Impact Flows Across Layers: How Data Quality Cascades into Agent Decisions into Systemic Stability

Abstract

We formalize the inter-layer impact propagation model underlying the Amplitude measurement architecture. Poor data quality (measured by Meridian) degrades agent behavioral trust (measured by Fidelity), which amplifies systemic risk (measured by Cascade), which in turn triggers governance accountability requirements (measured by Mandate). This directed flow across the Data, Agent, and Ecosystem layers establishes that AI impact measurement cannot be decomposed into independent dimensions - the layers interact through measurable causal pathways.

Background

The dominant approach to AI governance treats different dimensions of AI impact as separable concerns. Data quality is the province of data engineers. Algorithmic fairness belongs to ML researchers. Systemic risk falls to regulators. Human oversight is a compliance function. This separation mirrors organizational structures, but it does not mirror reality. In reality, a degradation in data quality does not remain contained within the data layer; it propagates through agent decision-making into organizational outcomes and ultimately into ecosystem-wide dynamics. Any measurement framework that treats these layers as independent will systematically underestimate the total impact of failures at any single layer.

The history of financial regulation provides instructive parallels. Before the 2008 financial crisis ^[1], risk was measured in silos: credit risk, market risk, and operational risk were assessed through separate methodologies with limited cross-contamination analysis. The crisis demonstrated that risks propagate across these categories through mechanisms that silo-based measurement fails to capture. A decline in housing prices (market risk) triggered mortgage defaults (credit risk), which caused counterparty failures (operational risk), which amplified market volatility (market risk again). The Basel III reforms explicitly acknowledged this inter-category propagation by introducing systemic risk buffers ^[2] and interconnectedness assessments.

AI systems present an analogous propagation challenge. An autonomous trading agent that receives degraded market data does not simply produce slightly less optimal trades. The degraded data affects the agent's internal model, which alters its behavioral patterns, which changes its interaction with other agents in the ecosystem, which can trigger cascading effects ^[3] that amplify far beyond the initial data quality problem. Measuring only the data quality, or only the agent behavior, or only the ecosystem dynamics provides an incomplete picture that can miss the most consequential failure modes.

The Amplitude measurement architecture addresses this challenge by explicitly modeling the directed flow of impact across three layers: Data, Agent, and Ecosystem. Each layer contains measurement frameworks designed to capture the phenomena specific to that layer, but the architecture also defines the causal pathways ^[4] through which scores at one layer influence scores at adjacent layers. This paper formalizes those inter-layer pathways and demonstrates that they produce measurable, predictable effects.

Approach

We model inter-layer impact propagation as a directed acyclic graph ^[5] with three primary layers. The Data layer, anchored by the Meridian framework, measures the quality of information flowing into AI systems across four dimensions: accuracy, completeness, consistency, and timeliness. The Agent layer, anchored by frameworks including Fidelity, Provenance, Threshold, and Drift, measures the behavioral characteristics of autonomous systems including trust, identity, security, and alignment. The Ecosystem layer, anchored by frameworks including Cascade, Harmony, and Torque, measures systemic properties including concentration risk, competitive dynamics, and economic efficiency. A fourth cross-cutting layer, Governance, anchored by Mandate and Parity, measures oversight and fairness properties that span all three primary layers.

The directed flow model specifies that Data layer scores establish boundary conditions for Agent layer scores. Formally, we define the conditional score function S_agent(x | D) where D represents the data quality environment in which the agent operates. When Meridian scores fall below critical thresholds, the maximum achievable Fidelity score for any agent operating on that data is mathematically bounded. This is not a modeling assumption; it is a consequence of information theory ^[6]. An agent cannot exhibit behavioral consistency that exceeds the consistency of its input data, just as a financial model cannot produce predictions more accurate than the data it consumes.

The Agent-to-Ecosystem propagation follows a different mechanism. Individual agent behavioral scores do not directly determine ecosystem scores; rather, the distribution of agent scores across a population determines systemic properties. A market in which 95% of agents have high Fidelity scores exhibits qualitatively different systemic dynamics than a market in which Fidelity scores follow a bimodal distribution. Cascade, which measures concentration and contagion risk, takes as input the joint distribution of agent-level scores rather than individual scores, capturing the emergent systemic properties that arise from agent interactions ^[7].

The Ecosystem-to-Governance propagation closes the loop by connecting systemic risk measurements to accountability requirements. When Cascade identifies elevated systemic risk, the Mandate framework's requirements for human oversight effectiveness increase proportionally. This formalization captures the regulatory intuition that higher-risk environments demand more rigorous oversight, but renders it quantitative rather than qualitative. The propagation is not merely correlational; it is a design-level requirement that ensures governance intensity tracks measured risk.

