BL-SD-022: Cascade Exhaustion Fade, Multi-Confirmed

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-07

Venue: Hyperliquid perpetuals Programme: I Verdict: FALSIFIED Axis of death: Idea Transferability: UNIVERSAL

1. Identification

S1-MC was designated the flagship of the multi-confirmation programme before a single event was measured, on the grounds that its primary signal was the only non-maker band compatible with the firm's reaction loop and that every piece of gate data it required was already flowing. It is included in this record under its original name and its original status: the flagship died at the first gate.

2. Economic thesis

Every position on Hyperliquid carries a publicly readable liquidation price, because margin and leverage per address are venue state rather than private information. Aggregating that state across all open positions in a coin produces a forced-flow density function: the notional that must be sold or bought, without regard to price, if the mark reaches a given level. When mark crosses a mapped cluster, a cascade of forced liquidations runs until the mapped notional is exhausted. The thesis is that displacement remaining after exhaustion is a dislocation with no forced seller or buyer left behind it, and that price should revert toward the venue's oracle-anchored fair value once the forced flow that caused the displacement has run out.

Who is forced to lose. Two parties, named explicitly at design time: the liquidated wallets, whose flow is forced and price-insensitive by construction, and a late chaser who enters in the cascade's direction after the forced flow has already exhausted, buying or selling into the peak of a dislocation that is about to correct. Both losses are structural. Liquidation is a venue mechanism no counterparty can decline, and there is always a fresh supply of over-leveraged positions to feed it.

The multi-confirmation doctrine layered four gates onto this primary signal, each drawn from a source the doctrine holds to be causally orthogonal to the others: aggressor identity, oracle anchor, book supply texture, and HLP inventory absorption. The doctrine's own rule states plainly that gates confirm a primary that already works; they do not resuscitate one that does not.

3. Falsifiable claim

Pre-registered in two stages. The event and frequency test, pre-registered 2026-07-09:

The cascade-exhaustion primary fires at a rate sufficient to reach a pre-registered sample of at least 100 events within the programme's runway.

The edge test, D-018, pre-registered before any reversion number was seen:

Post-exhaustion price reverts toward the oracle, net of the round-trip cost, and beats a dislocation-matched random-entry null. If the primary fails this test, the multi-confirmation gates are not applied; the doctrine's own rule forbids treating a gate ladder as a rescue for a primary that does not clear the null on its own.

4. Mathematical specification

Notation per BL-D-011.
 
Cluster mapping, at onset t0 of a liquidation episode:
 
  cluster_size(t0) = sum of |szi_j| * mark0
    over positions j with liquidation_px in [mark0, mark0 + d * beta * ATR]
    and correct side for direction d
 
  d           +1 for a short-liquidation episode (price rising),
              -1 for a long-liquidation episode (price falling)
  ATR         trailing 60-minute true range median, from 1-minute candles
  beta        pre-registered grid {0.5, 1.0, 2.0}, fit-window only
 
Realised forced flow within the episode, from forced-side aggressive trades:
 
  cum_liq_flow(t) = sum over forced-side trades i in [t0, t] of  px_i * sz_i
    forced side = 'A' (sell) for d = -1,  'B' (buy) for d = +1
 
Exhaustion event at t_ex, the earliest t at which:
 
  cum_liq_flow(t)  >=  alpha * cluster_size(t0)          AND
  remaining forced density in [t, t + beta*ATR] forward  <  (1 - alpha) * cluster_size(t0)
 
  alpha       pre-registered grid {0.6, 0.75, 0.9}, fit-window only
 
Signal at t_ex:
 
  fade_dir  = -sign( (mark(t_ex) - oracle(t_ex)) / oracle(t_ex) )
 
Entry:
 
  fire at t_ex in direction fade_dir
 
Exit:
 
  at t_ex + H*, H* = 120 s, fixed. Secondary horizons {30, 60, 300} s reported
  as robustness only, never as a decision.
 
Sizing:
 
  equal notional per event
 
Cost:
 
  c_rt = 6.0 bps, maker-taker round trip at venue tier 0
 
Net return per event:
 
  reversion_ret(t) = fade_dir * (price_{t_ex+H} - price_{t_ex}) / price_{t_ex} * 1e4
  net_ret(t)        = reversion_ret(t) - c_rt

Causality holds by construction: cluster_size is computed from position state as of t0 only; cum_liq_flow accumulates only trades at or before the evaluation time; the exhaustion condition looks forward only to confirm the absence of continuing forced flow, which is a post-hoc characterisation of the episode rather than an input to the entry decision at t_ex, since t_ex itself is defined as the earliest time the accumulated condition is met.

