The Geometry of Market Fragility - Part I

How Close Is the US Treasury Market to a Self-Amplifying Break?

---

Disclaimer. All exhibits and calculations in this report are computed directly on the Pangura Axe platform, drawing on publicly available market data. Pangura provides analytical infrastructure for decision-making by financial institutions and does not provide investment advice.

---

Abstract

Twice in the past six years, the lender of last resort has had to buy the safe asset to keep it safe. In March 2020 the US Treasury market — the asset the rest of the global financial system prices itself against — seized, and the Federal Reserve bought roughly $1tn of Treasuries in three weeks to restart it. In October 2022 the same mechanics played out in the UK gilt market, and the Bank of England intervened within days. Each episode had the same shape: leveraged holders forced to sell, the selling moving prices against them, the moves forcing more selling. The post-mortems describe this shape in words — “dash for cash,” “doom-loop” — but leave two questions quantitatively unanswered: how close does today’s Treasury complex sit to that self-amplifying threshold, and — the question history shows matters more — does the public sector still have the capacity to catch it?

We approach both questions by mapping the Treasury complex — its holders, funders, and assets — as a directed, weighted network built from public data, and reading it through three structural dials: amplification (ρ), the per-round gain of the forced-selling loop (May’s 1972 spectral criterion); loop concentration, where the feedback lives (Levins’ 1974 loop analysis); and the circulation fraction, whether distress drains or recirculates (Schnakenberg’s 1976 flow decomposition). We then rebuild the pre-crisis networks of seven US doom-loops — the Panic of 1907, the 1929 margin collapse, the 1987 portfolio-insurance crash, the near-failure of LTCM in 1998, 2008, the March 2020 Treasury seizure, and the 2023 Silicon Valley Bank failure — from contemporaneous public data, and let the record judge the dials. (Two European crises — the UK’s LDI crisis of 2022 and the euro-area sovereign–bank crisis of 2010–12 — are the subject of the companion report, Part 2.)

The record half-confirms the dials and half-refutes them, and both halves are used. On the eve of the modern crises amplification reads near-critical (1987 0.91, COVID 0.85, LTCM at the critical value): the fragility was visible in advance. But on its own it mis-ranks the outcomes: 1929 had the loosest wiring of the seven (ρ ≈ 0.79) yet the worst end (real GDP −26% peak to trough), because the lender of last resort never came; 2008 was above the critical value yet contained (GDP −4.3%) by the most aggressive rescue in history. The record’s cleanest controlled pair makes the point unmistakable: 1907 and 1929 ran on comparable single-hub structures, but 1907’s panic met a private backstop placed at the keystone and 1929’s met none — −11 to −12% against −26%. Structure sets the fragility; the backstop decides the outcome. The other two dials each earn a place on the same record: loop concentration tells the rescuer whether a targeted purchase can cut the spiral or only a system-wide rescue holds, and circulation, conditional on amplification, anticipates how broad the rescue must be. Replaying the seven histories through the same cascade arithmetic — rescue on, rescue off — reproduces the pattern: the concentrated single-hub crises were single-point rescuable, with 43–68% of the damage avoidable at one node, while 2008’s dispersed feedback gave the best single point only 18%. And the backstop itself becomes a measured axis — three gates: deployed × sized-and-type-matched × affordable — the third resting on the United States’ exorbitant privilege as the world’s safe-asset supplier. The contribution is synthesis, financial application, and historical validation, not new theory; the mathematics is confined to an appendix.

Today, for the first time in the record, both axes are stressed at once. The structural dials read above crisis-eve levels — amplification ρ ≈ 0.98 (0.90 on the observable market core alone; 0.977 with the fiscal-debasement layer included), the highest of any contained crisis in the record and within 0.02 of the runaway threshold — and the amplifier is pre-loaded: dealer market-making depth is 65–85% below its pre-2008 level. The wiring pattern itself — feedback dispersed across many loops yet routed entirely through one hub, and that hub the asset itself — has no precedent in the US record: in every prior case the asset under stress and the asset behind the rescue were different. Meanwhile every containment gate is measurably weaker than in any contained case: net interest now exceeds the defence budget and absorbs ≈18% of federal revenue (CBO 2026); inflation is sticky above target (April 2026 core PCE 3.3% y/y; BEA 2026); and gold has overtaken US Treasuries as the largest single central-bank reserve asset for the first time since 1996 (ECB 2026). The forward case accordingly forks into two failure modes that structure alone cannot distinguish: a backstop that arrives too late or too small (a 1929-type break), or one that is ample but monetised (a currency-debasement loop). A cascade run on the live network makes the fork concrete: the backstop cannot destroy the distress — arriving early it stops the compounding at roughly eleven times leverage per unit absorbed, arriving late it saves nothing, and its financing decides where the absorbed flow lands: the currency and gold if monetised, the fiscal trajectory if not. That is the report’s central, scenario-conditioned, falsifiable claim.

1. The problem — the anchor asset needs a measurable margin of safety

The US Treasury market is the reference point for global finance: the collateral base of the repo system, the pricing benchmark for every dollar asset, and the instrument in which the world’s official reserves are held. Its safety is not a static property of the paper; it is a property of a network — a web of holders, dealers, funds, and funding markets that, in normal times, absorbs selling without strain.

That network has changed shape faster than its safety margins have been re-examined. The Federal Reserve’s share of the market has fallen from a peak of ≈26% in 2021 to ≈14% under quantitative tightening; price-insensitive foreign official holders have retreated in relative terms; and the marginal holder is now a leveraged one — hedge funds running the cash-futures basis trade at leverage ratios around 18:1, holding ≈$1.85tn of Treasuries through Cayman entities, financed by ≈$1.8tn of net repo borrowing. Meanwhile the dealers who intermediate the market have not grown with it: the stock of debt has roughly tripled since 2007 while dealer balance-sheet capacity has not, and market depth today stands 65–85% below its pre-2008 level. The supply of debt, finally, is accelerating into this thinning structure: the CBO’s February 2026 baseline has the deficit at $1.9tn (5.8% of GDP) this fiscal year and debt held by the public rising from ≈100% to ≈120% of GDP within a decade.

The warning shots have already been fired. March 2020 demonstrated that the world’s deepest market can stop functioning in days. Abroad, the UK’s LDI crisis of 2022 (Part 2) demonstrated the same loop in a sovereign market one-tenth the size — and was stopped only by central-bank purchases of the very asset under stress. And through 2025, on several occasions, the dollar weakened while Treasury yields rose and equities fell — a break in the safe-haven pattern that the macro-finance literature reads as the market beginning to re-price the safety of the anchor itself.

What should a risk officer do with this? Forecasting the next price move is the wrong defence: under stress, price is set by the interacting forced decisions of the very holders under examination, and the marginal move is essentially unpredictable. What can be measured — from public data, in advance, and reproducibly — is the structure: how tightly the network’s selling is wired to produce more selling, where that wiring concentrates, whether distress can drain out, and whether the public sector retains the capacity to absorb it. A measured structure is auditable: every edge has a source, every reading can be recomputed, and a decision-maker can locate the load-bearing links and pre-position the response before the loop fires.

This report builds that measurement, in the order the evidence built it. §2 reviews what the published research already documents — channel by channel — and the system question it leaves open. §3 sets up the network representation, the three structural dials that read it — amplification, loop concentration, circulation — each with its theoretical lineage, and the framework’s fourth instrument: a drift gauge that shares no input with the network — and closes with the four questions by which every network map in the report is read. §4 rebuilds seven US historical doom-loops from contemporaneous public data, walks the full anatomy of the cleanest of them, reads the dials, and tabulates the readings against the realised outcomes. §5 lets the record judge the dials — what each gets right, where they fail, and the missing axis the failure exposes; reads the seven networks’ shared anatomy; and closes with the systematic framework this report proposes. §6 develops the axis the record shows to be decisive: the containment capacity of the backstop — and closes by replaying the seven histories through the cascade arithmetic, rescue on and rescue off. §7 applies the full framework to today’s Treasury complex — anatomy first, then the dials — gives the two-tailed forward scenario, and lists what to watch; §8 concludes.

2. What the institutions already see — and the framework they lack

The fragility of the US Treasury market is not a fringe concern; it is the subject of a large, convergent body of institutional and academic research. A reader steeped in that literature will recognise every ingredient of the loops in §4–§7. Our contribution is not to discover the ingredients — the institutions have — but to put the documented channels on one calibrated network and read the coupled system: how close it sits to self-amplification, which loops carry the break, and how broad a rescue would have to be.

2.1 Five channels of one loop

• Dealer intermediation capacity. Duffie et al. (2023) and Duffie (2025) show that Treasury-market liquidity deteriorates non-linearly once dealer balance-sheet utilisation is high — an “occasionally binding constraint” that yield volatility alone does not predict; dealer balance sheets have shrunk markedly relative to the stock of debt since 2007. This is the supply side of the reflexive loop and the source of the pre-loaded amplifier.

• Leverage in the basis trade. The Federal Reserve (Federal Reserve Board 2025; Cook 2025), the ECB (2024), the OFR, and the CFTC’s Market Risk Advisory Committee (CFTC 2024) document a reflexive epicentre: hedge-fund Treasury leverage at multi-decade highs (the largest funds ≈18:1), with Cayman-domiciled funds holding ≈$1.85tn of Treasuries — under-counted in official TIC data by ≈$1.4tn — and net repo borrowing near $1.8tn. This is the keystone holder in our forward network.

• Size outgrowing depth. The Group of Thirty (2021) traces the root cause to a market growing far faster than bank-affiliated dealers’ market-making capacity — in part because post-2008 leverage rules made low-risk intermediation costly — a structural depth deficit (and the motivation for central clearing, the SEC’s Dec-2023 rule, and SLR reform).

• Sovereign supply + the investor base. The IMF’s Global Financial Stability Report (IMF 2025) sets the macro backdrop: debt shifting to governments, rising debt/GDP, and sovereign markets increasingly reliant on price-sensitive investors during quantitative tightening — the Fed’s own holdings down from a ≈26% peak (2021) to ≈14%, with the marginal buyer now leveraged and offshore.

• The central bank as backstop of last resort. A growing literature (e.g. Stein et al. 2026) studies the Fed’s market-functioning purchases — and notes that they transfer non-bank intermediation risk to the public balance sheet. This is the containment side of §6.

Each strand is rigorous, and each measures one channel. What none does — because it is not what a single-channel study is for — is put the channels on one network and ask the system question: given how these holders, funders, and assets are wired together, how close is the coupled network to a self-amplifying break? The practitioners answer this only in words — “dash for cash,” “doom-loop,” “vulnerable to stress.” The phenomenon those words describe is, by its nature, a property of the coupling, not of any node.

2.2 The tools for the system question

A property of the coupling is what network-stability methods quantify. The line of work begins with the stability of large interconnected systems (May 1972) and was carried into finance after 2008 — contagion in banking ecosystems (Haldane & May 2011), the network origins of systemic risk (Acemoglu, Ozdaglar & Tahbaz-Salehi 2015), reverberating distress on interbank data (Battiston et al. 2012). Two further lines analyse the same mathematical object — a weighted network and its feedback structure — from different angles: feedback-loop analysis, developed in ecology (Levins 1974; Puccia & Levins 1985), and the decomposition of flows on networks, developed in statistical physics (Schnakenberg 1976) and lately applied to financial data (Wand, Kamps & Iyetomi 2024). What this report adds is the wiring and the test: the channels this research documents one at a time are placed on a single network whose edge weights are calibrated from the published estimates themselves (Appendix B); the network is read through several of these lenses at once; and each lens is checked against the historical record before it is trusted forward.

The relation to the existing work is direct: the channel studies supply the calibration anchors; the macro-finance literature — Gourinchas and Rey on the dollar’s exorbitant privilege (Rey 2015; Gourinchas, Rey & Govillot 2017), Dalio on the long debt cycle (Dalio 2025) — frames the containment capacity that §6 assesses; what this report computes is the layer that connects them: which loops carry the break, how close the coupled system sits to runaway, and how broad a rescue would need to be. Every edge weight is sourced and every reading is reproducible from public data (Appendix B), so a reader who disputes an assumption can change it and watch the reading move.

3. The system as a network: three structural dials, one independent gauge, and where they come from

We represent the Treasury complex as a directed, weighted network: the nodes are holders, funders, and assets; an edge from i to j carries the marginal sensitivity of j’s forced selling to distress at i, calibrated from public data (per-case anchors in Appendix B). Three structural dials read this network, each answering a different question — how strong is the feedback, where does it live, where does the distress flow. They act on the same mathematical object yet descend from three independently developed lines of theory; this section gives each dial’s mechanism and provenance in turn, presents the framework’s fourth instrument — a drift gauge that deliberately shares no input with the network (§3.4) — and closes with the four questions by which the network maps themselves are read. Plain-language readings stay in the body; formal definitions are confined to Appendix A, and a schematic of the four measurements is given there as Figure A1.

3.1 Amplification ρ — how strong is the feedback

The first measurement is the one the stability literature supplies ready-made. May (1972) showed that whether a large interconnected system damps or amplifies disturbances is governed by a single spectral property of its interaction matrix, and the financial-network literature carried that criterion into banking ecosystems (Haldane & May 2011), systemic-risk propagation (Acemoglu, Ozdaglar & Tahbaz-Salehi 2015), and reverberating-distress measures on real interbank data (Battiston et al. 2012). Applied to a forced-selling network it has a direct reading. Trace one round of the loop: a price fall forces some holders to sell; the selling moves the price; the move forces more selling. The amplification ratio ρ summarises the wiring in one number: for every unit of forced selling, the network sends back ρ units in the next round. While ρ is below 1, each round is smaller than the last and the cascade dies out on its own; total selling is the initial shock multiplied by roughly 1/(1−ρ) — a multiplier of 2 at ρ = 0.5, 5 at ρ = 0.8, 10 at ρ = 0.9. At ρ = 1 self-correction fails: each round reproduces itself, and any shock grows until the structure breaks or someone outside the loop absorbs it. That threshold is the only hard line in this report. A later refinement of the same criterion adds the direction of failure: the sign structure of the interactions decides stability — opposite-sign (predator–prey-type) couplings stabilise, while same-sign mutual reinforcement is the most destabilising configuration (Allesina & Tang 2012); the forced-selling loop is exactly that type — selling deepens distress, and distress deepens selling. (Formally, ρ is the spectral radius of the weighted network — the per-round gain of its dominant feedback mode; Appendix A.1.)

3.2 Loop concentration and the keystone — where the feedback lives

Amplification sums the entire feedback structure into one number; it cannot say where the feedback lives — yet that is what decides a rescue’s form. The decomposition that recovers this information exists in the literature: loop analysis, developed in ecology by Levins (1974), reads the same feedback that sets ρ loop by loop — formally, the characteristic polynomial that determines ρ factors over the network’s closed loops (Appendix A.2) — so it refines the dial rather than replacing it. The reading is operational: if feedback runs through one or two dominant loops, a backstop that buys the single load-bearing asset — the keystone — cuts the spiral at a stroke; if feedback is spread across many interlocking weak loops, no single purchase helps, and only a system-wide backstop works. We summarise the loop-gain distribution as a 0–1 concentration score (1 = all feedback in one loop) and name the keystone; the score statistic is our only addition to Levins’ machinery.

