Hyperliquid's own fill log puts the structural offspring multiplier no higher than 0.195 during the largest liquidation event on record. Both factors are visible directly in the log. The figure to remember is 0.195, measured on the one venue where every forced fill can be watched as it lands.
Building the multiplier from fills
The measure is a branching ratio in the Galton-Watson sense. Forced selling with notional V moves price by dp/p = kV, where k is impact per dollar of net aggressor flow, measured in inverse dollars. A relative move of size delta crosses the density of liquidation thresholds rho(p) across a band of width p times delta. That move forces another rho(p) p delta of notional onto the book.
Write rho-tilde for rho(p)p, the forced notional per unit relative move. The dimensionless offspring mean is lambda = k rho-tilde, with no free constant. The cascade condition has lambda going to one. An initial shock V0 is amplified to V0 divided by (1 minus lambda).
Most venues hide both factors, so the reflexivity literature usually estimates lambda from aggregate event counts. Hyperliquid publishes every open position and forced fill on chain. Its per-minute archive also reports impact prices directly, removing the need to estimate k by regression. Garcia Seuma takes both factors from that log. Since the record begins on 2025-05-25, the October 2025 crash, which began at 2025-10-10 20:50 UTC, is the only event for which this measurement is possible at all.
Across the event, the structural estimate moves from 0.031 in the Oct-9 baseline to 0.026 pre-onset, then 0.097 late pre-onset. It reaches 0.195 at nucleation, falls to 0.140 at peak, then records 0.032 late and 0.050 in the aftermath. Two other constructions also indicate subcriticality.
The amplification calculation sets $644M of forced notional in the nucleation window against $733M for the full cascade during the next 15.7 hours. That ratio produces A = 1.14 and an implied lambda of 0.122. A Hawkes-style flow estimator based on one-minute forced-sell counts moves the other way, falling from roughly 0.566 in calm markets to 0.283 through nucleation.
The paper states plainly that this third estimator has a mechanically inflated level. In individual six-hour calm windows, it rises above unity, reaching 1.63 coarse and 2.05 notional-weighted. Its trajectory is the usable result. The paper also warns that the three estimators lack statistical independence because they share the same event, fill log and price path. The abstract is even more direct: all three support subcriticality within the venue, "not on a common numerical level." Any individual number, including the 0.195 in my own headline, therefore comes from one estimator on one venue. That disclosure is appropriate. I give more weight to the structural and bookkeeping estimates, using the flow series only for direction.
The book saw a minority of the selling
Forced selling was heavily front-loaded. Of all post-onset volume, 87.8% arrived during the first thirty minutes and 96.5% within the hour. The $733M total crossed fifty-eight price buckets, each 0.25% wide. Across the broader Oct-9 to Oct-11 window, the same population moved $0.96B through eighty buckets between $100.7k and $122.7k. Such a delivery schedule fits generational propagation poorly.
The order book never received 62.6% of post-onset forced-sell notional. Hyperliquid's backstop vault took that flow off-book. During the worst minute, 21:19 UTC, forced selling reached $641M. The vault absorbed $576M, while $64M reached the book.
Hyperliquid's liquidation waterfall therefore dampens the feedback represented by the branching model, with its strongest intervention at the climax. Garcia Seuma makes the limitation explicit: the backstop "actively suppresses the very quantity we measure on the one venue where we can measure it." From there he derives a testable design prediction. Realized lambda should run hotter on venues that lack such a vault. The contrast case he gives is Binance's $283M reimbursement during the same event.
We could not test this mechanism with our own data. The required objects are perpetual-futures liquidation fills carrying liquidated-user attribution, quoted impact prices and venue backstop absorption. Our holdings are crypto spot bars. Any substitution would remove the mechanism under study instead of approximating it. An honest replication is therefore unavailable, and the paper itself is a diagnostic study without a strategy or cost accounting.
Seven events, one discontinuous pattern
The seven-event analysis tests the criticality interpretation through the correlation fabric rather than through a single price. It uses Binance USDT perpetuals sampled at five-minute resolution. The panel expands from N=22 in May 2022 to N=30 by 2024-2025.
At onset, mean pairwise coupling rises by +1.6 to +4.4 baseline standard deviations in six of seven events. The resulting phase has coupling near 0.85 to 0.90, with leading-eigenvalue weight around 0.9. May 2022 is the exception. It was a multi-day decline at minus 0.08 whose fabric had already become ordered. The susceptibility proxy, defined as N times the variance of couplings, drops in five of seven events. It reaches minus 3.4 sd in December 2024 and diverges in none.
