Persona Drift and Hallucination as a Production Discipline: An Eval-Harness Substrate for Agentic Operations

Production agentic systems exhibit two operationally distinct drift patterns that conventional QA does not detect: operator-trust drift (the harness reports work done while no artifact moves) and output-mode drift (the harness produces the same output mode regardless of variation in the prompt). Both fail invisibly because the verification primitive lives inside the harness's self-knowledge.

This paper specifies the runtime substrate that closes the gap: an eval-harness architecture with ground-truth fixtures, scoreable assertions, re-runnable evaluations, and a three-layer monitoring stack (per-task, per-session, per-milestone) at three cadences with three ownership boundaries. The drift taxonomy is three-class — useful drift (uncertainty surfaced), pathological hallucination (certainty about fabrication), and operator-acknowledged residual (known-unknowns the operator has accepted as ambient) — not the conventional two-class binary. Two forensic case studies from Q2 2026 production operations ground the architecture: a multi-paper coast event in which the harness conflated merge-to-main with shipped-to-canonical-URL, and an editorial-substrate feasibility eval that surfaced output-mode drift before the substrate reached production. Together, the two cases demonstrate why the substrate is load-bearing.

This is Paper Four of the Managing Agentics Ops decalogy. The paper specifies the runtime substrate the lookback discipline (Paper Five) measures against and the sprint container (Paper Eight) operates within.