SARC-DQ: Runtime Data-Quality Gating for Agentic AI: Silent Evidence Defects, the Incompetence Shield, and Downstream-Only Remediation
Source
arxiv.org
Author
Gaston Besanson
Date
Why it matters
Stale or superseded records are well-formed in the payload and betrayed only by freshness and lineage metadata the AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → never sees. Agents acted on them about 60% of the time with doubt markers at chance — and a 15x more expensive model did no better.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
articleOne Gate Is Not Enough: Composing Stateful Pre-Action Controls for Agentic AIGaston Besanson
articleAssurance Envelopes for Autonomous Coding Agents: Minimum-Cost Evidence for Software ChangeAnjan Goswami
articleFrom Traceability to Justifiability: Accountability Structures in Agentic Software EngineeringRashid Azarang