BL-SD-025: Oracle-Tether Premium, Characterised

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-08

Venue: Hyperliquid perpetuals Programme: I Verdict: CONFIRMED Blocking axis: not applicable, this is a fact rather than a hypothesis Transferability: not applicable, confirmed rows carry no transferability classification

1. Identification

This dossier is not a strategy verdict. It records a fact about the venue's own mechanism, measured before any trading design was tested against it, and it is the premise that the chase-entry variant in BL-SD-024 and the pre-positioned maker variant in BL-SD-027 were each built to monetise. Both failed. This entry exists so a reader with a different execution capability can evaluate the underlying fact on its own terms, separate from either failure.

2. Economic thesis

Hyperliquid's oracle is a liquidity-weighted median of centralised-exchange spot, and it governs both liquidation and hourly funding. Local flow on the venue pushes the venue's own mid away from that oracle; convergence back is not a statistical hope but a mechanism: cross-venue arbitrageurs, a funding penalty accruing every hour against the side sitting away from the anchor, and the structural fact that Hyperliquid's own flow does not move the spot median it is measured against. If that mechanism is real, the premium series should be stationary rather than a random walk, and it should revert on a timescale short enough to be relevant to a trading design.

This dossier does not name a forced loser, because it characterises a mechanism rather than proposing a trade. The question answered here is narrower and prior to that one: does the oracle tether actually produce a measurable, fast, deep reversion, on this venue, in this data era, before any capture method is designed against it.

3. Falsifiable claim

Pre-registered in the design document, section II.S3.A.2, before the characterisation was run:

There are at least 8 events per day, universe-wide, with a net margin at or above three times the maker-to-maker round trip, and a half-life at or below a 300-second holding-cap proxy. Passing this gate earns the right to proceed to a full screen; it is not itself an edge verdict.

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
 
Stationarity test, per coin, over the recorded series:
 
  ADF(prem)  ->  p-value against the unit-root null
 
Half-life, per coin, fit to an OU / AR(1) approximation of the recorded series:
 
  half_life = ln(2) / (- ln(phi)),   phi = AR(1) coefficient of prem_t
 
Event definition, at threshold thr and slippage-adjusted for test size s:
 
  event at t  when  |prem_t|  >=  thr,  after subtracting the impact-based
                     slippage bound for size s
 
Net-qualifying event:
 
  event qualifies  when  net margin (event magnitude, less round-trip cost)
                          >= 3 * maker-to-maker round trip (3.0 bps at tier 0)
 
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}

Causality holds by construction: every quantity in prem_t is observed venue state at time t, and the ADF and half-life estimates are fit on the trailing recorded series with no forward reference. There is no entry, exit, or sizing rule in this specification because this measurement characterises a series; it does not trade one.

5. Operating envelope

6. Instrument

The premium characterisation instrument, test-driven with 5 of 5 unit tests passing, offline and lake-only. Slippage is bounded conservatively as half the impact spread, an upper bound for every test size against the impact notional on this venue.

7. Results

Universe-wide across all 24 coins, ADF rejects the unit-root null everywhere, p at or below 1e-10, with a median half-life of 3.9 seconds and a range of 1.7 to 11.8 seconds. The mid - oracle comparison series is consistent with the API-exact field, for example BTC at 11.8 seconds against 13.0 seconds respectively.

8. Verdict, blocking axis, and transferability

Verdict: CONFIRMED.

This is a fact about Hyperliquid's own venue mechanism, not a strategy, and it carries neither an axis of death nor a transferability classification: those fields exist to describe why a trading hypothesis died and whether the finding generalises. Nothing died here. What was confirmed is that the oracle tether produces a stationary premium series, rejecting a unit root in every one of 24 coins tested, with a median half-life of 3.9 seconds, and that on the four major coins the reversion is both fast and deep: 22 to 43 percent of an initial dislocation remains after 5 seconds, and time-to-half is 2 to 3 seconds.

Why it could not be monetised as measured. Two capture mechanisms were built against this fact and both failed. A chase-entry variant, BL-SD-024, detects the dislocation and quotes a maker order at the moment of detection; it failed because its conservative fill model confirms entries only after price has traded through the quoted level, by which point the 2-to-3-second reversion this characterisation measured has largely already happened. A pre-positioned maker variant, BL-SD-027, quotes ahead of any dislocation and lets the displacing flow arrive; it failed because adverse selection against the maker's own book exceeded the spread available on this venue's majors, independent of the premium mechanism entirely. Neither failure contradicts the fact recorded here: the premium reverts exactly as measured. What is missing is an entry mechanism fast enough, or an execution structure cheap enough, to sit inside a 2-to-3-second window without absorbing more adverse selection than the reversion is worth.