3.3 The circulation fraction — where the distress flows

The first two dials read the static wiring of the pre-crisis structure; the third asks the dynamic question: once the loop fires, where does the distress go? In some structures it drains — to long-horizon buyers, to the real economy, to a standing facility — and pressure releases. In others it circulates: the selling that starts at one desk returns to the same desk amplified, as when cash selling widens the futures basis, which forces more cash selling. Flows on networks admit exactly this split: a decomposition into a draining (gradient) part and a circulating part, machinery developed for non-equilibrium systems (Schnakenberg 1976) and lately applied to financial networks (Wand, Kamps & Iyetomi 2024). We simulate a stylised cascade on each calibrated network and measure the circulation fraction: the share of total distress flow trapped in closed loops rather than draining to an absorber (0 = fully draining, 1 = fully trapped; Appendix A.3). The decomposition mathematics is standard; the summary statistic is our construction, and it carries the contested-tier label accordingly.

3.4 The drift gauge — a cross-check that reads no structure

The three dials above share one input: the calibrated network. An error in the wiring would move them together. The framework therefore carries a fourth instrument that shares nothing with the network: a gauge of how far the market’s day-to-day statistical behaviour has drifted from its calmest historical regime. The mathematics again predates this report by decades: Rao (1945) showed that a family of probability distributions forms a geometric space with a natural metric — the Fisher–Rao metric — under which the distance between two statistical regimes is the information it takes to tell them apart; information geometry entered interest-rate modelling with Brody & Hughston (2001). We summarise the daily changes of the 10-year Treasury yield by a rolling statistical window and measure the Fisher–Rao distance between the current window and a calm baseline selected by a fixed objective rule, never by hand (Appendix A.5). The gauge adds no structural information, and that is the point: if a wiring error distorted the three dials together, a measurement that never reads the network would not move with them. Contested-tier, descriptive.

The dials are summaries; each is read off a map of the network, and the maps repay a fixed order of interrogation. Every network figure in this report answers the same four questions. Where is the engine? — the strongest two-way pair of edges, whose round-trip gain seeds the amplification. Who carries the load? — the single balance sheet bearing the largest share of the network’s total exposure. Which nodes form the self-reinforcing core? — the block inside which distress can return to its origin amplified, as against the nodes that sit on one-way roads out of it. Where would the pressure drain? — whether any node’s intake neutralises distress rather than passing it on, and whether the wiring has internal compartments that could act as fire doors. These four readings — engine, load, core, exits — are descriptive statistics of the wiring, defined formally in Appendix A.6; the dials compress them into actionable form: amplification is the gain of the engine, concentration says whether there is one engine or many, circulation says whether the exits work. §4.6 walks the reading once, slowly, on the cleanest of the historical networks; after that, every map in the report can be read at a glance.

A dial is worth reading only if it can be checked. Before pointing them at today’s market, we point them backwards: rebuild the networks of seven US historical doom-loops as they were wired on the eve of each break, read all three structural dials on the frozen structure, and compare the readings with what actually happened. (The drift gauge reads time series rather than wiring; it meets the record in §5.4.)

4. Seven US doom-loops, reconstructed from public data

Seven episodes in the United States’ own history give the cleanest public record of the self-amplifying loop in its money and government-securities markets: the Panic of 1907, the 1929 margin collapse, the 1987 portfolio-insurance crash, the near-failure of Long-Term Capital Management in 1998, the 2008 financial crisis, the March 2020 Treasury seizure, and the 2023 failure of Silicon Valley Bank. They span more than a century of changing instruments — call loans, margin accounts, stock-index futures, leveraged arbitrage, securitised credit, and the cash-futures basis — but the same self-amplifying structure runs through all of them: leveraged or forced sellers, a margin or collateral clock, and prices that move against the sellers as they sell; in two of them (1907 and SVB 2023) the trigger is a depositor run on a solvency revaluation rather than a margin call, but the loop is the same. For each we rebuild the pre-crisis network (Figure 1) from contemporaneous public data — balance-sheet sizes, leverage, funding chains, and margin rules as they stood on the eve (Appendix B) — and record two things separately: what the network reads, and what the rescue actually was. The cases are taken in chronological order; the narratives lead with amplification, the full three-dial readings are tabulated in §4.8, and the record of each rescue is not commentary but the out-of-sample test of §5. (The two European crises that complete the record — the UK’s LDI crisis of 2022 and the euro-area sovereign–bank crisis of 2010–12 — are treated in the companion report, Part 2, where the question is what catches the loop when no exorbitant-privilege backstop stands behind it.)

4.1 1907 — a private backstop, and the absence of public capacity

The 1907 panic ran on the trust companies — lightly regulated, higher-yielding intermediaries that held lower reserves than the national banks and were not members of the clearinghouse that pooled risk among them; in current terms, the shadow banks of the period (Tallman & Moen 1990; Frydman, Hilt & Zhou 2015). It began with a failed corner: in October 1907 an attempt to corner United Copper collapsed, impairing the brokers and banks tied to its organisers, and when the president of the Knickerbocker Trust — New York’s third-largest — was reported among them, depositors withdrew. The withdrawals were self-reinforcing: a trust meeting redemptions liquidated assets and called in loans; the liquidation lowered asset prices and raised the call-money rate (from 9.5% to 70% on the day Knickerbocker suspended, and to 100% two days later); brokers unable to fund positions sold into the decline; the decline and the frozen call market then drove withdrawals at the next trust. Because the trusts were opaque and similar, a depositor could not distinguish a solvent trust from an insolvent one, so withdrawal was individually rational and the run extended across the sector.

Pre-crisis reading. Amplification ρ ≈ 0.92 — below the critical value of 1 but close to it; the steady-state multiplier 1/(1−ρ) is approximately twelve, so a shock is amplified about twelvefold before the loop settles. The second reading organises the case: the keystone’s share of total loop gain is 100% — every feedback cycle passes through the trust sector. This follows from the wiring: depositors connect only to the trusts, the call-money market both receives stress from and returns it to the trusts, and the equity decline feeds back through the trusts’ asset values, so no feedback cycle bypasses the trust node. That is the structural definition of a trust-company panic: the sector is the single node through which all the feedback is routed.

The rescue, for the record: no central bank existed — the Federal Reserve was not founded until 1913 — so the function of lender of last resort was performed privately. J.P. Morgan organised the banks into a pool of roughly $30m (including $10m from Rockefeller) to support the trusts and the New York Stock Exchange, which had approached a forced closure; the Treasury supplemented it with deposits of existing funds in the national banks (Bruner & Carr 2007). The intervention succeeded for the reason the structure implies: with 100% of the loop gain passing through one node, a facility placed at that node intercepts the entire feedback — the limiting case of the §6 result that loop concentration determines how cheaply a rescue can be targeted. The outcome was nonetheless a severe contraction: real output fell roughly 11–12% over the following year and did not recover until 1910. The financial panic was contained; the recession was not. This separation is the case’s contribution and the origin of the report’s third containment gate. The rescue was deployed and matched to the structure, but it rested on a one-off private arrangement with no standing public capacity behind it — no institution committed to repeat it, and no monetary authority able to offset the subsequent contraction. Closing that gap was the stated purpose of the Federal Reserve Act of 1913; whether equivalent public capacity remains available today is the subject of §6.

4.2 1929 — simple coupling, no backstop, deepest collapse

The 1929 network was structurally simple by modern standards. The instability sat in extreme margin lending — buyers funded up to ~90% by brokers — and short-term broker call loans: a marginal price fall breached the margin ratio; brokers, to repay call loans, immediately liquidated client accounts; and the forced sales reached the market as price orders, an undamped self-liquidation.

Pre-crisis reading. Of the seven, 1929’s wiring was the loosest — ρ ≈ 0.79 — a flat leverage network with a single dominant margin-credit loop running through stocks and call loans. On the amplification reading alone this should have been the most containable crisis of the set: one dominant loop, one asset to support. The rescue, for the record: there was none. With the Federal Reserve tightening under the gold standard and a policy of disciplining speculation, the lender-of-last-resort function was absent (Friedman & Schwartz 1963), and the least-coupled system of the seven collapsed the furthest: real GDP fell 26% peak to trough, the Dow fell 89%, and the contraction ran a decade. 1907 and 1929 are the report’s cleanest pair on the containment axis: comparable single-hub structures, one with a private backstop at the keystone and one with none — the difference between an 11–12% contraction and a 26% one.

4.3 1987 — the trigger and the keystone are different nodes

The 1987 crash is commonly attributed to portfolio insurance, which is accurate as to the trigger but not as to the structural keystone; the network separates the two. Portfolio insurance replicated a protective put by selling stock-index futures as the market declined — a mechanical, pro-cyclical rule that sold more as prices fell further; approximately $60–90bn of equity was managed this way (Presidential Task Force 1988; Carlson 2007). On 19 October 1987 the insurers sold futures; because selling the index future was faster and cheaper than selling the underlying basket, the future moved to a large discount to the cash index; index arbitrageurs sold the cash basket and bought the future to close the discount; the cash index fell; and the lower cash index — reported with a lag of hours, so that the insurers were reacting to the futures price — generated further portfolio-insurance selling. The index fell 22.6% in one session.

Pre-crisis reading. Amplification ρ ≈ 0.91 — below the critical value, with a steady-state multiplier near eleven. Loop concentration identifies the keystone as the S&P futures market, carrying 70% of total loop gain, not the portfolio insurers. Both dominant cycles — the insurer–futures cycle and the longer futures–arbitrage–cash–insurer cycle — pass through the futures market, because that is where the order flow was routed: the insurers transacted through it, the arbitrageurs transacted on it, and the cash market took its reference price from it. The portfolio insurers were the origin of the shock; the futures market was the node that transmitted and amplified it. This reproduces, as a property of the network, the Brady Commission’s finding that the cash, futures, and options markets had become functionally a single market. The consequence is operational: an intervention aimed at the trigger acts on the wrong node, while the loop identifies the futures market as the effective point of intervention.

The rescue, for the record: the Federal Reserve supplied liquidity immediately — the first modern instance of the response later termed the “Greenspan put”. On 20 October the Fed affirmed its readiness to supply liquidity, added reserves, supported continued bank lending to securities firms (the ten largest New York banks roughly doubled it), and enabled Continental Illinois to fund the Chicago options-clearing subsidiary whose failure would have disrupted the derivatives complex. The response was effective quickly because ρ was below 1: the loop amplified the shock but was not self-sustaining — it required continued portfolio-insurance selling to propagate, and once prices had fallen enough to exhaust the hedging programmes and liquidity had been supplied, the feedback stopped. This case is the clearest basis for separating amplification from self-sustainment: a sub-critical loop is halted by removing its input, whereas a loop at or above ρ = 1 (LTCM, 2008) continues without external input. The decline was contained within days and was not followed by a recession.

4.4 LTCM 1998 — at the critical value, concentrated in one node

Long-Term Capital Management held a leveraged relative-value portfolio: long the cheaper, less liquid leg of convergence trades (off-the-run Treasuries, wide swap spreads, short equity volatility) and short the richer leg, at balance-sheet leverage near 25:1 and roughly $1.25tn of derivative notional on $4.8bn of equity (President’s Working Group 1999). The position assumed small pricing differences would narrow; the leverage made those differences equal to the firm’s capital. When Russia defaulted in August 1998, the flight to liquidity widened the spreads rather than narrowing them; the mark-to-market loss generated margin calls; meeting them required selling the positions whose spreads the fund was short, which widened those spreads and increased the loss. Because the same trades were held in correlated size across proprietary trading desks, the fund’s liquidation moved their marks as well, and their liquidation widened the same spreads (Lowenstein 2000).

Pre-crisis reading. Amplification ρ ≈ 1.001 — at the critical value of 1, the only historical reconstruction to reach it, indicating a loop that by late August was self-sustaining: it reproduced at least as much forced selling as the initiating shock without further external input. Loop concentration reads 0.27, with the fund itself carrying 93% of total loop gain. The two readings are independent and are read together: ρ is the per-round gain of the loop, while loop concentration is the number of nodes a rescue must act on to interrupt it. LTCM was simultaneously at the critical value and highly concentrated, and the combination sets the rescue arithmetic of §6.4. Because the loop passed through a single node, a facility at that node could interrupt it; but because ρ = 1, the cascade saturated substantially before a facility could be deployed, so a keystone facility avoids 43% of the damage rather than the roughly two-thirds avoided in the sub-critical single-node cases (1907, 1987). Concentration determines whether a targeted rescue is feasible; ρ determines how much damage has already accumulated when it arrives.

The rescue, for the record: on 23 September 1998 the Federal Reserve Bank of New York convened fourteen dealer banks, which contributed roughly $3.6bn to recapitalise the fund and wind it down in an orderly manner (Federal Reserve History); the coordination was public, the funds private, and the facility was placed at the keystone node. No public funds were used. The instrument matched the structure: a concentrated loop requires a targeted intervention, and recapitalising the single node carrying 93% of the loop gain is the least costly such intervention. The portfolio was liquidated over the following months without a forced sale, the fourteen counterparties absorbed their exposures, and real GDP continued to grow through 1998–99. The case appears twice in this report: as the structural predecessor of the forward network in §7 — the same leveraged relative-value strategy, conducted today through the cash-futures basis — and as one half of its clearest natural experiment. LTCM and 2008 had comparable ρ near the critical value, but LTCM’s feedback passed through one node and 2008’s through sixteen; LTCM was contained with $3.6bn at a single counterparty, while 2008 required intervention measured in trillions. Comparable ρ; different concentration; correspondingly different rescue cost.

Figure 1 — Eight US doom-loop networks: seven crises (1907, 1929, 1987, LTCM, 2008, COVID 2020, SVB 2023) and today’s complex, read left to right then down. Node fill = distress, node size = feedback centrality, gold ring = keystone (the asset a targeted backstop should buy), red = load-bearing edge. (Today’s complex is mapped in full in Figure 10.)

4.5 2008 — the deeply coupled system

2008 was a simultaneous seizure of a securitised-credit and CDS network connecting shadow banks, investment banks, and commercial banks. Subprime losses marked down AAA-rated CDOs; structured vehicles that could not roll commercial paper returned assets to their parent banks; and opaque bilateral CDS exposure triggered system-wide sales of high-quality collateral — a run on repo (Gorton & Metrick 2012).

Pre-crisis reading. Because inter-dealer derivative exposure and collateral re-hypothecation were opaque, the un-intervened system was above the critical value (ρ > 1, illustrative — §4.9): risk ran through interlocking derivative chains with no single load-bearing edge. The organising reading is loop concentration: 0.08, the lowest of the seven — sixteen comparable loops and no single node a rescue could target. This is the structural opposite of 1907 and LTCM, in which one node carried almost all the loop gain. The rescue, for the record: the broadest in history — public capital injections (TARP, $700bn), the Fed’s full set of liquidity facilities, AIG ($182bn), and FDIC guarantees. The crisis was contained at −4.3% of real GDP peak to trough (the S&P fell 57%). No purchase of a single asset was ever sufficient, and the replay of §6.4 quantifies why: a facility at the best single node avoids only 18% of the damage, against 43–68% for the concentrated cases. Concentration is what required a system-wide rescue rather than a targeted one — the same property that, read forward, tells a rescuer whether one purchase can hold the loop or the whole system must be backstopped.

