BL-SD-024: Oracle-Tether Premium Reversion, Chase Entry

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-08

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

1. Identification

The characterisation gate ahead of this measurement passed with a wide margin: the premium is stationary, the median half-life across the universe is 3.9 seconds, and majors show 22 to 43 percent reversion within 5 seconds. This dossier tests one specific way of capturing that reversion. It is not a test of whether the premium reverts; that question is answered separately in BL-SD-025, and it is confirmed.

2. Economic thesis

Hyperliquid's perpetual price is structurally tethered to a liquidity-weighted median of centralised-exchange spot, used both as the oracle for liquidation and as the anchor for hourly funding. Local flow on Hyperliquid, retail bursts, small cascades, a single large market order, pushes the venue's own mid away from that oracle. Convergence back is not a statistical hope; it is a venue mechanism, driven by cross-venue arbitrageurs, an hourly funding penalty on the side sitting away from the oracle, and the fact that Hyperliquid's own flow does not move the oracle it is measured against. This variant detects the dislocation as it happens and chases it: quote a maker order at the current mid the instant the premium crosses a threshold, in the direction of convergence, and exit either at convergence or at a fixed time stop.

Who is forced to lose. The impatient local taker, retail momentum on Hyperliquid paying the taker fee plus slippage to chase a local move, and the bot that follows that move a beat late. Structurally, there is always local flow that does not price the basis to the oracle, and the venue guarantees the magnet through the oracle and funding mechanism, whether or not any single chase-entry attempt can capture it before it closes.

3. Falsifiable claim

Pre-registered in the design document, section 3.7, before the full screen was run:

Conditional reversion, net of the round trip and funding accrual, beats an unconditional reversion null with sufficient sample and without profit concentrated in a small fraction of events.

4. Mathematical specification

Notation per BL-D-011.
 
Premium (API-exact field, verified V0 item 3):
 
  prem_t = [ max(impact_bid_t - oracle_t, 0) - max(oracle_t - impact_ask_t, 0) ] / oracle_t
 
CEX-quiet gate, lake proxy for a live cross-venue feed:
 
  gate_t = 1{ |oracle_t - oracle_{t-w}| / oracle_t < 1 bp },   w in {1s, 3s, 10s}
 
Chase-entry signal, fired at detection time t0:
 
  fire at t0 when  |prem_t0| >= thr_bps   AND   gate_t0 = 1
  direction  d = -sign(prem_t0)              (fade toward the oracle)
 
Entry:
 
  quote maker at mid_t0 in direction d; fill recognised only when price trades
  THROUGH the quoted level, not merely touches it (conservative fill model)
 
Exit:
 
  at convergence, |prem_t| <= eps * |prem_t0|, or at a fixed time stop
  t0 + t_stop, whichever occurs first; exit leg crosses (taker)
 
Sizing:
 
  equal notional per event; walk-forward, fit on day d, test on day d+1
 
Cost:
 
  c_rt = maker entry (1.5 bps) + taker exit (4.5 bps) + funding crossings
       during the hold, before slippage

Causality holds by construction: prem_t0, the oracle history feeding the gate, and mid_t0 are all observable at or before t0; the exit condition is evaluated only on data after t0.

The fill model is deliberately conservative, requiring price to trade through the quoted level rather than merely touch it. That choice is not incidental to the result: it is the mechanism that section 7 identifies as the primary driver of the loss, because it means the fill is confirmed precisely when the premium has widened further against the position, not when it has begun to narrow.

5. Operating envelope

The three-day data era is short by the programme's own subperiod-stability standard, and that limitation is reported rather than smoothed over. It did not affect the direction of the verdict, because the failure shows up in-sample on every fold and every grid point before the out-of-sample test is even run.

6. Instrument

The premium-reversion screening instrument, test-driven with 4 of 4 unit tests passing, implements the event simulation, the funding-crossing count, the null-index sampler, the day-level grid builder, and the block-bootstrap confidence interval used throughout.

Every one of the 16 grid combinations, across all 10 in-sample folds, returned a negative result before the out-of-sample test was run. That is reported here because it is the strongest single fact in this dossier: the out-of-sample failure was not a surprise reversal from a promising in-sample fit, it is a design that never once produced a positive in-sample result on any grid point tested.

7. Results

Win rate on the conditional arm was 0.3 percent. The conditional variant did not beat the unconditional null; both are deeply negative, meaning the premium-threshold gate did not distinguish a capturable dislocation from a generic one. Every in-sample fold across every grid point tested was negative, which is the in-sample analogue of the same finding.

8. Verdict, axis of death, and transferability

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

Three kill criteria fire together: the net confidence interval lies entirely below zero on 720 out-of-sample trades, the conditional arm fails to beat the unconditional reversion null, and the loss is concentrated across every grid combination tested rather than in a fragile tail that a parameter change might avoid.

