Verifiable Cleanliness for LLMs: Data-SBOM, CAP Protocol, and a Poison Stress Test
Year: 2025
Type: article
Abstract: Small-budget training-data poisoning makes open LLM pipelines a governance problem, not just a technical one. This paper introduces a practical baseline: a machine-readable Data-SBOM (DSBOM), an operational CAP protocol (Cleanliness–Audit–Provenance), and a reproducible pre-deployment Poison Stress Test (PST), plus metrics (TIGI, R-Persist, S-Frag, A-Score) to measure susceptibility and auditability. Together these artifacts form a testable gate and deployment roadmap for enterprises, model hubs, and regulators to raise provable cleanliness and reduce poisoning risk.
Author: Raser, Jake
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7095006362
FWCI:
Cites:
Cited by:
Related to: