BL-SD-036: Vault Inventory Normalisation

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-08

Venue: Hyperliquid perpetuals Programme: I Verdict: OPEN Blocking axis: Sample Transferability: -

1. Identification

This dossier records a hypothesis that was designed, wired, and then never characterised, rather than one that was measured and found wanting. It belongs in the record for the same reason BL-SD-020 does: an unresolved line of research with a stated cost and a stated reopening condition is a finding in its own right, and hiding it would let a reader assume the question was either answered or never asked, when in fact it was asked carefully and then starved of the time it needed.

2. Economic thesis

The HLP vault is Hyperliquid's mechanical counterparty: it backstops liquidations and provides passive liquidity, and its aggregate position is public. Its inventory is therefore a mirror of two things at once, the crowd's net positioning and its own recent absorption of forced flow. Two distinct hypotheses follow from that mirror, and the design deliberately separated them.

The first, the strategy hypothesis: after the vault absorbs a liquidation cascade, its inventory sits away from whatever level it normally runs at, and if that displacement unwinds in a predictable direction over a predictable horizon, a participant could ride that normalisation rather than trade against it.

The second, the gate hypothesis, survives even if the first does not: HLP's skew is a measure of how crowded the market is on one side, independent of whether the vault's own unwind is itself tradeable. A crowding gate of that kind could sharpen other hypotheses in this programme's Hyperliquid family without needing to be profitable on its own.

Who would be structurally forced to lose, if the first hypothesis held. The participants whose liquidations the vault absorbed at a stressed price. They have already lost by the time their position becomes vault inventory; the thesis is not about trading against them, it is about anticipating the vault's own mechanical response to having absorbed them. The design document is explicit that this reasoning had not been verified against data at the time it was written, and states so as a caveat rather than a claim: this was the most speculative design in the family, and it was built specifically as a low-cost characterisation exercise for that reason, not as a pre-approved strategy.

3. Falsifiable claim

Pre-registered as a characterisation gate, not yet run against sufficient data:

With at least 30 days of recorded HLP inventory, no predictive structure exists if lead-lag or Granger causality from inventory delta to the next period's coin return is indistinguishable from zero, and if inventory unwind shows no autocorrelation. Absent that structure, S5 stops as a standalone strategy; its skew measure is retained only as a candidate gate for other hypotheses in the family if a separate informativeness test passes, and discarded entirely otherwise.

None of these thresholds were ever tested against data. The claim is reproduced here in the form it was written in, before any measurement, and is reported as untested rather than as failed.

4. Mathematical specification

This is a specification of the intended measurement, not of a strategy that was ever run, and it is stated as such rather than reconstructed after the fact. No signal fired, no position was sized, and no cost was charged, because the characterisation gate that would have preceded any of those terms was never reached.

Notation per BL-D-011.
 
Intended substrate, per poll interval (design cadence: 1 per minute):
 
  inv_t        =  signed inventory of the HLP vault, aggregated across its
                   seven child accounts, per coin, read from clearinghouseState
 
  d_inv_t      =  inv_t - inv_{t-1}                       inventory delta
 
Intended structural tests, none of which were run against sufficient data:
 
  (i)   autocorr( d_inv_t, lag )               for a range of lags,
                                                 unwind gradualness/predictability
 
  (ii)  lead_lag( d_inv_t -> r_{coin, t+h} )    Granger-style, horizons
                                                 h in {5m, 30m, 1h, 4h}
 
  (iii) d_inv_t | post-cascade(coin, t)          conditional normalisation
                                                  behaviour, gated on an S1
                                                  cascade event having just
                                                  occurred on the same coin

Causality would have held by construction had this run: inv_t is a point-in-time state read, not reconstructable retroactively, so every day of delay was a day of sample permanently lost rather than deferred. That property is precisely why the design document flagged accumulation as urgent even while treating the strategy claim itself as unverified, and it is why the gap between what was required and what was recorded, stated in section 6, is the entire finding this dossier has to report.

5. Operating envelope

6. Instrument

No dedicated instrument was built for this hypothesis. The design specifies the reads it would require, clearinghouseState polls against the confirmed HLP parent address, folded into the existing wallet-polling infrastructure under wiring item W4, but the structural tests in section 4, autocorrelation, lead-lag, and the post-cascade conditional, were never implemented because the substrate they would run against never reached a usable depth.