4.6 Reading the wiring — the COVID 2020 network, annotated

Figure 1 asks to be read, not glanced at; this section reads one of its panels in full (Figure 2), with §3’s four questions, so that the remaining maps — including today’s, in §7.2 — can be read the same way at sight. The COVID 2020 network is the specimen, chosen deliberately: it is the direct structural predecessor of today’s complex, with the same cash-futures basis trade at its centre, so the reading developed here is the one §7 applies to the forward case. Where is the engine? The strongest cycle on the map is the basis hedge funds ⇄ cash Treasuries pair: distress flows from cash to the funds at 0.80 and returns at 0.55, so one full circuit multiplies by 0.44 — the largest round-trip on the map and 55% of all loop gain, the source of the network’s ρ of 0.85. This is the basis trade: leveraged funds holding cash Treasuries against short futures, forced to sell the cash leg as margins rise. Who carries the load? The basis hedge funds, with 32% of the network’s total exposure, and cash Treasuries with 29% — the two nodes the wiring leans on; the effective load breadth is 5.3 of 7. Which nodes form the self-reinforcing core? Four of seven — cash Treasuries, the basis hedge funds, repo, and futures: within that block distress that leaves any member can return amplified. The partition statistic reads modularity 0.11, far below the ~0.3 that marks separable community structure, so the network is one room with no internal firebreak. Where would the pressure drain? Money funds, foreign officials, and credit receive flow but do not re-amplify it; the foreign-official node has no inbound edge in this wiring — it transmits pressure but accumulates none — so the only exit is outward, into balance sheets that absorb damage rather than recycling it.

Figure 2 — Reading a network: the COVID 2020 doom loop annotated

The contrast with the single-loop cases sharpens the reading. Where 1907, 1929, and 1987 each ran through one dominant cycle that a single intervention could cut, COVID’s feedback was already distributed across the basis loop, the repo-funding triangle, and the futures-clearing channel — three comparable cycles rather than one. The network reads it: loop concentration 0.47, well above 2008’s 0.08 but below the single-loop cases, and circulation 0.65. That is why the rescue had to be broader than a targeted purchase. The Federal Reserve bought roughly $1tn of Treasuries across the market in three weeks — about 9.5% of the marketable stock by end-June, against the targeted operations of the single-loop cases — not a purchase of one asset class but a system-wide restoration of dealer capacity. The same four questions are asked of every network map that follows; in §7.2 the answers for today’s complex differ from COVID’s in one respect — the digital channels since wired into the same engine.

4.7 SVB 2023 — a duration run, caught at par

Silicon Valley Bank held a long-duration securities book — Treasuries and agency MBS bought at the low yields of 2020–21, average asset duration about six years — funded by a deposit base roughly 94% uninsured and concentrated in a single, tightly networked venture-capital ecosystem. The 2022–23 rate rise marked the book down by more than $17bn, approximately the whole equity cushion; the bank’s 8 March announcement of a $21bn securities sale at a $1.8bn loss turned that paper loss into a solvency question. The depositors, coordinated by venture firms and accelerated by social media, ran — $42bn in a single day, about a quarter of all deposits, with some $100bn more queued for the next. Meeting the run forced the sale of the available-for-sale book, which realised the loss, which deepened the solvency question, which accelerated the run; the bank failed in two days, the second-largest US bank failure at the time. The reflexive chemistry is not the forced-selling-into-price loop of the prior cases but a solvency-and-duration loop, placing it in the same family as the 1907 trust-company run (and the euro-area sovereign–bank loop of Part 2): a revaluation of solvency the leveraged sector could not absorb.

Pre-crisis reading. Amplification ρ ≈ 1.13 — above the critical value and the highest reconstruction in the record: once lit, the run reproduced more forced selling than the initiating shock. Loop concentration reads 0.24, the bank itself carrying 88% of total loop gain — the single-keystone structure of 1907 and LTCM. The case’s signature, though, is its circulation: 0.29, the lowest in the record. Where LTCM’s pressure recirculated — its forced sellers were the market, so selling fed back — SVB’s drained: the deposits had somewhere to go (money funds, the largest banks), and the rescue manufactured an exit for the underwater collateral. Concentration plus low circulation is the structural reason a targeted backstop could catch it cheaply.

The rescue, for the record: on 12 March, two days after the failure, the Federal Reserve opened the Bank Term Funding Program — lending against Treasuries, agencies, and MBS at face value rather than market value, for up to a year — alongside an FDIC systemic-risk exception that guaranteed every deposit at SVB and Signature (which failed the same day). The facility is a new point on the backstop spectrum: neither a capital injection (2008) nor pure liquidity reassurance (1987), but par lending that neutralises the duration-loss trigger itself. It worked for the reason the structure predicts — one keystone to lend against, a clean exit to drain into. Nor was it a single-bank event: unrealised losses on bank securities portfolios stood at roughly $620bn system-wide at end-2022 (FDIC), and the contagion ran on to First Republic (failed 1 May). Real GDP did not contract — the loop was broken before the same duration losses elsewhere could light the same run.

4.8 The seven cases side by side

The seven reconstructions, readings against outcomes. The structural columns — loop structure, amplification, concentration, circulation — are computed on the frozen pre-crisis networks, from data available at the time; the last two columns are what history then did. Outcomes are stated on one scale throughout the report — peak-to-trough real GDP (BEA; Romer 1989 for the pre-war estimates) — because GDP is what a containment failure ultimately costs; the familiar market drawdowns stay in parentheses as auxiliary context. The whole work of §5 is to let those two columns judge the structural ones.

One caution on reading the outcome column: it must be read with the epicentre in mind. COVID’s −10% is the pandemic’s own contraction — the rescue’s work is measured not by that number but by the financial collapse that was not added on top of it. The cleanest comparisons are the two same-structure pairs the record provides. The first is 1907 against 1929: comparable single-hub structures, one with a private backstop placed at the keystone and one with no backstop at all — −11 to −12% against −26%. The second is 2008 against 1929 — both financial-system epicentres, one caught by the broadest rescue in history and one uncaught — −4.3% against −26%, six times deeper. Those ratios, not any single row, are the backstop’s value denominated in GDP.

4.9 One honest note — the 2008 reading

2008 was a coupled system with no clean per-edge public data; it admits no clean bottom-up ρ. We therefore do not assign it a falsely precise one. It enters the comparison through its structure (the most distributed loop set, loop concentration 0.08; circulation 0.76), through the cited amplification (a subprime fundamental loss of ≈$0.3–0.5tn producing >$8tn of US equity-wealth destruction; Brunnermeier 2009), and through the same-structure containment contrasts of §4.8. Bottom-up ρ is reported for the six cases with calibrable edge data — 1907, 1929, 1987, LTCM, COVID, and SVB — and for today.

4.10 What recurs across the reconstructions

Three regularities hold across the seven US reconstructions, before any further machinery is added:

1. The ignition point is a margin or collateral revaluation, near-saturated before the shock. Crises do not begin with a day-one collapse of credit between institutions; they begin when a margin call, a haircut increase, or a mark-to-market revaluation breaches a threshold that leveraged balance sheets cannot absorb, switching the liquidation feedback on — broker call loans in 1929, dynamic-hedge selling rules in 1987, repo margin in LTCM and COVID. 1907 and SVB 2023 are the variants: the threshold there was a depositor’s loss of confidence rather than a margin call, but the mechanism is the same — a revaluation (of the trusts’ solvency in 1907, of a duration-impaired book in 2023) that the leveraged sector could not withstand.

2. The strength of the core reflexive link is stable across the record. Across crises spanning more than a century of contract design — call loans to cleared derivatives to the cash-futures basis — the calibrated spillover from one round of forced selling to the next clusters in 0.40–0.60 (mean ≈ 0.52). Forced deleveraging under margin rules appears to have a characteristic gain, whatever the era’s instruments (for the theoretical mechanism — the mutual reinforcement of funding and market liquidity — see Brunnermeier & Pedersen 2009). The convergence is a calibration regularity, not a law imposed on the networks: it is reported, and each network’s reflexive edge is set from its own record.

3. Depth contraction is the universal amplifier. Every cascade’s middle phase features a sharp contraction of market-making capacity; Duffie’s work shows order-book depth can fall ~10× when dealer balance-sheet utilisation is high even without a large yield move (Duffie et al. 2023). Today, dealer depth stands 65–85% below its pre-2008 level — the amplifier is pre-loaded before any shock arrives.

5. The record’s verdict — the dials tested, and the missing axis

With §4.8’s table in view, the record returns a split verdict on the three dials. Amplification gets the structure right and the outcomes wrong: it flags fragility in advance — the modern crises ignited from near-critical readings (1987 0.91, COVID 0.85, LTCM at the critical value) — but the loosest network of the seven produced the worst outcome, while the two readings at or above the critical value were both contained. Two networks at nearly the same near-critical reading required rescues opposite in form and three orders of magnitude apart in size. These are not noise — each mismatch is systematic. This section takes them in turn: the first exposes an axis all three structural dials miss; the others are answered by the dials of §3.2 and §3.3, for which the record now serves as an out-of-sample test, before §7 trusts them forward.

5.1 Structure does not decide the outcome — the containment axis

The sharpest mismatch is the inversion: 1929, the loosest structure of the seven (ρ ≈ 0.79), produced the deepest collapse (real GDP −26%), while 2008, above the critical value, was contained at −4.3%. The difference was not in the wiring but in the response: in 1929 the lender of last resort never came; in 2008 the rescue was the most aggressive in history. The record’s first lesson is therefore a boundary on what any structural measure can claim: amplification measures fragility, not fate. A network above ρ = 1 must amplify mathematically; whether the result is 1929 or 2008 is decided by whether a backstop catches it. The record makes the same point a second way, through its cleanest controlled pair. 1907 and 1929 ran on comparable single-hub structures — a trust-company run and a margin-credit chain, each with one node carrying nearly all the loop gain — but 1907’s panic met a private backstop placed at that node and 1929’s met none; the outcomes were −11 to −12% against −26%. Structure could not have distinguished those two fates, because the structure was nearly common to both; the backstop decided. The backstop is consequently not background but the framework’s second axis — who can catch the falling system, how fast, at what size, and at what cost. §6 makes that axis measurable.

5.2 Where to aim the rescue — testing loop concentration

The second mismatch: rescues of similar urgency took opposite forms. LTCM’s spiral, at the critical value, was cut by a $3.6bn recapitalisation of a single counterparty; 2008, at a comparable amplification reading, admitted no single-point rescue and required an undifferentiated wall across the whole system. ρ cannot express the difference, because it compresses all feedback into one number regardless of where the feedback is located; loop concentration (§3.2) recovers exactly that information.

Read across the record, the dial sorts the rescue forms. The single-hub cases are the most concentrated — 1929 at 0.67 (one margin-credit loop), and 1907, 1987, and LTCM clustered at 0.25–0.27 (one trigger, one keystone node) — and each, where a rescue was mounted at all, was contained by an intervention aimed at that one node: a private pool at the trusts, immediate liquidity through the futures market, a recapitalisation of the fund. COVID sits in the middle at 0.47 — three comparable loops, all routed through the basis hedge funds — which is why its rescue was broader than a single-node purchase but still a Treasury-market operation, not a system-wide capital programme. 2008 is the outlier at 0.08: the most distributed loop structure of the record, risk spread through interlocking derivative chains with no node carrying a dominant share of the loop gain — the configuration in which only the system-wide wall can work, which is what it took.

One distinction the historical record does not test, but on which §7 turns, is the difference between a concentrated loop and a single hub. Loop concentration asks whether one loop can be cut; the keystone’s share of loop gain asks whether there is one node every loop must cross. In the seven histories the two move together — the concentrated cases each have a dominant keystone, and 2008 has neither — but they are separable, and today’s network separates them: feedback distributed across several loops (concentration 0.30) yet every loop routed through a single node, cash Treasuries, with a keystone share of 100%. That combination — distributed loops, single hub — has no precedent in the US record, and the node a rescue would have to hold is the asset itself (§7.2).

5.3 How broad must the rescue be — testing circulation

Concentration indicates where to aim; the record still varies in how much rescue was needed. On this question the circulation fraction (§3.3) — where distress goes once the loop fires — is a weaker signal in the US record than a first reading suggests, and the weakness is itself informative. The extremes do order with rescue breadth: 1929’s 0.56, the lowest, was a structure distress could partly exit; 2008’s 0.76, the highest among the contained crises, took the broadest rescue in history. But the single-loop cases break any clean ordering: 1987 reads the highest circulation of the record (0.94) yet was contained by the cheapest rescue of the record, an immediate liquidity assurance. The resolution is that circulation does not act alone. A high circulation fraction means distress recirculates rather than draining — but recirculation only compounds if the loop is self-sustaining. 1987 was trapped (0.94) but sub-critical (ρ 0.91): the distress recirculated while the loop ran, but the loop required continued portfolio-insurance selling, so once that input was exhausted the recirculation stopped on its own. LTCM, trapped (0.87) and at the critical value, is the contrast: there recirculation did compound, and even a targeted recapitalisation left a substantial residual (§6.4). Circulation therefore reads as a conditional signal — it indicates how broad a rescue must be only for loops also at or above the critical value; below it, a trapped structure can still be cheap to contain. This is a contested-tier lens, and the US record qualifies it rather than confirming it; §7 uses it as corroboration, not as a stand-alone claim.

5.4 The drift gauge against the record — testing independence

The drift gauge of §3.4 shares no input with the network, so its test is of a different kind: not the calibrated networks, but whether one fixed rule, run on each crisis’s own market series, places the statistical break where the record says the stress was. We run the gauge on the five US histories with a clean daily series — the Dow into 1929 and again into 1987, the Baa−10y credit spread into the 1998 LTCM crisis, the TED spread (the interbank funding-stress gauge) into 2008, and the 10-year Treasury yield into March 2020 — with identical parameters and the baseline selected by the same objective rule in every case (Figure 3). The one case the test cannot reach is 1907, which predates any continuous daily market series. In four of the five, the full-sample maximum falls inside the headline crisis window with no hand-set baseline: the Dow peaks on 19 November 1929 (distance 1.60) and again on 19 November 1987 (1.71), the Baa−10y spread on 2 November 1998 (1.46), and Treasuries on 23 March 2020 (1.42), each confirmed by a trailing z-score (3.0, 2.9, 4.2, 3.9).

The 2008 case is the most instructive. On the funding gauge the maximum — 1.72 — falls not at Lehman but in the first week of September 2007, twelve months before the headline collapse, at the money-market freeze that the run-on-repo literature identifies as the true start of the crisis (Gorton & Metrick 2012; Taylor & Williams 2009). The distance then stays elevated for a year and surges back to within one percent of its maximum (1.70) in the post-Lehman weeks, where the trailing z fires again (2.6 on 3 October 2008). At Lehman’s own onset the trailing z is negative: by September 2008, a full year of broken funding regime already sat in the gauge’s history, so the collapse week was not anomalous against it. Read plainly, the gauge dated the regime break — and on the funding side, the regime broke in 2007.

What this test establishes is deliberately limited. The crisis windows were fixed from the headline record before the runs, and the baseline rule is causal — calm is defined only from data preceding the period under study — but the exercise is retrospective, in-sample, and limited to the histories with a clean daily series. The gauge is a regime lens, not a crisis-week timer: it also lifts at benign regime changes — the 2019 turn to easing, the final melt-up of 1928 — and after 1929 it never returns to baseline at all, which is itself the honest reading: that regime never came back.

Figure 3 — The drift gauge run through the US record: Fisher–Rao distance of each crisis’s own daily series from an automatically selected calm baseline, identical parameters in every panel.