Two departures from the design document are recorded here rather than left implicit. First, cum_liq_flow was originally specified from position disappearance in the polled snapshot stream; that measure undercounted realised flow by the scanner's coverage gap and returned near-zero events across every grid tested. The amendment moved to forced-side trade notional, which has full coverage but is a measured upper bound on true forced flow, since it can include voluntary same-side aggression. Second, the identity gate G1 was never applied to this measurement, for the reason given in section 6.

5. Operating envelope

Eleven of the fourteen coins in the scanning universe produced too few liquidations to contribute a single exhaustion event at any grid point. The effective universe for this measurement is the three majors, and that concentration is reported here rather than smoothed over, because a reader substituting a smaller or newer venue should expect this primary to be even more event-starved.

6. Instrument

The cascade-event detector identifies the exhaustion event and computes cluster_size and cum_liq_flow; the reversion-edge evaluator computes reversion_ret and net_ret and applies the moving block bootstrap, BL-IS-001, to build the dislocation-matched null p-value. Both are pure functions, test-driven, and reused without modification across the primary and high-power grids.

The identity gate, G1, was pre-registered in the multi-confirmation doctrine as "residual aggressive flow not dominated by the winner cohort." Its prerequisite, a persistent winner cohort, was tested directly by BL-SD-023 and killed: rank correlation of realised skill across non-overlapping windows reversed sign rather than persisting. G1 as specified could not be constructed, and the orthogonality instrument, BL-IS-013, was run on the three gates that remained computable: anchor distance (G2), book supply texture (G3), and HLP absorption (G4). Co-fire at random times found G3 and G4 not redundant (phi = +0.261 on 1,177 joint firings, below the 0.30 discard threshold), while G2 and the primary itself fired too rarely in the random-time sample to adjudicate (fewer than 5 joint firings). None of this determined the verdict below: the primary was tested alone, because the doctrine's own rule forbids applying a gate ladder before the primary clears the null on its own merits.

Both samples are below the 100-event target set at pre-registration. The verdict below is reported as preliminary for that reason, and the review recorded in section 8 accepted the preliminary result rather than waiting for the full sample, on the grounds that two independent grids agreed in sign and in failing both kill criteria.

7. Results

Both intervals lie entirely below zero, and neither event mean approaches its null's 95th percentile; the p-values state that the primary underperforms a generic dislocation-matched fade on the large majority of resamples rather than merely failing to beat it.

Recovery fraction, the share of the initial dislocation that closed by the horizon, tells the same story from a different angle:

A positive recovery fraction would mean the dislocation closed toward the oracle, which is the mechanism the thesis names. A negative one means price moved further from the oracle after the primary fired. Both grids show the latter.

Robustness across the secondary horizons moves in the same direction and does not reverse at any point tested:

The primary's net return does not merely fail to turn positive at a longer horizon; it worsens.

8. Verdict, axis of death, and transferability

Verdict: FALSIFIED. Axis: idea. Transferability: UNIVERSAL.

Both pre-registered kill criteria fired on both grids tested. The confidence interval on net return lies entirely below zero, and the event mean fails to beat a matched null by a wide margin rather than a narrow one. The recovery fraction is negative rather than merely small, which rules out the softer reading that reversion exists but is too slow to monetise: price does not revert after cascade exhaustion on this substrate, it continues.

This is a mechanism failure, not a cost failure or a cadence failure. At zero cost the primary would still show a negative net return, since the failure is in the direction of the price move rather than in the margin available to capture it. No envelope substitution rescues a mechanism that moves against its own thesis.

The gate ladder that gave this hypothesis its name was never applied, and that absence is itself the finding this dossier exists to preserve. The multi-confirmation doctrine states explicitly that gates confirm a working primary and cannot resuscitate a failing one. S1-MC is the doctrine's own worked example of that rule, tested against its flagship candidate on the first measurement it produced.

9. Cost and capacity analysis

Net return, primary grid, H*=120s        -5.97 bps
Net return, high-power grid, H*=120s     -8.30 bps

No gross figure separate from net is reported in the source measurement; the cost model applies a constant 6.0 basis point round trip regardless of holding period, so the negative trend across horizons in section 7 reflects the direction of price movement rather than an escalating cost. Capacity was not measured. Event frequency on the effective three-coin universe was approximately 1.5 to 6 events per day depending on grid, which was adequate for sample accumulation but was never reached in a form requiring a capacity study, because the mechanism failed before frequency became the binding question.