We validate the directed flow model through simulation experiments in which we systematically degrade Data layer inputs and measure the downstream effects on Agent and Ecosystem scores. The simulations confirm that impact propagates through the specified pathways with measurable attenuation factors that vary by pathway. Data-to-Agent propagation exhibits approximately linear attenuation in the moderate-quality regime but nonlinear amplification below critical quality thresholds ^[8]. Agent-to-Ecosystem propagation exhibits threshold effects consistent with phase transitions in complex systems ^[9], where gradual degradation of agent scores can trigger sudden shifts in systemic risk measures.

Findings

The simulation experiments yield three principal findings. First, the magnitude of inter-layer propagation is substantial and cannot be dismissed as second-order effects. A 20-point reduction in Meridian data quality scores produces, on average, a 12-point reduction in Fidelity behavioral trust scores for agents operating on that data, and a 7-point increase in Cascade systemic risk scores when the degradation affects agents controlling more than 15% of market activity. These are not hypothetical numbers; they emerge from the mathematical structure of the scoring frameworks and the information-theoretic constraints on agent behavior.

Second, the propagation is asymmetric. Improvements in data quality propagate to agent behavior with diminishing returns, while degradations propagate with increasing severity. This asymmetry arises because high-quality data is necessary but not sufficient for high-quality agent behavior, whereas low-quality data is both necessary and sufficient for degraded agent behavior ^[10]. The practical implication is that maintaining data quality above threshold levels is substantially more important than optimizing already-adequate data quality to higher levels. Regulatory frameworks that establish minimum data quality standards are therefore better aligned with the propagation dynamics than frameworks that reward marginal improvements at the top of the quality spectrum.

Third, the propagation exhibits temporal dynamics that create measurement challenges for point-in-time assessments. Data quality degradation does not instantaneously propagate to agent behavior; there is a lag that depends on the agent's model update frequency, memory structure, and learning rate ^[11]. Similarly, agent behavioral degradation does not instantaneously affect ecosystem dynamics; the propagation depends on the agent's market share, the diversity of the agent population, and the coupling strength between agents. These temporal dynamics mean that a snapshot measurement at any single point in time may miss propagation effects that are in transit between layers.

The causal pathway analysis ^[12] also reveals feedback loops that complicate the directed-flow model. While the primary flow is Data-to-Agent-to-Ecosystem, ecosystem dynamics can feed back into data quality through market effects. For example, elevated systemic risk can reduce the willingness of data providers to share information, degrading data completeness scores, which further degrades agent behavior, creating a negative feedback spiral ^[13]. Our model accounts for these feedback loops by treating them as second-order effects that modify the primary directed flow rather than reversing it.

Implications

The formalization of inter-layer impact propagation has direct consequences for how AI impact measurement should be designed, implemented, and interpreted. The most fundamental implication is that AI impact measurement cannot be decomposed into independent dimensions without losing essential information about how impacts actually propagate. A measurement system that produces a data quality score, an agent behavior score, and a systemic risk score as independent numbers provides less information than a system that also captures the causal relationships between those scores. The Amplitude architecture therefore includes inter-layer propagation as a first-class measurement concept, not an afterthought.

For regulators, the propagation model implies that regulatory interventions at the data layer can have outsized effects on agent behavior and systemic risk. Data quality standards are not merely a concern for data engineers; they are a systemic risk management tool ^[2]. Conversely, governance requirements that focus exclusively on agent behavior without addressing the data quality environment in which agents operate will be systematically less effective than anticipated, because they address symptoms rather than root causes.

For organizations deploying AI systems, the propagation model implies that risk assessment must account for the full causal chain from data to agent to ecosystem. An organization that measures its agent's Fidelity score in isolation, without understanding the data quality dependencies of that score, may be surprised when data quality degradation causes a behavioral trust score collapse. The propagation model enables predictive risk assessment: given current data quality trends, what is the expected trajectory of agent behavioral trust, and what are the systemic risk implications?

For the measurement science community, the non-decomposability result establishes a research agenda focused on inter-layer dynamics. Current AI evaluation research overwhelmingly focuses on single-layer phenomena: data quality benchmarks, model capability evaluations, or systemic risk models. The propagation model demonstrates that the most consequential measurement challenges lie at the boundaries between layers, where the mechanisms of impact transmission are least understood and most consequential. Future measurement research should prioritize these inter-layer dynamics over marginal improvements in single-layer measurement precision.