BL-SD-027: Monetising the Identity Gap Through Selective Quoting

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 entry gate confirmed that adverse selection differs sharply by aggressor identity, which is a necessary condition for state-conditional quoting but not a sufficient one. This dossier answers the sufficiency question directly: can a maker that avoids quoting into large-account flow actually clear its costs, given the spread this venue's book actually offers.

2. Economic thesis

Hyperliquid's matching is price-time FIFO on-chain, with maker fees at or near zero at tier zero and rebates available at volume. A maker that knows, at the moment of a fill, which class of participant just took its quote, could in principle quote continuously but pull or widen around large-account activity, capturing the spread from uninformed flow while avoiding the worst of the adverse selection this family's own entry gate measured. If the spread available on this venue exceeds the adverse selection carried even by the safest class of filler, the arithmetic favours the maker; if it does not, no amount of selectivity rescues the position.

Who is forced to lose. The uninformed taker, retail flow and forced liquidation flow that demands immediacy without pricing information into its urgency, is named as the structural loser at design time: it pays the spread to a maker that has done nothing but wait. The question this dossier settles is whether that structural relationship is strong enough on this venue to overcome what the maker loses to the informed class it cannot fully avoid.

3. Falsifiable claim

Pre-registered before any virtual fill was simulated:

State-conditional quoting, avoiding fills when recent aggressive flow is dominated by large accounts, produces a mean net return per fill above zero at the maker-to-maker round trip.

4. Mathematical specification

Notation per BL-D-011.
 
Virtual quote, placed on a book snapshot subsampled every S = 5 s per coin:
 
  at t0:  quote bid at P_b = best_bid(t0),   queue-ahead  Q_b = sum sz at level P_b
          quote ask at P_a = best_ask(t0),   queue-ahead  Q_a = sum sz at level P_a
 
Fill condition, FIFO, evaluated only on trades after t0, within a wait window W_wait = 30 s:
 
  bid fills  when  sum{ sz of aggressive-SELL trades at px == P_b, after t0 }  >  Q_b
  ask fills  when  sum{ sz of aggressive-BUY  trades at px == P_a, after t0 }  >  Q_a
  no fill within W_wait  ->  quote cancelled, no PnL (realistic non-fill)
 
Fill outcome:
 
  side_sign =  +1  bid fill (maker goes LONG)
               -1  ask fill (maker goes SHORT)
  fill_px   =  P_b  or  P_a, whichever side filled
 
Exit, marked to mid at a fixed horizon from the fill:
 
  gross_bps = side_sign * ( mid(t_fill + Delta_exit) - fill_px ) / fill_px * 1e4
              Delta_exit = 5 s primary; {10, 30} s secondary, robustness only
 
Cost and net:
 
  net_bps = gross_bps - fee_rt
            fee_rt primary = 3.0 bps (maker-to-maker); sensitivity 6.0 bps
            (maker-to-taker) and 0 bps (rebate tier)
 
Arms:
 
  Arm A (unconditional):  every virtual fill
  Arm B (state-conditional):  quote suppressed when the share of notional from
    LARGE-class aggressive flow over a trailing R = 10 s window  >=  w_safe
    w_safe tested on the grid {0.3, 0.5}

Causality holds by construction: Q_b and Q_a are read from the book snapshot at t0, and fill evaluation uses only trades timestamped after t0. Cancel activity ahead of the virtual quote is not observed in the 250-millisecond book sample, which the design records as a known conservatism: the replay cannot see orders ahead of the virtual quote being cancelled, so it may understate how quickly the virtual quote would reach the front of the queue, and any fills the replay does produce are correspondingly a conservative estimate of true fill likelihood.

5. Operating envelope

6. Instrument

The identity-aware maker replay instrument, test-driven, implements fill-time detection, the corrected trade-through fill matcher, maker_pnl_bps, and reuses the identity classifier and bootstrap functions from earlier instruments in this family.

7. Results

Primary run, majors, gross return per fill before fee:

All three arms are negative before any fee is applied. State-conditional quoting narrowed the loss by roughly 0.01 basis points relative to the unconditional arm, which is not a material protection against a loss of this size.

Verification pass, 14-coin universe, per-coin gross returns:

No coin in the 14-coin universe produced a significant positive gross return. Mid and small-capitalisation coins performed worse than majors rather than better, reversing the hypothesis that a wider quoted spread on thinner books would compensate for adverse selection.

8. Verdict, axis of death, and transferability

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

The kill criterion fires cleanly: Arm B's mean net return is negative at the primary fee, and the negative result holds at the gross level, before any fee is even applied. This is a mechanism failure rather than a cost failure. Setting the round-trip cost to the most generous sensitivity tested, a full rebate at 0 basis points, still leaves every arm and every coin losing money, because the loss originates in the spread-versus-adverse-selection arithmetic itself, not in the fee schedule layered on top of it.