Six of seven jump. Five of seven collapse.
Subsampling from N=8 through N=28 shows extensive susceptibility: the December baseline increases from 0.18 to 0.77. The onset jump remains N-invariant. The paper uses "first-order" only as a signature-based classification, without an N to infinity limit or free energy.
One result surprised me. October 2025 was the most violent event under every liquidity measure, yet its fabric de-correlates into onset, with tau = minus 0.50 and p = 0.202. February 2025 produced no single-variable early warning in the author's Part I, but it has the strongest build-up: tau = +0.83 and p = 0.006. Event severity does not reveal the order of the transition.
A brief attribution comparison helps. Our earlier note on a correlation-fabric stress index found that its attribution layer reduced to a row sum (link). Garcia Seuma's order parameter openly measures the weight of the leading market mode. The quantity is measured rather than repackaged.
The branching model misses both pre-state predictions. Relative to placebo onsets matched by trailing three-day return, only four of seven events exceed the matched median, with Fisher p = 0.062. Across the placebos, lambda follows the trailing price path itself, showing corr with the three-day return of minus 0.38. February 2025 is the lone exception. It reaches matched percentile 1.00 and z = +4.25, which the paper treats as a case finding rather than a general rule.
The zero-parameter unit-slope severity regression is rejected. The price-conditioned estimate is minus 0.244 with interval [minus 0.767, +0.279], n=157 and simulated power 0.958. The other estimate is minus 0.152, with n=158 and power 0.970. Directly measured Hyperliquid impact does not save the prediction. The corresponding estimates are +0.015, with n=115 and power 0.995, and +0.376 with interval [minus 0.175, +0.928], n=188 and power 1.000.
Impact and threshold density offset each other. corr(log k, log rho) is minus 0.72, while rho adds zero to four decimal places of incremental R-squared over k. The leverage-times-fragility product carries no information about severity because its two factors move in opposite directions.
How far does 0.195 travel?
Garcia Seuma identifies venue scope as the central boundary on lambda. He writes that subcriticality within a venue can coexist with system-level amplification. The discussion concedes this, the limitations repeat it, and the closing sentence asks whether subcriticality survives the cross-venue loop.
The headline figure still tends to travel without its venue label. Yet the system-level loop operates through a price shared across venues, and the author has separately documented that mirrored coupling. In his own words, the network version of the branching measurement remains future work. The fill log establishes a narrower result: Hyperliquid's engine, including its vault, measured 0.195 at nucleation and 0.140 at peak against a critical boundary of 1. The clean range of 0.122 to 0.195 describes one event on a particular venue with a particular mechanism design. Treating it as a property of leveraged crypto markets goes beyond the evidence.
The claim that no pre-state measure grades severity needs the same boundary. The abstract says "none of the scalar pre-state measures we can construct." The paper identifies a candidate beyond the reach of its designs: the accumulated map of liquidation thresholds along the path. Its discussion describes this map as a measure swept by the path rather than a scalar, precisely the kind of state variable excluded from the scalar severity tests. The negative finding concerns scalars. The candidate those designs cannot exclude remains untested.
Readers of the quotable line, "leverage is the fuel, not the fire alarm, and the fire is in the book," should treat its second half as a hypothesis. The paper leaves two objects untested: the cross-venue backstop prediction and the threshold map.
The findings that hold up best for monitoring occur inside the cascade. Binance's Kyle regression finds six-hour mean impact elevated by a factor of 1.2 to 3.5 across all seven cascades. Hyperliquid reports impact rather than estimating it by regression. There, the cascade impact spread rises across all six major instruments by a mean factor of 3.2 to 9.1, and by 511 to 927 at the cascade-day maximum. The archive covers five of the events. Hyperliquid open interest clears to 0.30 to 0.55 of baseline, compared with minus 24.6% for Binance BTC. Those measures behaved consistently.
One appendix stands on its own. Binance metrics dumps timestamp the end of an interval, while klines mark its start. The taker ratio has correlation 0.996 at exactly plus five minutes and no more than 0.11 at every other lag. Naive pairing creates a significant cascade-week "dark causality" increase at p below ten to the minus four. Realigning the timestamps makes it disappear. Anyone running mixed-source high-frequency crypto pipelines should audit that alignment before making another spectral claim.
A network version of the measurement would change my view of the venue-boundary objection: apply the same fill-log construction to a second venue without a backstop vault over the same minutes. If lambda also levels off near 0.2, the subcritical interpretation generalizes. If it runs hot, the 0.195 was engineering.