Emergent Epistemic Drift in Multi-Agent LLM Pipelines: Transmission Bias Dynamics in a Collective Adaptation Testbed

Multi-agent large language model (LLM) pipelines are increasingly deployed as collaborative reasoning platforms, yet their epistemic dynamics — how reasoning patterns propagate and homogenise across interacting agents — remain poorly characterised. We introduce the Collective Adaptation Testbed (CAT), a six-agent LLM pipeline engineered as a self-observing complex adaptive system for measuring transmission bias, the mechanisms by which behavioural patterns spread within an agent collective. Drawing on the dual-inheritance distinction between frequency-dependent (FD; conformity-based) and payoff-biased (PB; performance-weighted) transmission, we operationalise these categories as classifier-labelled decision events. Across a four-day pilot (N = 46 runs; 282 classified decisions), we document a sharp, governance-driven epistemic reset: FD bias rose to 43.9% of classified decisions on Day 3 and collapsed to 0% on Day 4 (χ²(1, N = 137) = 14.05, p < .001; Cohen's h = 1.45). Critically, this reset is produced by a deterministic dissent-injection protocol rather than spontaneous self-organisation, and is therefore a proof-of-concept demonstration of an engineered FD-suppression mechanism, not organic emergence. Concurrently, the adaptive selector developed stable pipeline-quality preferences (spread 0.745) and logged 57 emergence events, and an expert bottleneck in the quality-gating agent — flagged by the system's NudgeEngine — was resolved through a human-in-the-loop rubric intervention that raised approval-threshold quality from 0.27 to 0.73 (a 170% gain). Together these results establish transmission bias as a measurable, non-stationary, and structurally governable property of multi-agent LLM systems, and motivate epistemic-drift monitoring as a design principle for self-correcting AI architectures that resist knowledge collapse. Data, code, and analysis scripts: https://github.com/volkanask/CAT-collective-adaptation-testbed and https://doi.org/10.17605/OSF.IO/ZWYPG