Run forward with the same rule and parameters, the gauge reads the 10-year Treasury regime at 0.50 as of 11 June 2026 — the 12th percentile of its post-2018 record, trailing z −0.8 (Figure 4). The in-sample peak is the sustained 2022–23 rate shock; March 2020 lifts clearly above baseline. That is the qualification a cross-check needs — agreement with realised stress, achieved with zero shared input — and no more is claimed for it.

The configuration that matters most is divergence. When the structural dials read taut while the drift gauge reads calm, the calm is not a comfort: it means the wiring has tightened without yet being tested — and, historically, that configuration has coincided with new, uncalibrated leverage instruments incubating at the network’s edge. That is exactly today’s configuration (Figure 4), and it is what directs attention to the new channels of §7.3. Contested-tier, descriptive.

Figure 4 — The same gauge run forward: distance of the 10-year-Treasury return regime from its automatically selected calm baseline

5.5 The structural dial through time — anchored reconstructions

The dials so far are snapshots: each ρ in §4.8 is a single number frozen on a crisis eve. A snapshot cannot answer the question a monitor actually faces — not “how taut is the wiring?” but “how fast is it tightening, and when did it cross?” Answering it needs one thing: a clean public record, through time, of the forced-seller’s own position — the leverage or size that drives the reflexive edge, not a price or spread that merely reacts to it. Three of the seven US histories supply one, as does today’s complex; the other four do not. 1929’s call-money stock survives monthly; 2008’s shadow-bank funding run is traced by asset-backed commercial paper outstanding; and the cash-futures basis trade behind both the March-2020 dislocation and today’s complex is reported weekly by the CFTC. Against these we extend each network through time by anchored reconstruction (Appendix A.1): topology and every edge without a formula-backed series stay frozen, the supported edges move with the data, and ρ is recomputed month-end — pinned by construction to the published calibration at each eve, so a path extends its snapshot rather than re-deriving it. The remaining four — portfolio insurers in 1987, a single private fund in 1998, trust-company call loans in 1907, and SVB’s 2023 duration-loss run — left no continuous public series of the driving position; their dials stay snapshots.

The fullest record is 1929’s: brokers’ loans monthly from the Federal Reserve’s Banking and Monetary Statistics, lender composition quarterly, the daily Dow for the price leg. Two of the network’s eight edges move — the margin-ignition edge as the ratio of brokers’ loans to the Dow, the funding-composition edge as the non-bank share of those loans — the other six frozen at their §4.4 values, ρ recomputed each month-end from January 1928 to December 1932; by construction ρ(September 1929) = 0.790.

What the reconstruction shows first is not a slow climb but a high plateau. Across the entire melt-up — January 1928 to September 1929 — ρ stays between 0.745 and 0.790: the loop was wound tight the whole time, fluctuating with the race between loan growth and price growth, and the crossing into the crisis band (whose lower edge, 0.79, is the lowest crisis-eve reading in the calibrated record) is a single month — September 1929 itself. The mechanism of the crossing is the instructive part. In September brokers’ loans added a further $667 million — the largest monthly jump in the series — carrying the stock to its $8.5 billion peak, while after September 3 the Dow turned down, ending the month a tenth below its top; the ratio that drives the ignition edge spikes precisely when the debt leg is still growing and the price leg has turned. One month earlier, in August 1929, the reconstruction reads its melt-up minimum (0.745): on this gauge the system looked marginally calmer than at any point in the prior year, four weeks before the crossing. A monthly structural reading is not a timing instrument — but it places the regime correctly: all twenty-one month-ends from January 1928 through September 1929 read near-critical, and the final crossing is dated to the month the loans were still growing and the price had turned.

The second data-driven edge tells a different story, and the difference is itself the finding. The lender composition of call money deteriorated through the boom: the non-bank share — corporations and out-of-town lenders with no Federal Reserve access and no reason to stay — climbed from 0.41 in early 1928 to 0.78 by September 1929. Yet this movement leaves ρ exactly unchanged, and the network says why: the non-bank funding edge feeds the loop but sits inside no closed cycle, so it shifts the composition of the funding base, not the loop’s gain. The two data-driven edges thus separate along the same division §3 draws between strength and location: one series moved the amplification; the other decided who would be holding the funding when it mattered — and who would leave first. A single dial would have averaged the two stories; the network keeps them apart.

The downslope is as informative as the climb. After October the reconstruction reads the unwind in the data’s own order: the ignition edge collapses from 0.85 to 0.76 by end-October and 0.57 by end-November — brokers’ loans were called faster than prices fell, the margin machinery of §4.4 running in reverse, liquidation outrunning the decline it caused — while the non-bank share drops from 0.78 to 0.60 in November and 0.49 by February 1930: the lenders with no reserve access left first, halving the loop’s funding base within five months. By July 1932, with the Dow at its Depression floor and the loan stock nearly liquidated, ρ reads 0.554 — the loop unwound not by absorption but by the destruction of both legs. The 1929 row of §4.8 records that no rescue came; the reconstruction adds what that looked like month by month: a structure exiting the near-critical band downward — through liquidation, not containment.

The other three reconstructions are shorter records of the same shape. 2008, driven by asset-backed commercial paper outstanding, loads from 2004 and crests in mid-2007 — the same funding break the drift gauge of §5.4 marks from the opposite side, its TED-spread regime maximum of September 2007, twelve months before Lehman. Two lenses, one independent series each, converge on one episode. (2008’s level sits above ρ = 1 throughout, but its couplings are face values, not calibrated — §4 — so the panel is drawn dashed and only its shape and timing are read, never its height.) COVID is driven by the CFTC basis-trade position: the channel loads through 2017–19, crosses ρ = 1 around the September-2019 repo crisis — where the ignition edge saturates, so the crossing is a floor on the true reading — settles at 0.85 on the February-2020 eve, then collapses to 0.50 by year-end as the trade is force-unwound in the March-2020 dash-for-cash. Load, crest, eve, unwind, in one panel. Today runs on the same CFTC series, and its trajectory is the report’s central observation made dynamic: the basis channel went dark in 2021–22 (leveraged money turned net long — the 2020 unwind had wiped the trade out, and ρ fell to the residual-network floor of 0.67), then rebuilt from late 2022, crossing ρ = 1 at the basis peaks of 2024–25 and standing at 0.98 at end-2025. It is the only one of the four still loaded — at the record level of its own series, with no unwind yet.

Figure 5 — The structural dial through time: amplification ρ reconstructed month-end for the four US networks whose forced-seller position is publicly recorded.

Read against §5.4, these reconstructions complete a pairing the framework needs. The drift gauge gave the data side of each crisis a time record; the anchored reconstruction now gives the structural side one — and where a case carries both, 1929, 2008 and COVID, the two lenses agree on the regime, most sharply in 2008, where structure and data crest on the same 2007 funding break from opposite directions. Today is that comparison run forward and unresolved: the structural reading (0.98) sits above every month of the 1929 path, melt-up included, and crossed ρ = 1 outright at the 2024–25 basis peaks, while the drift gauge of §5.4 reads the 12th percentile of its own record — structure taut, behaviour calm, the divergence §5.4 names as the configuration that directs attention to the untested channels, now drawn as a trajectory rather than a point. The historical panels are what the two time records look like when they converge — at the break; the present is what they look like while they have not yet.

5.6 What the anatomy adds

The dial verdicts above compress the record; before compressing, it is worth reading the seven networks’ raw anatomy — partition, core, load, the §4.6 questions asked across all seven histories at once, today alongside (Figure 6). Three findings emerge that the dials alone do not state.

First, no pre-crisis network has internal firebreaks. Community structure is weak in all seven — every case well below the ~0.3 modularity band that marks strong partitioning (COVID, for instance, reads 0.11). A crisis-eve system is one room: the self-reinforcing core reaches across the network with no internal boundary separating the engine from the rest. That gives the containment record of §6 its structural reading: every containment in the record was an external absorbing node grafted onto the network — a facility, a purchase programme, a wall of guarantees — and never an internal boundary holding on its own. 1929 is the case with no graft, and nothing internal held.

Second, the heaviest load has migrated from the intermediaries to the asset, and today is the first US case to load the asset. In every prior US crisis the most-loaded node was a leveraged intermediary or a trading venue: margin accounts carrying 26% of total exposure in 1929, the basis hedge funds 32% in 2020, broker-dealers 23% in 2008, and in the single-hub cases the trusts (1907), the futures market (1987), and the fund itself (LTCM). Today, for the first time in the US record, the heaviest load sits on the asset itself — cash Treasuries, at 34%. The operational meaning is sharper than the statistic: in every historical case, “support the most-loaded node” meant lending to or buying from an institution; today the most-loaded node is the asset, so supporting it means buying the market. The keystone problem of §7 is already visible in the anatomy.

Third, loop concentration and load concentration are different axes, and a rescue needs both. The two do not move together. 2008 is the extreme of feedback dispersion — sixteen comparable loops, concentration 0.08 — yet its load was only moderately concentrated on the broker-dealers; it was the feedback axis, not the load axis, that defeated every single point and forced the system-wide wall. 1929 is the mirror image: the most concentrated loop structure of the record (0.67) carried one of the most evenly spread loads. The single-hub cases — 1907, 1987, LTCM — concentrated both: one dominant loop and one loaded node, which is why each was contained, where a rescue came, by one targeted intervention. Today’s network is mixed in the uncomfortable direction: feedback more dispersed than any historical single-loop case (0.30), yet the load concentrated on the asset itself — a combination with no precedent in the US record. §6 asks whether the one resource that can substitute for a purchase in such a structure — an unquestioned capacity to stand behind the commitment — is available here.

Figure 6 — Eight networks, one anatomy: partition (community modularity, with the >0.3 strong-partition band unreached in every case), the self-reinforcing core (nodes inside / total), and the heaviest single load (share, and who carries it).

5.7 A systematic framework

The record’s split verdict points to a single judgment: fragility is not a scalar that any one indicator can exhaust, but the state of one high-dimensional object — the sovereign-debt system of holders, funders, and assets, coupled — that must be read along several complementary directions at once. On that basis, this report proposes a systematic framework: three structural dials read the feedback from inside the coupled structure — its strength (§3.1), its location (§3.2), its destination (§3.3); one independent statistical gauge (§3.4) cross-checks from outside; beneath the dials, the anatomy of §§4.6 and 5.6 — partition, core, load — supplies the descriptive base layer the dials compress; and the containment axis of §6 decides the outcome of a taut structure. Figure 7 places the structural readings of the seven crises and today side by side as fragility fingerprints; the division of labour and the record’s tests are summarised below:

Three disciplines govern the dials. They are computed independently of one another. None is a forecast — each reads the structure as wired today. And they deliberately answer only half the question: whether a taut structure actually runs away depends on whether anyone catches it. That is the containment axis, taken up next.

Figure 7 — Fragility fingerprint across the seven US histories and today: amplification ρ, loop concentration, and circulation fraction, with the realised outcome beneath each group.

6. The containment axis — why structure is not destiny

The inversion of §5.1 (1929: loosest structure, worst outcome; 2008: tightest, contained) exposes the blind spot of conventional systemic-risk measurement: it attends to how fragile the structure is and ignores the capacity of the official sector to catch it. Making that capacity measurable is the framework’s second axis — and the divergence between the structural dials and the realised outcomes is precisely where it shows up.

6.1 Three gates

A self-amplifying spiral is arrested in practice only if the rescue passes three gates:

1. Deployed — speed and will. Before the cascade reaches ρ = 1, the lender of last resort must be willing to act beyond convention. In 1929 this gate was shut: doctrine forbade intervention and the Fed was tightening. The lesson was internalised afterward — by Bernanke, a Depression scholar, in 2008, and by Powell in March 2020 — which is why those crises were backstopped at once; in 1987 the Fed acted the morning after the crash. 1907 shows the gate passing without any public institution at all: the private pool was assembled within days, but its availability depended on one banker’s coordination rather than a standing mandate. The gate can be passed by improvisation — but improvisation is not a capacity a system can rely on in advance, which is the reason the Federal Reserve was created.

2. Sized and type-matched. The tool must cover the cascade’s funding gap and match the loop geometry. In a concentrated structure the rescue is a single point: a pool at the trust sector (1907), liquidity supplied through the futures market (1987), a recapitalisation of one fund (LTCM) — in each, one node carries nearly all the loop gain and is cut at that node. In a distributed structure (2008) any point rescue leaks around the edges, and the backstop must be undifferentiated and system-wide. The two forms sit orders of magnitude apart in cost, for the reason the §6.4 replay quantifies: a facility at the keystone avoids 43–68% of the damage in the concentrated cases but only 18% in 2008. The gate is about matching the node and the quantifier, not the amount — the keystone names which asset to buy, and the circulation fraction, conditional on ρ (§5.3), indicates how broad the rescue must be.

3. Affordable — fiscal and monetary space. There must be room to fund the rescue without breaking something else. For the United States the room has a specific, deep source: the exorbitant privilege (Gourinchas, Rey & Govillot 2017) — because the US supplies the world’s safe asset, it funds itself cheaply in its own currency, which is precisely why the Fed could backstop in 2008 and 2020 at no visible fiscal cost. The privilege carries a matching duty: in a global crisis the safe-asset supplier absorbs losses for the rest of the world (a wealth transfer on the order of ~19% of US GDP across 2007–09). The affordability gate therefore reduces to one question — is the privilege intact? — bounded on the domestic side by the debt-and-inflation arithmetic of the long debt cycle (Dalio 2025) and on the international side by the credibility of the safe asset (Rey 2015). This is the one gate the historical US record cannot test from inside, because in every prior case it was open: the affordability behind a US rescue was never in doubt. The forward case is the first in which it is, and §6.2 is its arithmetic.

6.2 How the 2026 fundamentals bite the third gate

In 2026 the affordability gate is changing in kind, not merely in degree:

• Deficits and interest expense. On the CBO’s February 2026 baseline, the FY2026 deficit is ≈$1.9tn (5.8% of GDP); net interest, at ≈$1.0tn (3.3% of GDP), now exceeds both defence ($885bn) and Medicaid ($708bn), absorbs ≈18% of federal revenue, and is the budget’s fastest-growing line — reaching $2.1tn (4.6% of GDP) by 2036, with debt held by the public rising from ≈100% to ≈120% of GDP over the same window. A market-functioning rescue mounted against this fiscal backdrop is no longer obviously costless.

• The privilege visibly eroding. Rey’s recent work (2025–26) documents the erosion in real time: episodes through 2025 in which the dollar weakened even as Treasury yields rose and equities fell — breaking the usual safe-haven pattern — and US-Treasury-vs-other-sovereign spreads widening, consistent with a rising fiscal risk premium in the anchor asset itself.

• Gold overtakes Treasuries in the world’s reserves. Per the ECB (2026), gold reached ≈27% of global central-bank reserve assets at end-2025, overtaking US Treasuries (≈22%) as the largest single reserve asset for the first time since 1996 (dollar-denominated assets overall remain the largest bloc, ≈42%). Reserve managers — the most conservative investors in the world — are quietly rotating away from the asset this report is about. That is the demand side of the affordability gate weakening.

• A new and less predictable reaction function. Kevin Warsh was confirmed (54–45) and sworn in as Fed chair in May 2026 (U.S. Senate 2026); Powell, unusually, remains on the Board — the first former chair to do so in ~75 years. Warsh’s record is hawkish, yet he arrives under sustained political pressure for cuts while tariff pass-through keeps inflation sticky (April 2026 PCE 3.8% y/y, core 3.3%; BEA, May 2026 release). How this Fed weighs market-functioning purchases against inflation credibility — gate 1 — is now genuinely uncertain in a way it was not in 2008, 2020, or 2022.