The decisive fact is the gross figure. Mean gross return, before any fee, funding, or slippage, is -2.56 basis points. This is a mechanism failure, not a cost failure: the chase-entry design loses money even at zero cost, because the fill model's conservative through-price requirement means the entry is confirmed only after the premium has already widened past the quoted level, and the characterisation gate measured a time-to-half reversion of 2 to 3 seconds. On a 1-second detection and fill-confirmation grid, by the time a chase-entry fill is recognised, the reversion this design intends to capture has largely already happened, leaving only the continuation drift that follows it.

No envelope substitution rescues a design whose gross return is negative before any cost is applied. That places the transferability classification at UNIVERSAL rather than ENVELOPE-BOUND, by the same BL-D-002 section 4 test applied throughout this record: at arbitrarily favourable cost, latency, or capital parameters, this specific capture mechanism would still lose.

This finding is precise about what failed. The oracle-tether premise itself is confirmed independently in BL-SD-025: the premium is stationary and reverts fast and deep on majors. What failed here is the chase-entry method of capturing that reversion, not the reversion itself.

9. Cost and capacity analysis

Gross return, out-of-sample          -2.56 bps
Round trip, maker-to-taker            6.00 bps
Net return, out-of-sample            -11.65 bps

The gap between gross and net is roughly 9 basis points rather than the nominal 6-basis-point round trip, which reflects funding crossings accrued during the hold and is consistent with the design's own kill criterion on funding accrual. No capacity measurement was taken. The screen failed before frequency or size became a binding question, since a negative gross return cannot be rescued by capacity at any scale.

10. Envelope sensitivity and reopening conditions

No binding parameter exists to name, because the finding is a sign reversal at the gross level, not a magnitude shortfall against cost.

Reversal condition:  none computable.

Setting the round-trip cost to zero and removing funding accrual entirely still leaves a mean gross return of -2.56 basis points. No fee tier, no maker rebate, and no execution channel converts a negative gross move into a profitable trade. This is the same structural finding as BL-SD-022: the mechanism itself points the wrong way for this specific capture method, independent of any operator's cost schedule.

What would reopen it: a different capture mechanism entirely, not a retuned version of this one. The characterisation gate's own recommendation, made before this screen was run, is a pre-positioned resting maker order queued before the dislocation forms, so that the order receives its fill from the flow that is causing the displacement rather than chasing the displacement after detection. That design was built and tested separately as the identity-aware maker replay; see BL-SD-027, which also failed, on different grounds.

What would not reopen it: retuning the chase-entry threshold, gate window, exit fraction, or time stop on this design. All 16 grid combinations across 10 in-sample folds were tested and all were negative before the out-of-sample confirmation. This is not an underpowered result waiting on a better parameter; it is a design whose fill mechanics are structurally too slow for the reversion speed the venue's own oracle tether produces.

11. Reference implementation

Published as 05_INSTRUMENTS/strategies/python/src/oracle_premium.py with 40 tests, shared with BL-SD-025, as a worked example of a capture-method failure distinct from a premise failure. The asset register marks this tool as an avenue that should not be re-run on this design expecting a different answer.

The distinction between the two failures is the point of publishing it, and two properties of the instrument are where it lives.

Only the threshold-crossing tick fires an event, and the detector re-arms only once the premium returns inside the threshold. Without that, a premium sitting wide for two hundred seconds would produce two hundred observations of one event, and every interval computed afterwards would collapse because the observations are near-copies. The instrument answers how often the condition begins, not how long it lasts.

The quiet gate is on the oracle rather than on the premium, and it is a wrong-way risk control rather than noise reduction. A wide premium during a fast oracle move is a perpetual lagging a real price change on the reference venues, and fading it is a bet against that change. A wide premium while the oracle sits still is the case the mechanism actually claims something about. Gating on the premium's own stability would not separate the two, and the first is where the losses are.

12. Evidence register

This measurement is recorded in a narrative verdict report, backed by a machine-readable JSON output retained in the venue data lake, rather than as a registered evidence artifact, and is registered as D-013 in BL-A-006.

13. Lessons extracted

This hypothesis is the clearest demonstration in the Hyperliquid record of the difference between a premise failure and a capture-method failure. The premise, that Hyperliquid's premium reverts toward its oracle, is real and was confirmed with the strongest statistics in this venue's record before this screen was ever run. What failed is one specific way of monetising that reversion, chase entry on a 1-second grid, and the failure was diagnosed to a specific mechanical cause: the conservative through-price fill model confirms entries exactly when the market has moved against the position, and the reversion is simply faster than a chase can follow.

The exhaustive in-sample grid, 16 combinations across 10 folds with a uniformly negative result, is itself worth preserving as a rule of practice: a design that cannot produce a single positive in-sample cell before its out-of-sample test is not a design waiting on more data, and running the out-of-sample confirmation anyway is cheap enough to do for the record, but the verdict was effectively settled before it ran.

14. Cross-references

BL-SD-025 for the characterisation of the same premium that confirms the mechanism this variant failed to capture. BL-SD-027 for the pre-positioned maker capture design this falsification motivated, which also failed. BL-D-002 section 4 for the reversal-point test applied in section 10. Ledger row S3.

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