A second finding worth stating plainly: raw event frequency is a misleading guide to which coins are candidates. PYTH and FARTCOIN carry the most events per day, 2,419 and 2,025 respectively, but their reversion is shallow rather than deep, 0.66 to 0.76 of the initial premium still present at 300 seconds. A high event count on those two coins reflects a persistent premium bias, not a repeated dislocation-and-recovery pattern, and a reader selecting a universe from event count alone would be selecting the wrong coins.

For a reader with a different cost structure or execution capability, this dossier states the fact this programme could not turn into an edge: the reversion exists, is fast, and is deep on the four major coins tested. An operator with sub-second maker placement, colocated infrastructure, or a materially lower adverse-selection profile than this programme's measured execution should treat this as the starting characterisation to build against, not as a closed question.

9. Cost and capacity analysis

No cost or capacity analysis applies to a characterisation. The 3.0 basis point maker-to-maker round trip used in section 4 defines which premium crossings count as net-qualifying events; it is not a claim that any specific trade nets that amount, since no trade was modelled here.

10. Envelope sensitivity and reopening conditions

Not applicable in the form used for FALSIFIED or CLOSED rows. This dossier records a fact, not a verdict that reverses under a parameter substitution. The relevant sensitivity is already stated in section 8: the fact holds regardless of an operator's cost schedule; what varies by operator is whether their execution mechanism can act inside the 2-to-3-second reversion window this characterisation measured, on the four coins where the reversion is deep rather than persistent.

Three caveats belong here, carried forward honestly from the characterisation itself. Event counting uses a re-arming crossing definition, so a cluster of volatility can inflate the raw count without producing that many independent tradeable dislocations. The ratio path reported in section 7 is a measure of the premium's own decay, not a PnL path; a conservative maker fill model, funding accrual, and walk-forward testing belong to a full screen, not to this characterisation. And the data era is three days, so subperiod stability of the half-life and ADF result was not tested here.

11. Reference implementation

Published as 05_INSTRUMENTS/strategies/python/src/oracle_premium.py with 40 tests, alongside BL-SD-024, the strategy dossier built against this characterisation. The two must be read in that order: this row established that the premium exists, is large enough to matter, and mean-reverts, and it is filed CONFIRMED because a characterisation instrument answers no question about whether anything can be traded. BL-SD-024 then failed to capture it. Reading the second row as falsifying the first would discard a measured venue property on the strength of one design that could not harvest it.

One data rule in the published instrument does more for this row's credibility than any of its statistics. Sampling gaps are cut and never bridged, so each contiguous segment is resampled on its own. Interpolating across an outage would manufacture exactly the finding being looked for: a premium wide before a gap and narrow after appears, bridged, as a smooth reversion that nothing observed. The instrument would then report reversion strongest precisely where the data is worst, and the correlation between missing data and apparent mean reversion would read as a mechanism. Cutting costs sample, and a smaller honest sample beats a larger one whose extra observations were invented. BL-D-006 is the article.

Publishing it added one guard the source did not have, and the way it surfaced is worth recording. A test constructed its nanosecond timestamps by multiplying an index array, which on this platform defaults to a 32-bit integer, and an epoch in nanoseconds does not fit in one. The wrapped array came back as a single contiguous segment and looked entirely ordinary. segment_on_gaps now rejects a series that is not non-decreasing, which catches both that and the more obvious case of input that was never sorted. Neither would have failed loudly before.

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, and is registered as D-014 in BL-A-006.

13. Lessons extracted

The clearest lesson from this dossier is the separation of a premise gate from an edge verdict, stated explicitly in its own pre-registration: passing this gate earns the right to build a full screen, it is not itself a claim of profitability. Both full screens built against this premise subsequently failed, and the record is precise about attributing those failures to capture method rather than retroactively reading them back into this characterisation.

The second lesson concerns event-count selection. FARTCOIN and PYTH produced the highest raw event counts in the universe and would have been the obvious candidates for a full screen built on frequency alone. Both have shallow reversion and represent a persistent premium bias rather than a dislocation-and-recovery mechanism. Universe selection for this family was corrected to use reversion depth rather than event frequency before the full screen was built, and that correction is recorded here as the reason BTC, ETH, SOL, and XRP were the coins carried forward.

14. Cross-references

BL-SD-024 for the chase-entry capture attempt against this characterisation, falsified on execution speed. BL-SD-027 for the pre-positioned maker capture attempt, falsified on adverse selection. Ledger row S3-C.

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