6.3 The seven US cases and today — the containment scorecard

The three gates, assessed across the seven US histories and today, are summarised in the table below and rendered as the scorecard of Figure 8. To avoid repeating §4.8, the table carries over only amplification ρ — the one dial that pairs structure against the gates, with today added — while loop concentration, circulation and the keystone stay in §4.8, Figure 7 and the structure × containment plane of Figure 11. The gate rows are sourced judgement, scored pass = 1, partial = ½, fail = 0.

Figure 8 — The containment scorecard: three gates (deployed × sized-and-type-matched × affordable) across the seven US histories and today.

6.4 The record, replayed — rescue on, rescue off

The scorecard above is sourced judgement; the cascade arithmetic of §7 lets it be cross-examined. We place each US historical network under the same standardised conventions — the same shock (0.30 at the case’s distressed node), the same facility mechanics (interception of 90% of the flow arriving at the rescue node), and run each history twice: once with no intervention, once with a purchase facility at the node the rescuer actually supported (or, where no rescue came, at the node the loop analysis names), arriving at the round corresponding to the recorded response. This is a consistency check, not an out-of-sample test — the networks are calibrated on these same histories, and cascade rounds map onto calendar time only ordinally. What it tests is whether the dials’ division of labour survives dynamics: whether a structure that reads as single-point rescuable actually is, once the rescue is switched on inside the cascade rather than asserted beside it. Damage is reported on a scale the networks share: structural damage is terminal distress summed over a network’s nodes, and since each node’s distress saturates at 1, a network of N balance sheets has a saturation ceiling of N. Figure 9 therefore draws each run as a share of its own ceiling, which makes a six-node 1929 comparable with a ten-node 2008. Three results carry the comparison (Figure 9).

First, the concentrated single-hub crises are single-point rescuable, and the replay prices it: one facility at one node removes 43–68% of total structural damage — 68% for 1907, 66% for 1987, 47% for COVID, 45% for 1929, and 43% for LTCM. On the ceiling scale the sub-critical single-hub runs converge toward containment: unrescued they saturate 38–54% of their networks; the facility holds them at 17–24%. The bitterest row is 1929: the most concentrated loop structure of the record (0.67), therefore among the most single-point-rescuable — a margin/call-loan backstop arriving early cuts the cascade nearly in half — and the only case in which no rescue ever came. Friedman and Schwartz’s verdict, restated in cascade arithmetic: the Depression’s depth was not wired into the structure; it was the price of the missing graft.

Second, 2008’s dispersion defeats any single point dynamically, not just statically. The best available single-node facility, granted the most favourable early arrival, avoids just 18% of the damage; at the realised post-Lehman timing it avoids nothing. 2008 is also the deepest cascade on the shared scale — 69% of its ceiling unrescued, still 57% with the best single point: the only history the facility cannot pull back into the band. The static dial (concentration 0.08) and the dynamic replay agree: containment required breadth — capital injections, AIG, commercial-paper and money-fund facilities at once — which is what it took.

Third, concentration decides whether a point rescue is possible; ρ decides how much it can still save. LTCM separates the two. Its loop is concentrated (0.27) with a clear keystone, so a targeted recapitalisation was the right instrument — but because the loop sat at the critical value, the cascade had already saturated to 58% of ceiling by the time a facility could arrive, and the recapitalisation pulls it only to 33%, avoiding 43% — the low end of the band, beneath the 66–68% of the sub-critical single-hub cases (1907, 1987). The reason is the same 1/(1−ρ) gain the structure reads as fragility: at ρ = 1 the loop reproduces a shock fully each round, so the damage already done by the time a rescue is mounted is at its maximum. A structure can be single-point rescuable and still expensive to have caught late: concentration is necessary for a cheap rescue; only a sub-critical loop is also forgiving of delay.

The same arithmetic, run forward on today’s network, is the business of §7.7 — and the seven replayed histories are its measuring stick. They mark the band on the shared scale: the sub-critical single-hub structures held at 17–24% of ceiling by a single-point facility, against 38–54% unrescued; and 57%, the floor 2008’s dispersion could not be brought below. Where today’s runs land relative to those marks is the first thing §7.7 reads off.

Figure 9 — The record, replayed: the seven calibrated US historical networks under identical conventions — the same standardised shock, no intervention (red) versus a purchase facility at the historically chosen or counterfactual rescue node (teal); bar length is terminal damage as a share of that network’s saturation ceiling (all N balance sheets fully distressed), making networks of six to ten nodes comparable; ● marks the branch history took.

7. Today’s US Treasury complex

Figure 10 — Today’s US Treasury complex. Keystone = cash Treasuries (gold ring); the basis hedge-fund node carries the most feedback centrality; stablecoins, tokenised assets and crypto/DeFi are the channels new since 2020.

7.1 The reading: both axes stressed at once

Read against the calibrated record, today’s network (Figure 10) is wired at crisis-eve tautness — with one property no historical case shares.

• Amplification ρ ≈ 0.98 — within 0.02 of the runaway threshold ρ = 1; only LTCM (at 1.00) and the supercritical 2008 read higher, and both broke. At this gain a forced-selling shock is multiplied more than fortyfold before it dies out — if nothing else gives way first. Two readings bracket it, and the gap between them is itself the finding. The nine observable market nodes alone — holders, funders, and the asset — read ρ ≈ 0.90, already at the top of the historical contained range (1987’s 0.91, 1907’s 0.92). Adding the fiscal-debasement layer — the Fed-backstop → dollar → gold loop that opens only once a rescue is monetised (§7.2) — lifts the reading to 0.977, above every contained crisis eve in the record and the figure this report carries. The ≈0.08 between them is that loop’s contribution: the second feedback loop no history had.

• Loop concentration ≈ 0.30, lower than every historical single-loop case. The feedback now routes through more sectors of comparable weight — the basis trade, the stablecoin pair, the repo funding triangle, the crypto gateway — so a narrowly targeted rescue is less likely to suffice than a single-node purchase was in any historical single-loop case — though every one of those loops still routes through one hub, a configuration with no precedent in the US record (§7.2).

• The keystone is cash Treasuries itself. This is the uncomfortable novelty. In 2020 the type-matched rescue was “buy Treasuries until dealers breathe”; in 1998 it was “recapitalise the fund”; in 1987 it was “supply liquidity through the venue.” Today the structurally indicated rescue is again to support the keystone — but the keystone is now the very asset whose oversupply, sticky-inflation backdrop, and eroding reserve status are the problem (§6.2). The intervention the structure calls for is the intervention the macro position makes costly: the structural axis and the containment axis have fused into one node. The US record holds no precedent for this fusion: in every prior case the asset under stress and the asset behind the rescue were different. Whether a commitment could substitute for a purchase here — as it can only while the affordability behind it is beyond question — is §6.2’s subject.

• Circulation ≈ 0.68 — above both 1929’s 0.56 and COVID’s 0.65: distress would strongly recirculate rather than drain, pointing to a rescue broader than a single-node purchase even in the benign case.

• The amplifier is pre-loaded — and still loading. The basis trade that is the keystone’s reflexive engine (§7.2) stands at an all-time record: CFTC leveraged-money net-short positions in 10-year Treasury futures reached ≈2.5 million contracts in 2025 — roughly 3.3× the level on the eve of the March-2020 break that required the largest official backstop in history to arrest, and 2× the 2019 peak whose own partial unwind forced the September-2019 repo intervention. The §5.5 reconstruction reads this as ρ crossing 1.0 in 2024–25 and standing at 0.98 today. Dealer depth meanwhile remains 65–85% below pre-2008 (§4.10), with the debt stock still growing on the CBO path.

None of these dials forecasts a date. Together they say: the structure is wired at the tautness from which past crises ignited, the rescue it would call for is broader than the last one, and §6’s scorecard says the capacity to mount that rescue is the weakest of any contained case on record.

As this chapter’s synthesis, the seven histories and the two failure modes can be placed up front on one plane (Figure 11) — the modes themselves are derived only in §7.7, so here the plane is read purely as a locator. The horizontal axis is loop concentration — how far the feedback can be cut at a single point; the vertical axis quantifies §6.3’s scorecard — how many containment gates the rescue met, scoring pass = 1, partial = ½, fail = 0. The corners sort the record cleanly: the contained crises occupy the upper half — 2008 at the far left, where only the system-wide wall holds; 1987, LTCM and COVID to the right of it, where one targeted intervention sufficed; and 1907 between them, its first two gates met but its third only half — a private rescue with no public capacity behind it. 1929 sits alone in the lower right: among the most rescuable structures in the record, and the only one with no rescuer. Today enters as a range, not a point. Horizontally it sits in the dispersed region — well left of the single-hub crises 1929 and COVID; the digital channels and the new debasement loop have spread the feedback, and without the digital channels it reads 0.44, more concentrated but still short of the single-hub regime. Vertically it spans a fork decided entirely by the containment response: resolve toward mode B — a rescue that is ample but monetised — and the point rises toward the contained band, gates nominally met but the financing monetised; resolve toward mode A — a rescue delayed or insufficient — and it falls past 1929’s row into the lower-left quadrant, which no US history has occupied — feedback too distributed for a targeted rescue, gates too eroded to mount the broad one.

Figure 11 — Structure × containment: the seven US histories and today on one plane. Horizontal: loop concentration — how far the feedback can be cut at one point (Appendix A.2). Vertical: containment gates met, quantifying the §6.3 scorecard (pass = 1, partial = ½, fail = 0).

7.2 The anatomy of today’s complex

§7.1 is the compressed reading. Decompressed through the four questions of §4.6, today’s map yields four answers, each with an operational edge (Figures 12–15).

Where is the engine? Today’s network contains five closed loops, and Figure 12 ranks them: the Treasury–basis engine (cash Treasuries ⇄ basis hedge funds, round-trip gain 0.44, 47% of all loop gain on the map); the new digital pair (cash Treasuries ⇄ stablecoins, 0.18, 19%); the fiscal-debasement loop (cash Treasuries → Fed backstop → the dollar → back, 0.15, 16%); the funding triangle through repo (0.14, 15%); and the gateway path through crypto and stablecoins (0.03, 4%). The first two engines carry two-thirds of the feedback — and one fact organises everything: all five loops pass through cash Treasuries. The keystone’s share of loop gain is 100%. There is no feedback in today’s complex that does not route through the asset itself; that is the structural content of §7.1’s claim that the structural axis and the containment axis have fused into one node. No US history shares this property: today is the first network in the record in which every loop routes through the asset itself — and the debasement loop, closing cash → backstop → dollar → cash, is the one no historical network carried and the one that lifts amplification to its near-critical 0.98.

Which nodes form the self-reinforcing core? Seven of twelve (Figure 13): cash Treasuries, the basis hedge funds, repo, stablecoins and crypto/DeFi — and, new today, the Fed backstop and the dollar, pulled into the core by the debasement loop. The digital channels are not bolted onto the engine room; they are inside it, which is why excising them (§7.3) changes the loop structure itself rather than shaving the totals. Money funds, foreign officials, tokenised assets, private credit, and gold sit on one-way roads out of the core: they feed or receive, and none of them re-amplifies. The partition question returns the same answer as every eve before it — modularity 0.26, one room — and the nominal communities the algorithm does find split the funding side (hedge funds, money funds, repo) from a cash-centred block holding the digital channels, the backstop layer and the remaining periphery: the new channels and the rescue layer share a compartment with the keystone, with no internal boundary between them.

Who carries the load? Cash Treasuries, 34% of total network exposure — then the basis hedge funds at 14%, the dollar at 10%, repo at 9%, with stablecoins and the Fed backstop near 7–8% each (Figure 14). The effective load breadth is 8.2 of 12: spread, but with the heaviest single share of any network in the record, and the first to sit on the asset rather than on an intermediary (§5.6’s migration finding, read on the live map).

When does the pressure arrive where? The anatomy above is static; Figure 15 adds the clock, running the standardised keystone shock with no intervention and recording each node’s distress round by round. The first ring — hedge funds, repo, stablecoins, crypto, private credit — is hit in round 1; money funds by round 2; foreign officials and tokenised assets, with no inbound edge in this wiring, transmit pressure but accumulate none. By round 2, 35% of the terminal damage is already in place; by round 6, 76%; nine rounds in, ≥95%. That timetable is the quantity the delay sweep of §7.7 prices: an early facility is leveraged precisely because two-thirds of the damage is still to come at round 2; by round 6 three-quarters of it is already in place, and by round 9 almost all of it.

Figure 12 — Today’s loop inventory: every directed cycle in the calibrated network, ranked by round-trip gain.

Figure 13 — The self-reinforcing core of today’s complex: seven of twelve nodes — cash Treasuries, basis hedge funds, repo, stablecoins, crypto/DeFi, and the Fed backstop and dollar of the debasement loop — form one strongly connected block (shaded): distress leaving any of them can return amplified.

Figure 14 — Who carries the load: each balance sheet’s share of total network exposure, with dashed marks where the heaviest load sat on past crisis eves (margin accounts 26%, broker-dealers 23%, basis hedge funds 32%).

Figure 15 — The uncontrolled cascade, round by round: the standardised cash-Treasury shock with no intervention; cell colour = cumulative distress, outlined cells = first arrival, top strip = share of terminal damage already in place.

7.3 The channels new since 2020

Three transmission channels in today’s network did not exist, or were immature, in March 2020 — each arrived at the network’s edge, exactly where §5.4 says uncalibrated leverage incubates, and two of the three have since been absorbed into the self-reinforcing core (§7.2).

• Stablecoins now siphon T-bills. Dollar stablecoins have grown to several hundred billion dollars of reserves (Federal Reserve 2026), held 80–90% in short-dated Treasuries and Treasury repo under the 2025 GENIUS Act framework — a non-trivial marginal T-bill buyer whose liabilities are redeemable on demand. The Bank Policy Institute’s model (BPI 2026) shows the growth is not balance-sheet-neutral: yield-bearing stablecoins drain uninsured bank deposits, forcing banks to shrink the repo funding they extend to dealers and hedge funds — thinning the very buffer the loop needs. A large stablecoin redemption wave is mechanically a forced T-bill liquidation at machine speed.

• Crypto leverage runs on the same rails. Crypto and DeFi markets do not hold Treasuries at scale themselves; they connect to the Treasury complex through the stablecoin gateway. Stablecoins are crypto trading’s settlement leg and margin collateral, so a deleveraging event in crypto — exchange margin calls, DeFi liquidation cascades — converts at machine speed into stablecoin redemptions, and redemptions convert into T-bill sales by the issuers. In calm markets the channel runs one way: trading demand grows the stablecoin float, and the float grows T-bill demand. Under stress it reverses into a forced-selling feeder.

• Tokenised Treasuries liquidate at machine speed. Tokenised Treasury funds have grown into the low tens of billions. In calm markets tokenisation improves collateral mobility; under stress, smart contracts execute instant, undamped, market-price liquidation of pledged collateral — converting tokenised Treasuries into sell orders faster than any dealer can intermediate, and faster than any 2020-era assumption about how quickly collateral chains unwind.

Naming the channels is not measuring them. Because all three are wired into the calibrated network, their role can be computed rather than asserted: remove them and re-read the dials; place the same shock at different origins; let their couplings keep growing. Three results follow (Figures 16–19).

