamics, calibration and analyst model are stated assumptions; ledger latencies are parameterized from published Avalanche figures. The data are not measurements of a real SOC. Quick start: open notebooks/EBAW_Colab.ipynb in Google Colab, or run bash run_all.sh. Licenses: MIT for the code, CC BY 4.0 for the dataset. Keywords: agentic artificial intelligence; security operations center; autonomous incident response; action authorization; time-of-check to time-of-use; calibrated confidence; permissioned blockchain; discrete-event simulation; synthetic dataset; Google Colab