Network Topology and Systemic Contagion Among Alternative Asset Managers: Evidence from GP-LP-Bank Interconnections

Version 3.0 (2026-08-19). This revision replaces the simulation with an observed network as the paper's primary object: 20,623 filed BDC-to-issuer edges across 174 BDCs and 11,171 issuers carrying USD 446.14 billion of fair value, a 361-facility bank layer with USD 221.97 billion of committed capacity of which 153 facilities worth USD 130.58 billion are attributed to a named bank, and 94 adviser complexes. The maximum-entropy IPF reconstruction that carried the deposited version is demoted to a compared baseline and backtested against that network: it recovers real counterparty links 2.727 times better than chance at an edge F1 of 0.0550, with the lift confined to 2.28-3.00x at every size tested. Centrality does not predict gating order - size does - which reverses a claim of the deposited paper, so the D-SINBFI designation proposal is reformulated with the eigenvector-centrality term dropped in favour of a common-exposure term and the illustrative weight vector withdrawn, no artifact supplying a weight for the new term. Table 4's Panels A and C are withdrawn outright: the Spearman association they carried was a free parameter of an uncalibrated synthetic AUM tail, since only 21 of the 179 GP nodes carry a disclosed AUM and moving the fabricated tail's starting value moves rho from 0.313 to 0.036; restricted to the disclosed-AUM nodes the statistic is -0.017 at p = 0.949. A pre-registered hypothesis failed and is reported as failed: H6 predicted the reconstruction would beat chance but stay below twice chance on an independent N-PORT network, and the observed median lift was 10.9x. Two funds leave the gating cohort under the evidence class observed, no gate evidenced, because all twelve never-gated ledger funds across 93 panel rows disclose no tendered-share count, so a derived not-prorated label is not a documented non-gate. What the text below no longer describes accurately. The tripartite GP-LP-Bank network described below is a maximum-entropy reconstruction; this version demotes it to a compared baseline and measures it against an observed filed network, so it is no longer the paper's primary object. The D-SINBFI framework described below is reformulated here: the eigenvector-centrality term is dropped in favour of a common-exposure term and the illustrative weight vector is withdrawn, with no artifact supplying a weight for the new term. Table 4's Panels A and C are withdrawn outright. The retirements already recorded in the Version 2 paragraph below - the betweenness-predicts-gating claim and the 6.2x amplification headline - remain retired. The earlier text is left standing unedited as the record of what those versions said. The manuscript's own title in this version is: A Calibrated Tripartite Network Simulation of Alternative-Asset-Manager Contagion, with an Observed BDC-Bank Facility Anchor - Corrected Cohort and Private-T Integration with the K-A-T Programme. The record title above is retained unchanged so that the concept DOI keeps a stable citation string across versions. Version 2.1 (2026-06) — presentation correction. The Section-8 heading and the conclusion are brought into line with the abstract's honest framing: the gating-order predictor is a SIZE (AUM) effect — a regression on AUM-share alone reproduces the identical fit (Spearman rho=0.301, logistic pseudo-R^2=0.27) — so network centrality carries no predictive content independent of sponsor size. No numbers or results change; this removes an internal abstract-vs-conclusion inconsistency present in v2. Version 2 (2026-06) — supersedes v1 (which remains citable at its version DOI). Corrections from the programme's data-integrity revision: (1) the Q1-2026 gated cohort is corrected to the SEC-verified seven-fund cap-exceeding set; (2) the v1 claim that betweenness centrality predicts gating order (rho=0.195) and the 6.2x amplification headline are RETIRED — centrality is disclosed as a deterministic transform of AUM with no independent predictive content; (3) the bilateral bank-facility matrix is corrected to the equal-syndicate-split aggregate of $27.84B (JPMorgan $5.10B, 18.3%; HHI 1214); the stale orphan '$40.67B / 29.9%' total is removed; (4) the DebtRank targeted-vs-random amplification (2.37x) and all robustness statistics are reproduced by a deterministic, seeded replication script (build_network_and_regression.py). No result is asserted without producing code. Constructs a tripartite GP-LP-Bank network (179 GPs, 113 LPs, 35 banks; 327 nodes) using SEC Form ADV filings, 13F holdings, and BIS bank-NBFI lending statistics. Bilateral exposures reconstructed via maximum-entropy iterative proportional fitting from aggregate marginals, with power-law sparsification calibrated to scale-free patterns. Applies DebtRank methodology (Battiston et al. 2012) to derive per-node systemic-importance scores. The March 2026 private-credit crisis serves as a natural experiment: betweenness centrality predicts gating order (Spearman rho = 0.195, p = 0.009); Blackstone ranks first by eigenvector centrality. Monte Carlo stress tests (10,000 scenarios) show simultaneous distress at the three most central nodes generates 11.6% of network value vs 1.9% for random nodes, a 6.2x amplification. Proposes a Designated Systemically Important Non-Bank Financial Institution (D-SINBFI) framework calibrated to network topology metrics. Limitations section discusses bilateral-reconstruction caveats, network coverage limits, DebtRank monotonicity assumption, cross-channel transmission magnitude uncertainty, and single-episode anchor.