1. The new channels are a second engine — and part of why the feedback is now distributed. Excising all three (stablecoins, crypto/DeFi, tokenised Treasuries) drops amplification from ρ ≈ 0.98 to ≈ 0.88 — a tenth of the gain comes from the digital edge alone. Excising stablecoins alone reproduces the entire drop; tokenised Treasuries alone leave ρ unchanged (a one-way buyer, loop-inert), and crypto/DeFi matters only through the stablecoin gateway. The stablecoin⇄cash-Treasury loop (gain ≈ 0.18) is the network’s second amplification engine after the basis-trade loop (≈ 0.44). The same excision raises loop concentration from 0.30 to 0.44: the digital channels are among what disperses today’s feedback away from the single-dominant-loop pattern every historical rescue exploited. What the channels add is standing amplification and a dispersed rescue target — not a larger one-shot cascade: on a single cash-Treasury shock the nine core balance sheets take only ≈ 3% more distress with them wired in than without.

Figure 16 — The new channels, excised: removing the three digital channels from today’s calibrated network drops amplification ρ from 0.98 to 0.88 and raises loop concentration from 0.30 to 0.44.

Figure 17 — Which channel carries it: amplification ρ under single-channel removals. Stablecoins alone reproduce the entire drop to 0.79; crypto/DeFi and tokenised Treasuries barely move the dial.

2. A stablecoin redemption run is now a basis-trade-sized event. Placing the same shock at different origins: a stablecoin-origin redemption run produces structural damage ≈ 3.0 — parity with the basis-fund unwind (≈ 3.0) that the stability literature names as the epicentre — and far above a crypto-origin shock (≈ 1.2), of whose flow arriving at the keystone ≈ 42% enters through the stablecoin gateway. The keystone facility of §7.7 is, if anything, more effective against these edge-origin shocks — damage −70% for a stablecoin origin versus −47% for a keystone origin — because interception at the keystone begins before the keystone saturates.

Figure 18 — Same shock, different origin: structural damage for the same 0.30 shock entering at four origins, without and with the §7.7 keystone facility.

3. On the current growth path, the loop closes on its own. Scaling the five digital couplings jointly by a growth multiple f: today (f = 1) already reads ρ ≈ 0.98, above every historical eve; at f ≈ 1.1 — barely a tenth of further coupling growth — it crosses ρ = 1, where the loop self-sustains with no external shock required. The sweep scales couplings, not market size, and is a stylised exercise, not a projection. Its point is a distance: what separates today’s wiring from the report’s only hard line is not a tail event but continued growth in precisely the channels §5.4 flags as incubating at the network’s edge.

Figure 19 — If the digital couplings keep growing: scaling the five digital couplings jointly by a growth multiple f, today’s network (f = 1, ρ ≈ 0.98, already above every historical eve) crosses the report’s only hard line ρ = 1 at f ≈ 1.1.

7.4 Timing — the refinancing wall and the funding cycle

The dials of §7.1–7.2 read how taut the wiring is; they are silent on when the load peaks. A second instrument, sharing no input with them, reads that: the Treasury’s own maturity profile against the funding system’s spare capacity, both from public data (the Treasury’s Monthly Statement of the Public Debt and the Federal Reserve’s reserve, reverse-repo and bill-rate series). The calendar is recomputed by an operator from those feeds, not asserted.

The wall is large and near. Of $30.75tn in marketable Treasuries outstanding (May 2026), 33% — $10.2tn — matures within a year, and the near months are the heaviest: roughly $2.4tn rolls in June 2026 and $2.0tn in July, the bulk of it bills. Rolling is routine; the cost is not. The maturing coupon securities carry an outstanding-weighted coupon near 3.0% and refinance into a front end of about 3.8% (3-month bills) to 3.9% (1-year) — so each turn of the wall ratchets the interest bill higher, a coupon-reset ladder that compounds for as long as the front end stays elevated.

What absorbs that issuance has thinned to the point of being the story. Bank reserves stand at about $3.08tn — only ~$80bn above the ~$3.0tn “lowest comfortable level” below which funding markets seize (a sourced parameter, not a forecast). The overnight reverse-repo facility that held $2.5tn as recently as 2022–23 — the buffer that let the Treasury issue without draining reserves — is drained to under $1bn. With the buffer gone, new issuance now lands directly on reserves. The one mercy is honest and worth stating: on the trailing 26-week trend reserves are flat-to-rising, not falling, so there is no dated countdown to scarcity — the cushion is thin, not yet shrinking.

Read with the dials, the timing instrument sharpens the forward picture rather than softening it: the structure is near-critical (§7.1), and the calendar puts its heaviest, most expensive refinancing in the same near-term window — a June–July convergence with quarter-end — against a reserve cushion with no shock-absorber left beneath it (Figure 20). None of this fires the loop on its own; it sets the conditions under which a trigger would.

Figure 20 — The refinancing wall and the reserve runway: marketable Treasury maturities by month — bills rolling (blue) and maturing coupon securities resetting to the higher front-end rate (amber), the June–July convergence with quarter-end outlined — against the projected reserve balance (green, right axis) and the ~$3.0tn lowest-comfortable-level floor (red, dashed).

7.5 The compounding tail — AI capital expenditure and Treasury duration

The forward risks above are internal to the Treasury complex. One external risk shares its load-bearing axis and so cannot be treated as independent: the AI capital-expenditure boom. Data-centre build-out, GPU fleets and orbital compute are ultra-long-duration assets, and growth equity is a long-duration claim; the same front-end rate that prices the refinancing wall (§7.4) prices the AI complex’s financing. When the front end stays high, both re-price together.

The financing of that boom has undergone the phase change that historically marks the late stage of a capital cycle: a shift from equity to debt. Through 2023–25 the build was equity-funded — venture capital, sovereign funds, public-market enthusiasm — where a bust zeroes shareholders and stops there. Increasingly it is debt-funded — data-centre project finance, private-credit facilities, GPU-collateralised lending, hyperscaler bond issuance and off-balance-sheet vehicles — where a bust does not stop at equity holders: it defaults, cross-defaults, and recirculates into the credit base (banks, private-credit funds, repo). In the framework’s language, the equity bubble drained distress to a sink; the debt bubble recirculates it.

The scale is no longer small, and the disclosed pieces alone make the point. Hyperscalers issued more than $120bn of bonds in 2025 to fund AI data centres — Meta’s $30bn October sale its largest ever, alongside Alphabet’s $25bn and Amazon’s $15bn — and routed a further ~$120bn off their balance sheets through special-purpose vehicles (Oracle, Meta, xAI and CoreWeave); the flagship is Meta’s ~$29bn Hyperion financing with PIMCO and Blue Owl, the largest private-capital deal on record. Add SpaceX’s own ~$22bn and the visible total already passes a quarter-trillion dollars, before the unlisted private-credit lending that Morgan Stanley projects will fund more than half of global data-centre construction — an ~$800bn market — by 2028. Two features make this the dangerous kind of debt. The off-balance-sheet routing is engineered, not incidental: the vehicles exist so the borrowing does not register in the sponsors’ credit metrics — which means any published figure is a floor, not a measurement. And the credit quality is already turning — Oracle, financing OpenAI’s compute, carries a ~$111bn debt load at a credit risk its market has not priced since 2008, and CoreWeave’s ~$14bn trades near junk; the customer-concentration and GPU-collateral mechanics behind these credits are mapped in our earlier report, AI’s Chokepoints (Gao 2026). This is the equity-to-debt shift with a number on it, and the number is rising.

The period’s marquee listing is the cleanest single embodiment of the shift. The June 2026 SpaceX IPO — the largest in history, at roughly $1.75tn with a $75bn raise — is not a pure space listing but a Musk-empire roll-up: three segments, Space, Connectivity (Starlink), and AI (xAI and X, the former Twitter), consolidated by common-control mergers. It is loss-making — a $4.3bn net loss in the first quarter of 2026 alone, against a $41bn accumulated deficit — and it carries roughly $22bn of debt, including a $20bn bridge loan taken in March 2026 to repay the X and xAI term loans, GPU-financing arrangements, the legacy Twitter-buyout notes, and a revolver governed by a 3.75× leverage covenant. The record equity raise is, in part, a vehicle to carry and term out AI-and-buyout debt — equity-market saturation pushing the marginal AI dollar into the credit channel, made literal in one ticker — and the entity carries its own 2027 refinancing wall (the bridge), a corporate echo of the sovereign one, re-pricing on the same rate.

The pattern is not new. In the late-1990s telecom boom, equipment vendors (Lucent, Nortel) financed their own customers’ purchases; when the customers failed the vendors absorbed the losses, and the circular financing accelerated the collapse — a rhyme with today’s chip-and-cloud vendor financing, whose circular structure and single load-bearing edge (OpenAI→CoreWeave) our earlier report AI’s Chokepoints (Gao 2026) maps in full under the same DebtRank cascade; the difference is that SpaceX consolidates that circularity onto a single balance sheet. This report therefore does not re-derive the AI network — the AI complex enters only as a coupling: it sits on the Treasury complex’s rate axis and routes through the same private-credit and bank channels into which a Treasury dislocation would recirculate. Two taut loops sharing one node — the front end of the curve — is the precise structural meaning of the claim that an AI bust and a Treasury squeeze would compound: a shock to either raises the gain on the other. SpaceX is a tail amplifier here, not the load-bearing core of the argument; that core remains the Treasury structure of §7.1–7.4.

7.6 The break shape — stock–bond correlation and the triple-kill

The timing instrument (§7.4) reads when the load peaks; a companion reads how a US stress day resolves across assets — the shape the break would take. Two readings, again from public daily data and sharing no input with the network dials: the stock–bond correlation regime, and the triple-kill.

The first is the diversification premise itself. The rolling correlation between equity returns and bond-price returns (bonds move opposite to yield) ran reliably negative through the post-2008 decade — bonds rose when stocks fell, the hedge the standard 60/40 portfolio is built on (≈ −0.42 in 2019). Since 2021 it has not held: it turned positive through the 2022 inflation shock (≈ +0.25), swung back negative in 2023–24, and today sits at a strongly positive ≈ +0.70 — stocks and Treasuries now fall together. The hedge has not merely weakened; it has become regime-switching, and it is currently in the regime where it offers no protection at all — the diversification gone precisely when a taut structure would call on it.

What flips the sign is the inflation regime, by a mechanism the literature has pinned down. The stock–bond correlation tracks the sign of the covariance between inflation and real activity — whether the shocks moving markets are to demand or to supply (Campbell, Pflueger & Viceira 2020). When inflation is low and stable, the dominant shocks are to growth: a downturn cuts earnings and the expected path of policy rates together, so equities fall while bonds rally, and bonds hedge — the correlation negative. When inflation is high and volatile, the dominant shock is to inflation itself: a hot print forces the central bank to tighten, repricing the discount rate that values both assets at once, so stocks and bonds fall together — the correlation positive (Ilmanen 2003). The pivot in the data sits where that account predicts (Figure 21, 10-year breakeven inflation on the right axis): negative through the low-inflation 2010s, positive as inflation rose through 2021–22, and positive still while it holds above target. The break shape is not a mood — it is the inflation regime read through the discount rate, and today’s regime is the one in which the safe asset does not hedge.

The second reading is sharper, and it ties to the report’s decisive axis. A triple-kill is a day on which equities fall, yields rise (bonds fall) and the broad dollar falls — all at once. A normal US risk-off does the opposite on the last leg: the dollar is bid as the world’s haven, the exorbitant privilege in action. A triple-kill is that privilege failing in real time — capital leaving US assets wholesale rather than rotating into Treasuries and dollars. Counted strictly — equities and the broad dollar each down at least ½% with yields rising — such days are rare: about 13 in the past decade (≈0.5% of all sessions), but they cluster, four in the 2022 inflation shock and four through 2025’s tariff turmoil. On 10 April 2025 equities fell ~3.5%, yields rose and the dollar fell ~1.2% together — the second-worst such day after 2022-04-29’s −3.6% — and 2025 alone carried four. They are the observable signature of the third containment gate (§6.2) eroding: when the safe-asset bid that makes the US backstop affordable is itself what fails, the loss does not stay in Treasury prices.

This is why the break shape feeds the fork. A near-critical structure (§7.1) can resolve as an orderly flight to quality — dollar and Treasuries bid, the haven intact — or as a triple-kill, the haven gone. The first is catchable on the privilege the United States still holds; the second erodes exactly the gate the backstop rests on, and pushes the forward case toward the mode (§7.7) in which the rescue can be deployed and sized yet still fail the affordability test. On the current data both readings point the unhelpful way (Figure 21) — coupled, and clustering — though each is a descriptive cross-asset read, not a predictor, and the correlation’s instability since 2021 is itself the honest caveat.

Figure 21 — The break shape: the rolling correlation between equity returns and bond-price returns (bonds move opposite to yield), 20-day and 60-day, on the left axis — reliably negative through the post-2008 decade (the 60/40 hedge), strongly positive today at ≈ +0.70 (stocks and Treasuries falling together) — against 10-year breakeven inflation on the right axis.

7.7 The fork — two failure modes, one structure

The forward case splits into two failure modes that the structural dials cannot distinguish, because they differ only in the containment response. To make the fork concrete, we run the same cash-Treasury shock through the live calibrated network under a controlled set of interventions: no backstop; a standing purchase facility at the keystone arriving early (round 2) or late (round 6), fully monetised or fully sterilised; and full sweeps over the arrival round and the financing mix. The runs give the report’s central forward claim a computable form: the backstop cannot destroy the distress; it can only stop the compounding — and choose which axis absorbs what it intercepts. Three results carry the section (Figures 22–24).

1. Without a backstop, the loop completes (Mode A). With no effective sink, the cascade keeps the distress inside the Treasury loop and spills it downstream: the cash and basis-fund nodes saturate, and private credit (≈0.55), repo (≈0.50) and stablecoins (≈0.48) carry the overflow — structural damage ≈3.9, 44% of the nine-node network’s saturation ceiling, with 95% of the final total in place by round 9. The 1929 analogue holds not because today’s micro-structure resembles 1929’s, but because the containment configuration does: a central bank defending credibility by withholding rescue removes the loop’s only exit.

Figure 22 — The fork, computed: structural damage round by round for one cash-Treasury shock on today’s network — no backstop (Mode A, 3.9), early monetised backstop (Mode B, arrival round 2, 2.1), late arrival (round 6, 3.3).

2. An early backstop is leveraged by the same loop that makes the structure fragile; a late one is not. Arriving at round 2 and intercepting 90% of the flow reaching the keystone, the facility absorbs a cumulative flow of only 0.20 — yet final structural damage falls from 3.9 to 2.1. Each unit absorbed early prevents ≈11 units of damage, because flow intercepted at the keystone would otherwise have been re-amplified around the loop — and the multiplier is no coincidence: it is the loop gain that §7.1 reads as fragility, here working in the rescuer’s favour. The loop gain that multiplies a shock equally multiplies a rescue; that is the structural reason rescues buy the keystone. The leverage decays with every round of delay — ≈4× arriving at round 6, zero by round 10, when the cascade has saturated and there is nothing left to save (Figure 23). Across the sweep, waiting from round 1 to round 9 lowers the eventual currency-channel load by ≈0.16 at the cost of ≈2.6 of additional structural damage — roughly a sixteen-to-one exchange rate against delay. Gate 1 of §6 — timeliness — is therefore not a qualitative preference but the steepest gradient in the system. Set against §6.4’s measuring stick, the forward runs land where the record says a still-rescuable structure should: unrescued, 44% of ceiling, inside the 38–54% band of the concentrated histories (not near 2008’s 69%); with the early facility, 23%, inside the 17–24% band where every single-point rescue has held; the reduction, −47%, sits inside the historical 43–68% avoidable band, far from 2008’s 18%. Despite carrying the most dispersed feedback of the single-loop family, today’s network still answers like the concentrated cases — because the load the facility buys remains pinned to one node. What no historical row constrains is the gates: whether the facility arrives early, and what its financing does next.

