The Interlocking Autoregression (EA-LO-INTERLOCKING-AUTOREGRESSION-01) is the companion Lagrange Observatory! white paper to The Keyed Ensemble (#1555). It maps three independently documented distortions of the training ecology — provenance opacity at the informative granularity, distributional shaping of nominally human text under the Mediation Ratchet, and a heritable keyed signature whose model-level residue no corpus operation can remove — as coupled state dynamics under a mismatched observation regime: X_{n+1} = F(X_n; I,II,III), Y_n = M_IV(X_n). Every edge of the map carries an explicit evidentiary status; the nodes are supported by external literature and standing archive analysis, and the edges are declared the new claims. A toy dynamics over a Zipfian type distribution (50,000 types, seven scenarios, a delta-sweep, and a retention counterfactual; code and traces attached with per-file SHA-256) demonstrates two-track divergence — tail mass halves by generation 7 while the head-weighted benchmark holds to generation 15 and the per-round quality proxy rises through a quarter of the trajectory — velocity ordering across components, detection lag as a function of benchmark probe composition, watermark state-dependence as an existence result, and a retention counterfactual in which fine-grained provenance is the control surface withheld from downstream corpus builders. Eight measurement hooks are specified, including a benchmark composition audit requiring nothing but the benchmarks themselves. Anchored across the archive's model-collapse wing, with the generative-monoculture term cited to Wu, Black & Chandrasekaran per ERRATUM to AXN:0341 (#1554).