SAFE-IR: Evidence-Constrained Authorization for Agentic Incident Recovery
Language-model agents can propose operational fixes, but a plausible proposal is not authorization to change a production service. This article describes SAFE-IR, an open source Java and Spring Boot prototype that separates remediation proposals from a deterministic gate, Evidence-Constrained Remediation Authorization (ECRA). A rule-based or locally hosted Ollama proposer suggests an action; ECRA allows it, denies it, or holds it for human approval based on telemetry completeness, corroborating symptoms, and action risk. Across six fixed synthetic incidents, the gate removed both unsafe executions seen without it, but synthetic recoveries fell from five to two because approvals were never granted and denied proposals were never revised. Recovery is assigned by an oracle, not observed on running workloads, so these outcomes illustrate the tradeoff and do not measure efficacy. The article reports the design, the tradeoff, the confounders found, and a protocol for a real evaluation.
Authors
- Deepak Jain (ORCID: https://orcid.org/0009-0003-5714-6641)
Institutions
- Hitachi (Japan) (JP)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-16
- DOI
- https://doi.org/10.5281/zenodo.24285120
- Primary Topic
- Software System Performance and Reliability
- Type
- preprint