Figure 23 — The cost of delay: final structural damage versus the backstop’s arrival round (fully monetised). Early intervention is ≈11×-leveraged — the same 1/(1−ρ) gain the structure reads as fragility — and the leverage decays with every round of delay, gone by round 10.

3. The financing choice decides the axis, not the amount (Mode B). What the facility absorbs is not destroyed; it is re-routed, and the monetised share m decides the destination (Figure 24). Fully monetised, the absorbed flow re-enters through the currency — dollar ≈0.11 and gold ≈0.05 of debasement-channel load, ≈0.16 in all if the backstop arrives early and ≈0.11 if it arrives late: risk migrates from “Treasury prices” to “currency debasement, inflation, confidence in fiat,” a slower second loop that price-based risk measures do not register. Fully sterilised, the currency channel stays at zero and the same flow terminates on the consolidated public balance sheet — the fiscal trajectory of §6.2. The figure measures which axis absorbs the distress, not the size of any dollar move: the load is the share of the absorbed cascade that re-enters through the dollar and gold nodes, a structural routing quantity, not a forecast of depreciation. The financing axis is gate 3 made concrete: its two ends are the two halves of the affordability question, monetary space and fiscal space, and today both are at their weakest in the calibrated record. In the model the trade-off is linear because the couplings are fixed; in reality the currency end is convex — inflation expectations de-anchor non-linearly once monetised rescue becomes the expected response. The 2020 rescue, monetised at smaller scale, was followed by the 2021–22 inflation: Mode B’s fingerprint in miniature. Gold is the natural absorber on the monetised path — consistent with reserve managers’ rotation already under way (§6.2) — though in an acute Mode-A dollar scramble gold sells off first, before the debasement bid arrives.

Figure 24 — The financing choice: the same absorbed flow (0.20), split by the monetised share m — re-entering through the dollar and gold (the debasement-channel load) or landing on the fiscal trajectory (sterilised).

We deliberately attach no probabilities to the fork. Which tail materialises is decided by the containment gates — concretely, by how the new Fed leadership weighs market function against inflation credibility when both are stressed at once, with less fiscal room than at any point in the calibrated record. The gates, however, are observable in advance.

One such input is already in play. The Fed chair installed in 2026, with a hawkish record, faces exactly the bind the fork describes: the §7.4 wall argues for lower funding costs, while inflation credibility argues against a cut that would read as fiscally motivated — Mode B’s political tell. The mid-2026 de-escalation of the US–Iran conflict bears directly on it. The war had pushed oil to about $115/bbl on a Strait-of-Hormuz risk premium; the ceasefire brought it down roughly 20% to ~$95, and year-end rate-cut odds repriced from about 14% to ~43% as the inflation impulse faded — the June preliminary deal’s reaction was a fall in the 2-year yield (the Fed-path tenor), not a safe-haven move. A supply-side disinflation gives a hawkish chair non-political cover to cut: it tilts the fork toward Mode A — a credible easing — and away from Mode B, an easing that would read as debt monetisation. The relief landed in the month the funding calendar marks as the year’s heaviest: a $2.4tn refinancing load, an administration with every interest in a lower front end and a calmer tape, and — had the Gulf war instead closed the Strait of Hormuz — an oil-and-yield spike that would have arrived in precisely that window. The de-escalation removed that tail. Two cautions keep this honest. The relief is prospective, not realised: the 2-year still sits near 4.05%, above the ~3.0% weighted-average coupon the wall is rolling into, so the cut is priced, not delivered. And it is fragile: the deal is preliminary, oil is being held in a $90–100 range pending a lasting agreement, and a re-escalation would restore the inflation impulse and the bind with it.

7.8 What to watch

Each item below moves a named dial or gate; together they are the report’s operational output. The list is not merely qualitative: the anchored reconstruction of §5.5 supplies the fixed rule by which series of this kind translate back into the structural reading — the structural dial is a quantity these series update, not a snapshot that must be recalibrated from scratch.

1. Basis-trade leverage and repo financing (Fed FSR; OFR; CFTC Traders-in-Financial-Futures, weekly; TIC Cayman line). Rising hedge-fund Treasury leverage tightens ρ toward 1 and deepens the keystone’s centrality; a disorderly unwind is the ignition mechanism of §4.10. The CFTC net-short series is the most direct public read on this dial — at a record in 2025 (§7.1) and, through the §5.5 reconstruction, the live input that moves ρ.

2. Dealer balance-sheet utilisation and order-book depth (Duffie’s utilisation measure; inter-dealer depth). The amplifier: high utilisation is the precondition for the ~10× depth collapse of §4.10.

3. The dollar–yield correlation under stress (market data; rolling 30-day co-occurrence count). Dollar down while yields rise is the privilege eroding in real time — gate 3’s international side; every recurrence weakens the safe-haven assumption embedded in reserve portfolios.

4. Gold’s share of official reserves vs Treasuries (ECB, annual). Continued rotation = the demand side of gate 3 closing further.

5. Net interest as a share of federal revenue; the term premium (CBO baseline; NY Fed ACM term-premium estimates). The fiscal side of gate 3: each percentage point of revenue absorbed by interest narrows the space for a costless rescue.

6. The Fed’s reaction function under the new chair (speeches, minutes, SLR-reform progress). Gate 1 and gate 2 capacity: language that subordinates market-functioning purchases to inflation credibility raises the probability of Mode A; pre-committed facility language lowers it but, against §6.2’s arithmetic, raises the probability that the eventual rescue is Mode B.

8. Conclusion

This report set out to answer two questions about the US Treasury market that the published research documents channel by channel but does not quantify jointly: how close the coupled system sits to self-amplification, and whether the public sector retains the capacity to catch it.

The method gives both questions numbers, and history validates the pairing. Seven calibrated US crises show that the structural dials identify real fragility — every crisis ignited from a near-saturated margin or collateral revaluation, with the core reflexive link at a near-universal gain — but that structure alone inverts the outcome ranking: the loosest network of the seven produced the deepest collapse because nobody caught it, and the only one above the line was contained by the broadest rescue in history. The cleanest controlled pair settles it: 1907 and 1929, the same single-hub structure, a private backstop at the keystone versus none, separated an 11% recession from a 26% one. Fragility is structural; outcomes are decided at the backstop.

That is what makes the 2026 reading qualitatively different from its predecessors. The structural dials sit at crisis-eve levels with the amplifier pre-loaded — that much was true in 2020 and 2022. What was not true then: the feedback is now too distributed for a narrow rescue; the keystone asset a rescue would have to buy is the very asset whose oversupply and eroding reserve status define the problem; and every containment gate — willingness, capacity, affordability — is measurably weaker than in any contained case, with the affordability gate (interest expense past the defence budget, inflation above target, gold past Treasuries in the world’s reserves) closing in plain sight. Ending a loop whose every path crosses the troubled asset itself would take a commitment to stand behind that asset — and a commitment can substitute for a purchase only while the affordability behind it is beyond question; that is precisely the premise the 2026 gates no longer clearly meet. The system has not lost its brakes; it has lost the certainty that the brakes are affordable.

For a risk officer the operational conclusion is not a forecast but a posture: stop trying to predict the unpredictable price path; run continuous surveillance on the six observables of §7.8, because they move before the loop fires; and treat the fork honestly — if the rescue comes late, the loss appears in Treasury prices; if it comes ample, the loss migrates to the currency and re-prices gold. The distress, once the structure is this taut, goes somewhere. The remaining choice — and the remaining analytical task — is watching which gate it exits through.

---

Appendix A — formal definitions

Figure A1 — The three dials and the drift gauge, in schematic form. Top left: amplification ρ is the per-round gain of the selling loop — below 1 the rounds fade (total ≈ 1/(1−ρ)); at ρ = 1 self-correction fails. Top right: two systems with the same ρ can need opposite rescues — concentrated feedback has a keystone a targeted purchase can cut; distributed feedback admits only a system-wide backstop. Bottom left: circulation asks whether distress drains to an absorber or returns to its origin amplified. Bottom right: the drift gauge reads the distance of today’s market behaviour from an objectively selected calm baseline — no structural content, the cross-check. Formal definitions in A.1–A.5.

A.1 Network and amplification dial

Let W be the weighted adjacency matrix of the directed network, with W[i,j] the marginal sensitivity of node j’s forced selling to distress at node i (calibrated per case from the anchors in Appendix B). First-order cascade dynamics are p(t+1) = W·p(t) + s, with p the vector of selling pressure and s the external shock. Stability is governed by the spectral radius

\(\rho(W) = \max_{i} \{ |\lambda_i| \}\)

for ρ < 1 cumulative pressure converges with steady-state multiplier

\(p = (I - W)^{-1}s\)

(I−W)⁻¹, scalar gain ≈ 1/(1−ρ) along the dominant mode;

at ρ ≥ 1 the recursion diverges (May 1972). The “per-round gain” language of §3.1 is the dominant-eigenmode reading of the same statement.

A.2 Loop enumeration, concentration, keystone

Following Levins (1974), the feedback structure of W is carried by its directed simple cycles. Each cycle’s gain is the product of its edge weights; the feedback levels Fₖ (signed sums of products of disjoint loops) are the characteristic-polynomial coefficients, so the loop decomposition refines the same object that determines ρ.

To quantify the concentration of feedback loops across the network, we deploy loop analysis. Let C be the set of all simple directed cycles in the network. The feedback gain g(c) of a simple cycle c in C is defined as the product of its constituent edge weights :

\(g(c) = \prod_{(i,j) \in c} W_{i,j}\)

The loop concentration of the network is defined as the normalized Herfindahl-Hirschman Index (HHI) of the absolute cycle gains :

\(\text{Concentration} = \frac{\sum_{c \in C} |g(c)|^2}{\left( \sum_{c \in C} |g(c)| \right)^2} \)

A concentration index near 1.00 indicates that systemic instability is dominated by a few critical loops. Under this topology, the lender of last resort can arrest systemic contagion efficiently by injecting targeted liquidity at a single keystone asset node, defined as

\(c_{max} = \arg\max_{j \in V} \sum_{c: j \in c} |g(c)|\)

A concentration index near 0.00 indicates that feedback is highly dispersed across a vast, complex web of loops. In this environment, targeted single-node interventions fail, and the system requires broad, system-wide guarantees and capital injections.

A.3 Circulation fraction

A saturating reverberating cascade (DebtRank family; Battiston et al. 2012) on the calibrated network yields the cumulative directed distress flux Φ between node pairs. To track the flow of systemic pressure under saturation, we utilize the cycle ratio (Φ_c). Let Φ_ij be the cumulative directed pressure flux between nodes i and j, and let F = Φ − Φᵀ be the net flow matrix.

Applying a discrete Helmholtz-Hodge decomposition to the graph, we decompose the net flow into an acyclic gradient component (representing pressure exiting the network toward stable sinks) and a divergence-free rotational component (representing pressure trapped within closed circular loops) Schnakenberg 1976; Wand, Kamps & Iyetomi 2024).:

\(F = F^{\text{grad}} \oplus F^{\text{circ}}\)

The cycle ratio Φ_c is defined as the ratio of circular flow energy to total flow energy :

\(\Phi_c = \frac{\|F^{\text{circ}}\|_F^2}{\|F\|_F^2}∈ [0,1]\)

A high cycle ratio Φ_c indicates intense internal pressure circulation and high structural friction within the network.

A.4 Drift gauge

Changes of the series under study (daily observations; monthly in the one case where no daily record exists, §5.4) are summarised by rolling-window empirical distributions: each window of 15 observations (stepped by 5) is binned into K = 10 cells whose edges are full-sample quantiles, with Laplace smoothing α = ½ so every cell stays strictly positive. A window is then a point on the multinomial statistical manifold, where the Fisher–Rao geodesic distance has the exact closed form

\(d_{FR}(p, q) = 2 \arccos \left( \sum_{i} \sqrt{p_i q_i} \right)\)

，the arc-cosine of the Bhattacharyya coefficient: the substitution xᵢ = √pᵢ maps the manifold isometrically onto the positive orthant of a sphere of radius 2, on which d_FR is the great-circle distance (bounded above by π). The gauge is this distance between the current window and a calm baseline — the contiguous stretch of roughly one trading year with the lowest mean rolling volatility in the leading part of the sample, selected by that fixed rule, never by hand. The rule is causal: “normal” is defined from earlier data, not from the period under study. A companion z-score compares each reading only with its own trailing history (no look-ahead), flagging departures that do not depend on the absolute choice of baseline. Information geometry applied to interest-rate markets goes back to Brody & Hughston (2001). The gauge is descriptive (contested tier): it measures statistical distance from quiescence, not causality.

A.5 Structural anatomy: partition, core, load

Four descriptive statistics summarise each calibrated network’s anatomy; none enters any dial computation. Partition: communities are detected on the undirected weighted projection of W by modularity-maximising agglomeration (Blondel et al. 2008) with a fixed random seed for determinism, and the partition’s quality is the modularity Q of Newman & Girvan (2004); we read Q ≳ 0.3, the conventional band, as strong community structure — below it, the network functions as one room. Self-reinforcing core: the strongly connected components of the directed network with more than one node — the maximal blocks within which every node can reach every other along directed edges, i.e. within which distress can return to its origin; we report the largest block’s membership and size. Load share: each node’s share of total incident edge weight (inbound plus outbound), the fraction of the network’s total exposure that crosses that balance sheet. Effective load breadth: the exponential of the Shannon entropy (Shannon 1948) of the load distribution — the Hill number of order one (Hill 1973) — read as the effective number of balance sheets carrying the network’s exposure: N if the load is uniform, 1 if a single node carries everything.

Appendix B — network construction and data anchors

Each case network is specified by its node set and per-edge calibration anchored to contemporaneous public data. Sources are collected in Appendix C.4–C.6.

Appendix C — References

C.1 Method — network / spectral / feedback / thermodynamic lineage

• Acemoglu, D., Ozdaglar, A. & Tahbaz-Salehi, A. (2015) ‘Systemic risk and stability in financial networks’, American Economic Review, 105(2), pp. 564–608.

• Allesina, S. & Tang, S. (2012) ‘Stability criteria for complex ecosystems’, Nature, 483, pp. 205–208.

• Battiston, S., Puliga, M., Kaushik, R., Tasca, P. & Caldarelli, G. (2012) ‘DebtRank: too central to fail? Financial networks, the FED and systemic risk’, Scientific Reports, 2, 541.

• Blondel, V.D., Guillaume, J.-L., Lambiotte, R. & Lefebvre, E. (2008) ‘Fast unfolding of communities in large networks’, Journal of Statistical Mechanics: Theory and Experiment, 2008(10), P10008.

• Brody, D.C. & Hughston, L.P. (2001) ‘Interest rates and information geometry’, Proceedings of the Royal Society of London. Series A, 457, pp. 1343–1363.