Two findings sharpen why. First, the state-conditional arm barely differs from the unconditional one, -0.340 to -0.346 basis points against -0.349, which means avoiding the highest-adverse-selection class of filler does not meaningfully change the outcome. The reason is arithmetic, not a flaw in the conditioning logic: on this venue's majors, half-spread is thinner than the adverse selection carried even by the lowest-adverse-selection class measured in the entry gate, retail at 0.32 basis points and forced flow at 0.10 basis points. A maker loses to retail flow alone before large accounts are even considered, so filtering out large accounts removes a loss that was never the binding one.

Second, the verification pass closes the obvious escape route, that a thinner and less competitive market might offer wider spreads uncontested by professional makers. Mid and small-capitalisation coins have half-spreads five to twenty-five times wider than the majors, yet every one of them produced a more negative gross return, not a less negative one. Adverse selection rises faster than spread widens in a thinner book, which is a statement about venue efficiency rather than about liquidity: the market prices its own spread close to its own adverse selection, on every coin tested, and there is no coin where the gap favours the maker.

No envelope substitution rescues this result, which places it in the UNIVERSAL class by the BL-D-002 section 4 test. A gross return already negative at zero fee has no reversal point in cost, and a universe already searched from majors to small caps has no reversal point in venue selection within the instruments available to this programme.

A cross-check worth recording: the gross loss on majors, approximately -0.34 to -0.35 basis points, is close to the negative of the retail-class adverse selection measured in the entry gate, +0.32 basis points, once the near-zero half-spread on majors is added back. The two independently built instruments, the entry-gate markout measurement and this virtual-fill replay, agree with each other to within noise, which is evidence that the finding is a real property of the venue rather than an artifact of either instrument.

9. Cost and capacity analysis

Gross, Arm A, primary majors run      -0.349 bps
Gross, Arm B, w_safe = 0.5            -0.340 bps
Fee, maker-to-maker (best case)        3.000 bps
Net, best arm, best fee               -3.340 bps

No capacity measurement was taken. Frequency was never the binding question: at 33,140 virtual fills on three coins over the data era, sample size was more than adequate, and the finding failed on unit economics per fill rather than on scale.

10. Envelope sensitivity and reopening conditions

No binding parameter exists to name, because the finding is negative at the gross level, before any fee is applied.

Reversal condition:  none computable.

Setting fee_rt to zero, the most favourable value reachable on this venue's published schedule, still leaves every arm and every coin tested with a negative return. The mechanism itself, spread priced close to adverse selection on every coin in the universe searched, does not reverse under a cost substitution, which is why this hypothesis carries the UNIVERSAL rather than ENVELOPE-BOUND classification, following the same test applied throughout this record.

What would reopen it: a coin or a venue where the half-spread available to a maker exceeds the adverse selection of the lowest-adverse-selection filler class by a wide margin, which was not found in the 14-coin universe searched here. A materially different queue-priority mechanism, one that let a maker's order reach the front of the queue meaningfully faster than the 250-millisecond book sample used in this replay can observe, is a second candidate, untested because it would require tick-exact book data this programme's recorder did not capture at the time of this measurement.

What would not reopen it: moving to a thinner or less liquid coin expecting a wider uncontested spread. That was tested directly across 14 coins and the relationship runs the opposite direction: thinner books carried worse gross returns, not better ones.

11. Reference implementation

Published as 05_INSTRUMENTS/strategies/python/src/maker_replay.py, with 21 tests. The asset register marks this family, alongside BL-SD-024, as an avenue that should not be re-run on the same design expecting a different answer.

Both fill matchers are published side by side on purpose. The original was FIFO only: accumulate opposing prints at our level until the queue ahead is consumed. The audit of 2026-07-14 found it missed every print that swept through the level, which is a certain fill whatever the queue position and the most adverse case, since the resting order fills into a move against it. That was 55.7 percent of fills. The corrected matcher treats a through print as an immediate fill, and every arm this row measured came out gross negative once it did. One test hands both matchers the same prints and shows the original reporting no fill where the corrected one reports a fill into a loss.

12. Evidence register

This measurement is recorded in the pre-registration ledger rather than as a separately registered evidence artifact, and is registered as D-021 in BL-A-006.

13. Lessons extracted

This dossier, together with BL-SD-026, is the clearest worked example in this venue's record of the distinction between a confirmed microstructure fact and a monetisable edge. The fact, that identity separates adverse selection with overwhelming statistical power, is real and would matter on a venue where the spread-to-adverse-selection ratio ran the other way. It did not on this venue at this measurement, and the two independent instruments cross-validating to within noise is the finding worth carrying forward as a method: when a strategy fails, a second independently built instrument agreeing on the magnitude of the failure is stronger evidence than either instrument alone.

The verification pass across 14 coins closes a specific temptation that recurs whenever a strategy fails on liquid majors: retreat to a smaller, less-scrutinised market and hope the competition thins out enough to leave an edge. Here it did not, and the reason it did not is recorded precisely rather than left as an assumption for the next hypothesis to inherit.

14. Cross-references

BL-SD-026 for the entry-gate confirmation this screen was built to test against. BL-SD-024 for the other Hyperliquid maker-side design falsified on execution mechanics rather than on identity. BL-D-002 section 4 for the reversal-point test applied in section 10. Ledger row S4-R.

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