BL-P1-004: The Hyperliquid Track
BIJECTIVE LABS
BL-P1-004: The Hyperliquid Track
Status: RELEASED
Classification: PUBLIC
Revision: 1
Date: 2026-09-07
Period: 2026-07-02 to 2026-07-18 Hypotheses resolved: eleven Outcome: three exit mechanisms closed in five days; one fact confirmed and unmonetisable
1. Why a second venue
Not because the first one was exhausted. Because of a structural asymmetry that no amount of research effort was going to change.
On a centralised venue at this capital scale the programme faced three disadvantages it could not close: a fee tier requiring 5,000,000 USD of monthly volume, participants executing from colocated positions, and no visibility into queue position at all. Every passive strategy the maker era had tested lost to at least one of them.
A venue with deterministic on-chain ordering and public position data changes the question. Queue mechanics are observable. More importantly, so is who holds what and at what liquidation price, which makes a class of hypothesis testable that simply cannot be posed on an anonymous venue.
The trade is depth against richness. The centralised venue offers years of history and nothing about participants. This one offered identity down to the wallet over roughly three weeks. Neither substrate answers the other's questions, and the record is explicit about which one carries each finding.
2. What was built first
Five strategy designs and an overlay, then a verification runbook before any of them ran.
The runbook is the notable part. Its first stage answered six factual questions
about the venue by pulling live data rather than reading documentation, and its
output is published as evidence artifacts E-0055 through E-0057. Nothing was
built on an assumption about the venue that had not been checked against the
venue.
The read-only substrate that followed was a shadow recorder with a zero-order guarantee, writing a partitioned data lake. It placed no orders at any point in the programme.
3. Two premises tested before any strategy
Skill persistence. The hypothesis was that wallets with persistent realised profit lead price when they act together. Testing it requires only that skill persist across time, and that is a rank correlation on 23 wallets across three non-overlapping windows.
Correlation ran +0.37 in the first transition and -0.32 in the second. Neither was significant and the direction reversed. The premise failed and no strategy was built on it.
Oracle-tether premium. The hypothesis was that venue price mean-reverts to an externally anchored oracle. Characterisation passed cleanly:
| Statistic | Value |
|---|---|
| Half-life, median across 24 instruments | 3.9 s, range 1.7 to 11.8 |
| Stationarity, all instruments | ADF p at or below 1e-10 |
| Net events per day, universe-wide | 1,928 |
The reversion is real, fast, and frequent. The strategy that tried to trade it by chasing the dislocation returned -11.65 basis points net across 720 out-of-sample trades, and -2.56 gross, meaning it lost before any fee was applied. The win rate was 0.3 percent.
That pairing is instructive on its own. A characterisation can pass decisively and the strategy built on it can fail before costs, because a mean reversion with a 3.9 second half-life is not reachable by an operator arriving at 96 milliseconds without colocation and quoting rather than crossing.
4. Three exits, five days
The flagship hypothesis was cascade exhaustion: after forced liquidation flow exhausts, price dislocation reverts. There are only three ways to exit such a position, and the programme tested all three.
Fade it. Enter against the cascade and wait for reversion. Net -5.97 basis points, interval [-11.8, -0.2], failing a matched null at p=0.56. A high-power grid confirmed it at -8.30 with p=0.96.
The diagnostic mattered more than the verdict: the recovery fraction came out negative at -0.675. Price was not reverting after exhaustion. It was continuing. The premise was inverted, not merely unprofitable.
Follow it. If price continues, ride the continuation. Gross negative at both entry points, -1.59 and -5.84 basis points, before any fee.
Quote into it. Post passively rather than crossing, and let the fee schedule do the work. Every arm gross negative at -0.35 basis points on majors, and mid-caps worse rather than better at -1.93 and -1.43.
Three mechanisms, three closures, five days. That rate is the return on the apparatus built in June. The same work in May would have taken a month per hypothesis.
5. The fact that was real
One measurement in this period passed everything and could not be monetised.
Adverse selection differs by aggressor identity. Large participants impose 0.906 basis points more markout on a passive counterparty than retail flow does, with a 95 percent interval of [0.757, 0.983], a permutation p-value of zero, on 4.51 million fills. The effect persists for 10 to 30 seconds.
That is as well-established as anything in this archive. It says that if you can tell who is hitting your quote, you can price accordingly.
The replay that tried to monetise it returned gross negative across every configuration and every instrument tested. A 0.906 basis point differential does not cover a round trip, and selecting harder does not raise it.
The finding is recorded as CONFIRMED in the Kill Ledger, with no verdict and no
transferability classification, because it is a fact about the venue rather than a
hypothesis about a strategy. It is one of the more valuable things in the record
for anyone whose cost structure is different.
6. The audit that made the verdicts worse
Before the venue track was paused, the accumulated verdicts went through adversarial review.
The review found that the passive-fill replay had been undercounting trade-through fills, where a trade sweeps past a resting level without producing a discrete event at that level. At one instrument, 55.7 percent of adverse fills were being missed.
Every missed fill was an adverse one. The reported gross figures had therefore been an optimistic upper bound rather than an estimate.
The matcher was corrected with eleven tests and the verdicts were re-run. All of them were upheld, with worse numbers than before.
An audit that only ever improves your results is not an audit. This one is recorded because it went the other way.
7. Three timing levers, closed
A remaining hypothesis held that session or time of day was a hidden lever: cheaper spreads at some hours, more events at others, or better reversion quality in one session.
| Lever | Measurement | Result |
|---|---|---|
| Cost | Half-spread across 24 hours | Flat; 90th to 10th percentile ratio 1.01 |
| Event supply | Cascade events by session | Proportional to session clock share, 0.80 to 1.20 times |
| Reversion quality | By session, pre-registered event study | No session pair with non-overlapping intervals; pooled +4.39 bps, CI95 [-0.3, +9.0] |
All three closed. The first two are pure characterisation with no forward return computed at all, which is the cheapest way to close a lever and was deliberate.
8. How the track ended
Paused, not closed, on 2026-07-14.
One hypothesis remained open. Passive entry after cascade exhaustion filled 17 of 21 events at an 81 percent fill rate and returned -2.47 basis points at 120 seconds with an interval of [-9.21, +5.57]. The first kill criterion fired, because the interval spans zero. The second did not: the result beats a matched null at p=0.0102, and exhaustion conditioning contributes +4.8 basis points over a generic fade.
Under R-06 that is NOT PROVEN with the mechanism intact, not a falsification.
It stayed open, and it is one of the two leads the programme carried out.
9. Cross-references
BL-D-007 the doctrine governing the gates used here. BL-D-010 section 2.3 for
the matcher correction. BL-FM-002 sections 4 and 5 for the venue's latency and
cadence parameters. Ledger rows S1-MC, S2, S3, S3-C, S4-G1, S4-R,
M3B, IB1, WC0, WC1, BR1, TOC1, TOC2, TOC3, S5.
Copyright (c) 2026 Bijective Labs. All rights reserved.