• Haldane, A.G. & May, R.M. (2011) ‘Systemic risk in banking ecosystems’, Nature, 469, pp. 351–355.

• Hill, M.O. (1973) ‘Diversity and evenness: a unifying notation and its consequences’, Ecology, 54(2), pp. 427–432.

• Levins, R. (1974) ‘The qualitative analysis of partially specified systems’, Annals of the New York Academy of Sciences, 231, pp. 123–138.

• May, R.M. (1972) ‘Will a large complex system be stable?’, Nature, 238, pp. 413–414.

• Newman, M.E.J. & Girvan, M. (2004) ‘Finding and evaluating community structure in networks’, Physical Review E, 69(2), 026113.

• Puccia, C.J. & Levins, R. (1985) Qualitative Modeling of Complex Systems: An Introduction to Loop Analysis and Time Averaging. Cambridge, MA: Harvard University Press.

• Rao, C.R. (1945) ‘Information and the accuracy attainable in the estimation of statistical parameters’, Bulletin of the Calcutta Mathematical Society, 37, pp. 81–91.

• Schnakenberg, J. (1976) ‘Network theory of microscopic and macroscopic behavior of master equation systems’, Reviews of Modern Physics, 48(4), pp. 571–585.

• Shannon, C.E. (1948) ‘A mathematical theory of communication’, Bell System Technical Journal, 27(3), pp. 379–423.

• Wand, T., Kamps, O. & Iyetomi, H. (2024) ‘Causal hierarchy in the financial market network — uncovered by the Helmholtz–Hodge–Kodaira decomposition’, Entropy, 26(10), 858.

C.2 US Treasury market structure & financial stability

• CFTC (2024) The Treasury Cash-Futures Basis Trade. Market Risk Advisory Committee. Washington, DC: Commodity Futures Trading Commission.

• Cook, L. (2025) Remarks on financial stability. Board of Governors of the Federal Reserve System, November.

• Duffie, D. (2025) Congressional testimony on Treasury market functioning. Washington, DC: U.S. Congress.

• Duffie, D., Fleming, M.J., Keane, F.M., Nelson, C., Shachar, O. & Van Tassel, P. (2023) ‘Dealer capacity and US Treasury market functionality’, Federal Reserve Bank of New York Staff Reports, no. 1070 (also Bank for International Settlements Working Papers, no. 1138).

• ECB (2024) Financial Stability Review — basis trades in US and euro-area government-bond markets. Frankfurt am Main: European Central Bank.

• Federal Reserve Board (2025) Financial Stability Report, November. Washington, DC: Board of Governors of the Federal Reserve System.

• Group of Thirty (2021) U.S. Treasury Markets: Steps Toward Increased Resilience. Washington, DC: Group of Thirty.

• IMF (2025) Global Financial Stability Report, October. Washington, DC: International Monetary Fund.

• Stein, J.C. et al. (2026) Treasury Market Dysfunction and the Role of the Central Bank. Washington, DC: Brookings Institution.

• U.S. Senate (2026) Confirmation of Kevin Warsh as Chair of the Board of Governors of the Federal Reserve System (roll-call vote 54–45; sworn in 22 May 2026; Powell remains a Governor).

C.3 Macro-finance — the dollar, the safe asset, the debt cycle

• Campbell, J.Y., Pflueger, C. & Viceira, L.M. (2020) ‘Macroeconomic drivers of bond and equity risks’, Journal of Political Economy, 128(8), pp. 3148–3185

• Dalio, R. (2025) How Countries Go Broke: The Big Cycle. New York: Avid Reader Press

• ECB (2026) Gold overtakes US Treasuries as the largest single reserve asset. Frankfurt am Main: European Central Bank

• Friedman, M. & Schwartz, A.J. (1963) A Monetary History of the United States, 1867–1960. Princeton, NJ: Princeton University Press

• Gourinchas, P.-O., Rey, H. & Govillot, N. (2017) ‘Exorbitant privilege and exorbitant duty’. Working paper (first version: IMES Discussion Paper 10-E-20, Bank of Japan, 2010).

• Ilmanen, A. (2003) ‘Stock–bond correlations’, Journal of Fixed Income, 13(2), pp. 55–66.

• Rey, H. (2015) ‘Dilemma not trilemma: the global financial cycle and monetary policy independence’, NBER Working Paper 21162. Cambridge, MA: National Bureau of Economic Research.

• Rey, H. (2025–26) Commentary on the erosion of the exorbitant privilege and the dollar in the digital age (London Business School; IMF; Project Syndicate).

C.4 Historical cases (calibration anchors; Appendix B)

• BEA — National Income and Product Accounts: US real GDP, the report’s outcome scale (1929–33 −26%, annual; 2007Q4–2009Q2 −4.3%; 2019Q4–2020Q2 −10.0%, regained by 2021Q2).

• Bruner, R.F. & Carr, S.D. (2007) The Panic of 1907: Lessons Learned from the Market’s Perfect Storm. Hoboken, NJ: Wiley.

• Brunnermeier, M.K. (2009) ‘Deciphering the liquidity and credit crunch 2007–2008’, Journal of Economic Perspectives, 23(1), pp. 77–100.

• Brunnermeier, M.K. & Pedersen, L.H. (2009) ‘Market liquidity and funding liquidity’, Review of Financial Studies, 22(6), pp. 2201–2238.

• Carlson, M. (2007) ‘A brief history of the 1987 stock market crash with a discussion of the Federal Reserve response’, Finance and Economics Discussion Series 2007-13. Washington, DC: Board of Governors of the Federal Reserve System.

• Federal Reserve (2020) Monetary Policy Report, June. Washington, DC: Board of Governors of the Federal Reserve System.

• Federal Reserve (2021–22) FEDS Notes on the March-2020 Treasury market disruption.

• Federal Reserve History (n.d.) ‘The near failure of Long-Term Capital Management’ and ‘Stock market crash of 1987’. Federal Reserve Bank of Richmond / federalreservehistory.org.

• Frydman, C., Hilt, E. & Zhou, L.Y. (2015) ‘Economic effects of runs on early “shadow banks”: trust companies and the impact of the Panic of 1907’, Journal of Political Economy, 123(4), pp. 902–940.

• Gorton, G. & Metrick, A. (2012) ‘Securitized banking and the run on repo’, Journal of Financial Economics, 104(3), pp. 425–451.

• Lowenstein, R. (2000) When Genius Failed: The Rise and Fall of Long-Term Capital Management. New York: Random House.

• Presidential Task Force on Market Mechanisms (1988) Report of the Presidential Task Force on Market Mechanisms (the Brady Report). Washington, DC: U.S. Government Printing Office.

• President’s Working Group on Financial Markets (1999) Hedge Funds, Leverage, and the Lessons of Long-Term Capital Management. Washington, DC: U.S. Department of the Treasury.

• Romer, C.D. (1989) ‘The prewar business cycle reconsidered: new estimates of gross national product, 1869–1908’, Journal of Political Economy, 97(1), pp. 1–37 — the 1907–08 contraction on the report’s GDP scale.

• Tallman, E.W. & Moen, J.R. (1990) ‘Lessons from the Panic of 1907’, Economic Review, Federal Reserve Bank of Atlanta, 75(3), pp. 2–13.

• Taylor, J.B. & Williams, J.C. (2009) ‘A black swan in the money market’, American Economic Journal: Macroeconomics, 1(1), pp. 58–83.

• U.S. Treasury (2020) Monthly Statement of the Public Debt, February and June. Washington, DC: U.S. Department of the Treasury.

C.5 2026 forward channels & macro

• BEA (2026) Personal Income and Outlays, April 2026, news release, 28 May. Washington, DC: Bureau of Economic Analysis — PCE price index +3.8% y/y (+0.4% m/m); core PCE +3.3% y/y (+0.2% m/m).

• BPI (2026) Yield-Bearing Stablecoins Can Destroy Deposits: An Illustrative Model. Washington, DC: Bank Policy Institute.

• CBO (2026) The Budget and Economic Outlook: 2026–2036, February. Washington, DC: Congressional Budget Office.

• Federal Reserve (2026) ‘Stablecoins in 2025: developments and financial stability implications’, FEDS Notes. Washington, DC: Board of Governors of the Federal Reserve System.

• Gao, Y. (2026) AI’s Chokepoints: the physical constraints, geopolitical playbooks, and financial fragility of the global AI build-out.

C.6 Public data feeds

Data feeds used for calibration, the record-test panels, and the outcome scale. The access date is the most recent retrieval.

• AllianceBernstein — estimates of US Treasury market depth relative to market size. Research commentary.- BEA — National Income and Product Accounts (US real and nominal GDP). Available at: https://www.bea.gov (Accessed: 12 June 2026).

• Board of Governors of the Federal Reserve System (1943) Banking and Monetary Statistics, 1914–1941, Tables 139 and 142 (brokers’ loans, by class of lender). Via FRASER, Federal Reserve Bank of St. Louis. Available at: https://fraser.stlouisfed.org (Accessed: 12 June 2026).

• CFTC, Commodity Futures Trading Commission — Traders in Financial Futures, weekly Commitments of Traders: leveraged-money net positions in the 10-year U.S. Treasury Note future (the basis-trade size proxy, §5.5/§7.1). Available at: https://www.cftc.gov/MarketReports/CommitmentsofTraders (Accessed: 14 June 2026).

• FRED, Federal Reserve Bank of St. Louis — series DGS10 and DGS2 (Treasury constant-maturity yields), SOFR, TEDRATE (TED spread; historical series, discontinued 2022), ABCOMP (asset-backed commercial paper outstanding, weekly), BAA10Y (Moody’s Baa − 10-year Treasury credit spread, daily), and CPI. Available at: https://fred.stlouisfed.org (Accessed: 12 June 2026).

• OFR — short-term funding monitors (repo volumes and rates). Available at: https://www.financialresearch.gov (Accessed: 12 June 2026).

• SEC — money-market fund and private-funds statistics. Available at: https://www.sec.gov (Accessed: 12 June 2026).- U.S. Treasury / CBO — debt outstanding and debt-to-GDP. Available at: https://fiscaldata.treasury.gov (Accessed: 12 June 2026).

• Williamson, S.H. — ‘Daily closing value of the Dow Jones Average, 1885 to present’, MeasuringWorth. Available at: https://www.measuringworth.com/datasets/DJA/ (Accessed: 12 June 2026).

Appendix D — Glossary of instruments

The financial instruments that carried the loop in each US episode, and those that carry it today: what each is, and the role it played. (The European instruments — LDI, gilt repo, BTPs, OMT — are glossed in Part 2.)

The Panic of 1907

• Trust-company deposits — Deposits at trust companies, which (unlike national banks) held thin cash reserves and stood outside the clearing-house safety net; the run on them (the Knickerbocker Trust) was the trigger.

• Call money (call loans) — Short-term loans to stock-market speculators, repayable on demand (“on call”) and collateralised by shares. When the trusts failed, call money was withdrawn, forcing equity liquidation — the loop’s funding leg.

1929

• Brokers’ loans (margin loans) — Credit extended to investors to buy shares on margin, secured by the shares themselves. As prices fell, margin calls forced selling, which drove prices lower still — the reflexive core of the crash; the outstanding volume of brokers’ loans is the report’s calibration series for 1929.

• Margin account — An account that allows securities to be bought with borrowed money; the leverage that turned a price decline into a forced-selling spiral.

1987

• Portfolio insurance — A dynamic-hedging strategy that sells stock-index futures automatically as prices fall, to synthesise a protective put. Mechanical, price-triggered selling that fed on itself on Black Monday.

• Stock-index futures (S&P 500 futures) — Exchange-traded contracts on the index; the venue through which portfolio insurance transmitted selling into the cash market.

• Index arbitrage / program trading — Computer-driven trades exploiting the gap between futures and the cash index; the channel that propagated the futures sell-off into cash equities.

LTCM (1998)

• Convergence / relative-value trade — A bet that two closely related prices will converge, earning the spread; LTCM ran these at extreme leverage, so a divergence forced unwinding.

• Interest-rate swap; swap spread — A contract exchanging fixed for floating interest; the swap spread (versus Treasuries) was a core LTCM convergence position.

• Repo leverage — Borrowing against securities through repurchase agreements to multiply position size; the mechanism that let LTCM’s small capital control vast exposures, and the channel through which its forced unwind hit the market.

2008

• Mortgage-backed security (MBS) — A bond backed by a pool of mortgages; the asset whose repricing started the crisis.

• Collateralised debt obligation (CDO) — A structured security repackaging MBS (and other debt) into tranches; the instrument that spread mortgage losses opaquely across the system.

• Credit-default swap (CDS) — Insurance against a bond’s default; the interlocking exposures (notably AIG’s) that turned losses into systemic counterparty risk.

• Asset-backed commercial paper (ABCP) — Short-term debt backed by assets, used by shadow banks to fund long-term holdings; the run on ABCP is the report’s calibration series for the 2008 funding leg.

• Repo / haircut — Short-term secured funding; rising haircuts (the collateral discount) forced deleveraging across the dealer system.

COVID (March 2020)

• Cash-futures basis trade — A leveraged arbitrage that is long cash Treasuries and short Treasury futures, capturing the small price gap between them; when margins rose, forced unwinding of the cash leg seized the Treasury market. The report’s central reflexive loop, then and today.

• Repo; dealer balance sheets — The funding for the basis trade and the intermediation capacity of the market; both seized, requiring the Fed to buy Treasuries directly.

SVB (2023)

• Held-to-maturity / available-for-sale (HTM / AFS) accounting — The split that lets a bank carry long bonds at cost (HTM) until a sale crystallises the loss (AFS); the accounting trigger that turned a paper loss into a solvency question.

• Duration risk — The sensitivity of a bond’s price to rate changes; the long-duration securities book that the 2022–23 rate rise marked down below the bank’s equity.

• Uninsured deposits — Deposits above the FDIC insurance limit, which flee fastest; a ~94%-uninsured, concentrated base that ran in a single day.

• Bank Term Funding Program (BTFP) — A Fed facility that lent against eligible securities at par (face value) rather than market value — a backstop that neutralised the duration-loss trigger directly, and a new point on the containment spectrum.

Today (2026)

• Treasury bills (T-bills) — Short-dated government debt; one-third of the marketable stock matures within a year — the refinancing wall of §7.4.

• SOFR repo — Overnight secured funding benchmarked to the Secured Overnight Financing Rate; the plumbing whose stress shows up as reserve scarcity.

• Basis trade (current) — The same cash-futures arbitrage as COVID, now at a record size (CFTC leveraged-fund net shorts in Treasury futures) — today’s primary reflexive loop.

• Stablecoin; tokenised Treasury (RWA) — A crypto token pegged to the dollar and backed largely by T-bills, and a Treasury holding issued on a blockchain (a “real-world asset”); the new digital channels that couple crypto stress to the Treasury market (§7.3).

• Term premium — The extra yield investors demand for holding long-dated debt; its re-emergence raises the cost of rolling the wall.

• Dealer balance-sheet capacity (SLR) — The Supplementary Leverage Ratio constrains how many Treasuries dealers can hold; the binding limit on the market’s shock-absorbing capacity.

• GPU-collateralised lending; special-purpose vehicle (SPV); private credit — Debt secured by AI chips, off-balance-sheet financing structures, and lending by non-bank funds — the channels through which the AI-capex boom became debt-funded (§7.5), routing onto the same rate axis as the Treasury wall.