Counterparty-Level Bank Exposure to Private Credit and Tail-Risk Bounds: Evidence from the February-April 2026 Gate Cascade

Version 4 (2026-08-19) — supersedes version 3.0 (2026-08-05) and withdraws its identification of a largest bank counterparty. Across the three join rules that share the current extraction, two different banks take first place on identical filings, and JPMorgan Chase's own attributed book spans a factor of 1.23; including a fourth, superseded extraction vintage widens that to three banks and 1.38x, but that is an extraction difference rather than a join rule and the paper now says so. In place of the 21-facility hand-curated matrix the primary object is a 361-facility machine-extracted panel covering 121 BDCs and USD 221.97 billion of committed capacity, of which 61.7 percent of dollars is attributed to one of 29 named banks. The surviving positive result is the gate cross: USD 36.55 billion of committed capacity at ever-gated funds, USD 16.70 billion of it attributable to six banks, with Sumitomo Mitsui carrying USD 7.45 billion. That cross is invariant across the implemented join rules, and the paper now states the true reason - 56.0 percent of it resolves on a named identity token and the remaining 44.0 percent is measured to be invariant rather than guaranteed to be, because three of the eleven attributed cells carry no identity token at all. DebtRank amplification falls from 1.814 under the deposited normalisers to 1.512 on real CET1 and 1.538 on the observed panel, so the working finding that propagation roughly doubles the first-round hit does not survive - and no deposited version carries a DebtRank leg at all, so that computation is new here and what it corrects is our own first configuration of it rather than a published claim; the disclosure-anchored DFAST leg is reread as a descriptive boundedness statement rather than an identified stress test. Every printed quantity resolves to one of 465 machine-generated macros with per-macro provenance, and the bundle's macro audit ships with it. Note on the manuscript PDF (added at deposit). The PDF's title block describes itself as a "working revision v2" that supersedes "the deposited version of June 2026". That was written before this deposit and is superseded by it: the file is no longer a working revision but the deposited version, archived in this record. Its internal "v2" label counts revisions of the manuscript text and does not correspond to this record's version number, which is 4.0. Version 3 (2026-08) — CORRECTION of the published v2 record. (1) The hand-curated bilateral matrix is corrected from USD 27.84B to USD 27.61B; JPMorgan's equal-split share from USD 5.10B / 18.3% to USD 4.79B / 17.3%; HHI from 1214 to 1063 equal-split and from 2011 to 1657 lead-arranger-only; Wells Fargo's lead-arranger book from 33.3% to 28.5%. The v2 figures rested on four facility rows whose lender identities traced to no filing. (2) Those four rows are re-attributed and verified against live SEC EDGAR: TSLX USD 1.675B (Truist administrative agent, J.P. Morgan syndication agent); TCPC Citibank N.A. USD 265M within a USD 765M derived leverage total; BBDC USD 825M (ING Capital sole agent); BCSF USD 855M (SMBC administrative agent, MUFG joint lead arranger). (3) Control-group contamination is now disclosed: Bank of America, Citigroup and Goldman Sachs hold USD 57.0B, or 37%, of the exposure the control group was assumed to lack, so the reported CAR gaps are lower bounds. (4) Table 3 is recomputed on the corrected matrix. The v2 table was not generated from the facility matrix at all but from hand-assigned arranger weights over a ten-fund sample, and it conflated maximum loss (JPMorgan) with maximum CET1 drawdown (Wells Fargo): JPMorgan's drawdown is 1.08pp, not 1.95pp, and the binding Fed DFAST participant is Truist at 1.87pp. (5) The v2 description's "USD 40.7bn in committed credit lines, JPMorgan dominant primary arranger at USD 12.15bn" figures are WITHDRAWN. (6) Scope caveat added: Truist is not an FSB G-SIB, and roughly USD 1.3B across the re-verified rows sits with undisclosed syndicate members, so row-level G-SIB attribution beyond the named agent is not supportable from 10-K disclosure. Empirical companion to the K-A-T monolith. Provides the first publicly available facility-level bilateral matrix linking individual private-credit lenders to individual bank credit-line providers, hand-curated from EDGAR 10-K filings for fourteen U.S. BDCs. The matrix documents USD 27.61B in committed credit lines across 21 facilities (18 bank credit lines), with JPMorgan the largest equal-split holder at USD 4.79B (17.3%) and Wells Fargo the largest lead-arranger book at USD 7.88B (28.5%); HHI 1063 equal-split, 1657 lead-arranger-only, computed on the fully disaggregated 15-lender vector. Scaled to bank-side disclosures from April 2026 G-SIB Q1 earnings, the matrix anchors a disclosure-calibrated DFAST stress test recovering a 2.18 percentage-point CET1 drawdown at Wells Fargo in the catastrophic 40%-default scenario, breaching the Federal Reserve 2025 DFAST baseline by 0.58pp — narrowly, and concentrated at one institution. The February 19 to April 8 gate cascade is treated as a three-wave natural experiment; treated alternative managers accumulated CAR of -10.33% against +2.10% for low-exposure financials (difference -12.42pp, permutation p = 0.059), which the control-group contamination noted above makes a lower bound. Proposes a Contingent Private Credit Repo Facility (CPCRF) with a specified haircut schedule (15% IG / 30% sub-IG), penalty rate (SOFR+100bps floor) and eligibility conditions; the facility's legal-authority analysis remains outstanding. Replication code and the EDGAR provenance record are deposited with this record. Cite the concept DOI 10.5281/zenodo.20095768, which always resolves to the latest version.