10. Envelope sensitivity and reopening conditions

No binding parameter exists to name, because the finding is a sign reversal, not a magnitude shortfall.

Reversal condition:  none computable.

A cost substitution cannot reverse this result. The measured net return is negative because the underlying price path moves away from the oracle after exhaustion, and no round-trip cost, however small, converts a negative gross move into a profitable trade. Setting c_rt to zero and adding back the 6.0 basis point cost to both reported net figures still leaves the primary grid at a level far short of the mechanism the thesis requires, and the recovery fraction, which carries no cost term at all, remains negative on both grids.

This is stated plainly because it is the harder and more useful finding for an outside reader. BL-D-002 section 4 asks whether the result would be positive under arbitrarily favourable envelope parameters. It would not: the mechanism itself points the wrong way. That is what places this hypothesis in the UNIVERSAL rather than the ENVELOPE-BOUND class, and it is why the gate ladder, however carefully specified, could not have saved it. A gate narrows a population; it does not change the sign of the underlying move.

What would reopen it: a demonstration, on data this programme has not seen, that the event definition itself was miscalibrated in a way that produces a systematically early or late t_ex, such that the true exhaustion point lies at a horizon this measurement did not observe. That is a testable claim and it was not tested here, because doing so on the same data the verdict was drawn from would be horizon shopping.

What would not reopen it: applying the gate ladder to the existing sample. The primary failed the null on its own; gates filter a population and cannot change the direction of a move that already contradicts the thesis.

11. Reference implementation

The cascade-event detector is published as 05_INSTRUMENTS/strategies/python/src/cascade_events.py with 27 tests. The reversion-edge primitives it feeds are published as part of 05_INSTRUMENTS/strategies/python/src/session_structure.py, which is where BL-SD-035 reuses them. BL-IS-013, the orthogonality instrument referenced in section 6, remains specified rather than published, under the policy in BL-IS-000 section 8. All are pure-function libraries with unit coverage predating the data run.

Two design choices carry the detector's honesty and both are pinned.

The trailing volatility measure is a median of true ranges, not a mean. A cascade is by definition a period in which true range explodes, and this statistic scales the band that decides which events exist at all. Built as a mean over a window containing the cascade it measures against, it would grow with the event, widen the band, admit more forced flow into the cluster, and make the event look larger than it was. One test constructs that window and shows the mean moving by an order of magnitude while the median does not move.

The exhaustion test requires both halves: most of the mapped fuel burned, and little left in front of price. Either alone fires on the wrong thing. Cumulative flow alone fires in the middle of a large move, where more fuel lies ahead than behind. Remaining density alone fires on a symbol that never had much forced flow mapped, where the emptiness ahead means nothing happened rather than that something finished.

The cartography looks only ahead of price, and that is what makes the cluster a measure of fuel rather than of open interest. A falling price liquidates longs below it; a rising price liquidates shorts above it. Positions on the other side cannot be triggered by the move in progress.

12. Evidence register

This measurement is recorded in narrative verdict reports rather than as an emitted machine-readable artifact and is registered as D-018 in BL-A-006.

13. Lessons extracted

This hypothesis is the clearest demonstration in the archive of BL-D-007 section 5, the marginal-contribution ladder rule: a gate ladder is only ever applied to a primary that has already cleared its own null. S1-MC was designated the flagship of the entire multi-confirmation programme before a single event was measured, and it died at the first gate the doctrine itself imposes, before any of the four confirmation gates it was named for were touched.

The secondary lesson concerns instrument correctness under time pressure. The first measure of realised forced flow, drawn from position disappearance in a polled snapshot, returned near-zero events on every grid and would have produced a frequency-based kill if accepted at face value. The correction to trade-based flow measurement, made before any PnL number was seen, is what allowed the mechanism itself to be tested rather than an artifact of scanner coverage. BL-D-006 on data integrity exists for exactly this failure mode.

14. Cross-references

BL-SD-023 for the winner-cohort persistence test whose failure closed gate G1 before this measurement began. BL-SD-028 for the momentum-continuation follow-up this falsification motivated directly. BL-D-007 sections 5 and 8 for the marginal-contribution ladder and its limits. BL-D-005 for the axis taxonomy applied in section 8. Ledger row S1-MC.

Copyright (c) 2026 Bijective Labs. All rights reserved.