The gap between required and recorded substrate is not a measurement result. It is the reason no measurement was possible, and the two are kept distinct throughout this dossier.

7. Results

There is no results table. No structural test in section 4 was run against data, because none reached the pre-registered 30-day floor. The only number this dossier has to report is a substrate depth against a substrate requirement, and that comparison is given in section 6 rather than manufactured into a results table it does not belong in.

8. Verdict, blocking axis, and transferability

Verdict: OPEN. Axis: sample. Transferability: not applicable.

This is neither a kill nor a validation, and treating it as either would misrepresent what happened. The hypothesis was designed with a specific, dated characterisation gate; the substrate that gate required a minimum of 30 days to accumulate; the research runway was reallocated to a single priority hypothesis, BL-SD-020's Hyperliquid sibling, on 2026-07-05, roughly two days into that accumulation. The decision to reallocate is recorded in the runway document as deliberate rather than accidental: with research runway constrained to under a month, the governing decision named S5 as strategy work to cut, explicitly, while leaving the recording wiring in place.

The axis of death is sample rather than idea, cost, or data, because nothing about the mechanism was tested and found wanting. Nothing in this dossier supports a claim that vault inventory does or does not normalise predictably; the honest statement is that the question was never put to data at the power its own pre-registration required. No transferability classification applies, because nothing closed.

9. Cost and capacity analysis

Not applicable. No trade was modelled, and no capacity question arises from a characterisation gate that was never evaluated.

10. Envelope sensitivity and reopening conditions

Binding parameter: substrate depth.

Reversal condition:  days of recorded HLP inventory  >=  30

The gap is stated exactly as it stood at the point research runway moved elsewhere: approximately 2 days recorded against a 30-day requirement, a shortfall of roughly 28 days at recording cadence. Because the vault's inventory is a point-in-time state that cannot be reconstructed retroactively, closing that gap requires resuming the poll and waiting; it cannot be closed by reprocessing existing data, unlike a characterisation that depends on trade or book history the shadow lake already holds.

What would reopen it: resuming the W4 poll, already wired, for a contiguous 30-day window, followed by running the three structural tests named in section 4 exactly as specified: inventory-delta autocorrelation, lead-lag or Granger causality against next-period coin return, and the post-cascade conditional normalisation test. Any one of the three producing a result distinguishable from zero would move this from OPEN toward a settled verdict on the strategy hypothesis. All three returning null would close the strategy hypothesis cleanly while leaving the gate hypothesis, HLP skew as a crowding feature for other hypotheses in the family, to be evaluated separately, exactly as the original design specified.

What would not reopen it: running the structural tests against the roughly 2 days actually recorded. The design's own floor was set at 30 days specifically because inventory dynamics operate on a slower clock than tick-level order flow, and evaluating the gate below that floor would not produce a defensible answer in either direction; it would produce a guess with a p-value attached to it.

11. Reference implementation

None. No analytical instrument was built beyond the W4 polling wiring itself, which is infrastructure shared with other hypotheses in the family and is not specific to this one.

12. Evidence register

No evidence artifact records a structural test result, because none was run.

13. Lessons extracted

A characterisation gate with a stated minimum sample is a commitment, not a suggestion, and a runway decision that reallocates effort before that minimum is reached should say so plainly rather than let the hypothesis quietly disappear from the record. The design document and the runway decision that cut it both did that here: the cut is dated, attributed to a named constraint, capital-limited research time, and distinguished explicitly from a finding that the mechanism failed. That is the discipline this dossier is written to preserve. A programme that lets an unstarted characterisation gate blend into the archive alongside its falsified and closed hypotheses teaches its readers the wrong lesson about what the absence of a result means.

14. Cross-references

BL-SD-020 for the sibling hypothesis whose prioritisation is the direct cause of this dossier's unresolved state. BL-D-004 section 5 on verdict vocabulary, specifically the distinction between an unstarted or underpowered characterisation and a settled verdict in either direction. BL-SD-000 section 6, which records this row alongside the programme's other open leads. Ledger row S5, status OPEN.

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