Inceptive RAG: A Rigorous Gradient-Based Attribution Metric for Programmatic Data Monetization

As Retrieval-Augmented Generation (RAG) systems increasingly
mediate access to expert knowledge, a critical “Inceptive Gap” has
emerged: data creators remain uncompensated because current
systems lack a rigorous mechanism to measure the utility of re-
trieved context. Existing attribution methods rely on heuristic prox-
ies like attention weights, which often fail to capture true causal
influence, misattributing credit to superficial token overlaps rather
than semantic logic. In this work, we propose Inceptive RAG, a
framework utilizing the Relative Contribution Index (RCI) to
quantify the causal utility of retrieved documents for Small Lan-
guage Models (SLMs). Unlike prior approaches that validate against
static citations, we validate RCI against the gold standard of Leave-
One-Out (LOO) Perturbation. Experiments on a scaled dataset of
2,000 diachronic queries demonstrate that our gradient-based met-
ric correlates strongly (𝑅 = 0.89) with the actual probability drop
observed when context is removed, significantly outperforming
attention baselines (𝑅 = 0.54). We further demonstrate robustness
at scale, showing that RCI maintains fidelity even with large context
windows (𝑘 = 50), whereas baseline metrics degrade rapidly due
to probability dilution. To address the computational overhead of
inference-time gradient calculation, we introduce an Asynchro-
nous Audit Protocol, decoupling the attribution process from
real-time user latency. By shifting the computational burden to an
offline audit layer, we enable a precise, programmatic monetization
model specifically designed for high-stakes, on-premise domains
where white-box auditability and fair compensation are paramount.