BL-SD-046: Earning the Spread as a Maker

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-08

Venue: Bybit USDT perpetuals Programme: II Verdict: CLOSED Axis of death: Adverse selection Transferability: ENVELOPE-BOUND

1. Identification

This row exists to record a sequence rather than a single measurement, and the sequence is the point. Door 3 of BL-P2-001 was closed twice. The first closure was reached on seven instruments and was wrong. The second was reached on all 676 and held.

2. Economic thesis

A resting order at the best bid and offer earns the quoted spread from liquidity takers who pay for immediacy, provided the spread exceeds the round trip cost of the two legs. That is the entire proposition: no prediction of direction, no view on the market, only compensation for standing ready to transact.

Who is forced to lose, if the mechanism works as stated. Uninformed takers who value immediacy enough to cross the spread rather than wait. The reciprocal risk, and the one that ultimately closes this door, is that the maker itself becomes the forced loser whenever the taker on the other side of the fill is not uninformed at all.

3. Falsifiable claim

No pre-registration exists for this row; a cost-floor check precedes the pre-registration budget by design under BL-D-008. The claim is the two-gate criterion the layer imposes, stated before either gate was run:

Making is worth pursuing only if the universe quotes spread in excess of the maker round-trip cost (gate 1), and only if that spread survives what happens to price after the fill (gate 2). Gate 2 is the one that decides, because a maker without an information edge can only ever earn the first two components of the spread and never the third.

4. Mathematical specification

Notation per BL-D-011.
 
Quoted spread, instrument i, snapshot t:
 
  spread_i(t) = ( ask_i(t) - bid_i(t) ) / mid_i(t) * 1e4
 
Gate 1, eligibility to quote:
 
  eligible_i(t) = 1   if spread_i(t) >= c_rt_maker
                = 0   otherwise
 
  c_rt_maker = 2 * fee_maker      (round trip, both legs resting)
 
Fill classification, for trade k against a resting maker order:
 
  taker_dir_k = +1  if the aggressor bought (maker is short from the fill)
              = -1  if the aggressor sold   (maker is long from the fill)
 
Markout at horizon tau, maker side:
 
  markout_k(tau) = -taker_dir_k * ( mid(t_k + tau) - fill_price_k ) / fill_price_k * 1e4
 
Realised adverse selection at tau, against the half-spread earned at fill:
 
  AS_i(tau) = half_spread_proxy_i - markout_i(tau)
 
Net expected result per fill, at horizon tau:
 
  E[r_net] = half_spread_proxy_i - AS_i(tau) - c_rt_maker
           = markout_i(tau) - c_rt_maker
 
Sizing:  equal notional per fill, no inventory skew modelled
Cost:    c_rt_maker = 4.0 bps at Bybit tier 0 (2.0 bps per side, unverified, see section 5)

Causality holds for the entry decision: eligible_i(t) depends only on book state observable at t. markout_k(tau) requires the forward mid at t_k + tau and is not available at fill time; it enters nowhere in the entry decision and exists purely to evaluate, after the fact, what a maker who had already been filled would have realised. This is the same structure used for the exhaustion timestamp in BL-SD-022: causal at the decision, retrospective at the evaluation.

Where the original measurement left a term implicit: no queue-priority model is applied. The markout assumes every resting order at the touch is filled with certainty, which is optimistic; a maker who only fills when the queue clears to it is filling disproportionately into the trades that moved the book enough to reach it, which is a worse selection than the average reported here, not a better one.

5. Operating envelope

The out-of-era nature of gate 1 is itself an envelope fact worth stating plainly. It answers exactly one question, whether any part of the universe quotes wider than the maker round trip today, and it answers nothing about history.

6. Instrument

Gate 1: the spread census, BL-IS-011, with the spread primitive (14 tests), pulling L1 order book, instrument tick size, and 24-hour turnover from three independent venue endpoints and cross-checking two of them against each other.

Gate 2: the markout instrument, BL-IS-006, with the markout primitive (16 tests), computing markout from the public tick archive. The sign convention, that a taker-buy fill leaves the maker short, was verified empirically against 44,969 trades on a single instrument before a line of the markout logic was written, using the PlusTick and ZeroPlusTick fields as an independent check.

The control instrument, the venue's most liquid perpetual, gives the sanity check the instrument had to pass before any candidate was trusted: its quoted spread of 0.02 bps guarantees a negative result if the instrument is built correctly. It returned negative on 5 of 5 days.

7. Results

Gate 1, spread census across the full universe, cross-check between two independent endpoints agreeing to a median relative error of 0.23 percent:

An earlier snapshot taken 29 minutes before the one above returned a median of 5.70 bps. The artifact value of 5.26 is the one carried forward, per the correction recorded in BL-SD-000 section 7; the gap between the two snapshots is itself informative, since a dollar-pool estimate built on either one moved by roughly 40 percent between passes while the qualitative reading, that most of the universe quotes above the fee floor, held in both.

Gate 2, adverse selection, control and five mechanically selected candidates, 5 days, horizon 10 seconds and 60 seconds, fee grid centred on 2.0 bps per side:

The one instrument that appeared to survive, AXTIUSDT, carried 72.5 percent of its trades in a single day of five, with daily net swinging from -12.96 to +7.79 bps. A result built on one day is not a market-making result, and it is the same class of error the record already retracted once, in the dislocation verdict now marked WITHDRAWN.

The decisive comparison, adverse selection against the half-spread actually available, on the four instruments not excluded for concentration or thin sizing:

Adverse selection exceeds the half-spread on every row. The headline comparison, the control at the 60-second horizon: markout -0.53 bps against a half-spread proxy of 0.14 bps, on 9,352,722 trades.

8. Verdict, axis of death, and transferability

Verdict: CLOSED. Axis: adverse selection. Transferability: ENVELOPE-BOUND.

The door closed on the correct gate this time. Gate 1 established that most of the universe quotes spread comfortably above the maker round trip; the earlier closure on a seven-instrument sample understated the true median by a factor of four and was wrong on its own terms. Gate 2 then closed the door for a different and better reason: the price movement that follows a maker's fill consistently exceeds the half-spread that fill was meant to earn.

Applying the BL-D-002 section 4 test: would arbitrarily favourable envelope parameters rescue this? A cheaper fee moves the eligible universe in gate 1 but does nothing to gate 2, since adverse selection scales with how attractive the quote is rather than with the fee paid to post it. What would rescue it is a faster cancellation path, which is an envelope parameter, not a market fact. That is what places this on the adverse-selection axis rather than the idea axis, and ENVELOPE-BOUND rather than UNIVERSAL.

9. Cost and capacity analysis

Fee floor alone, fee 2.0 bps per side (RT 4.0)   412 of 676 instruments net-positive
Adverse selection, control, 60 s                 -0.53 bps markout vs +0.14 half-spread
Net after adverse selection, control             markout - c_rt = -0.53 - 4.0 = -4.53 bps

Capacity was never reached as a binding question, because the mechanism failed before size became relevant. One general figure from elsewhere in the programme is worth carrying into this row's cost picture rather than treated as unrelated: measured hybrid execution on this venue, a passive wait budget followed by a cross, ran to 8.85 to 12.17 bps round trip against a quoted taker floor of 11.0, and a pure-maker fill with certainty priced to roughly zero round trip only under the unrealistic assumption of a 100 percent fill rate [E-0013, E-0014]. The quoted 4.0 bps maker round trip used in section 4 is the optimistic bound, not the expected one.

10. Envelope sensitivity and reopening conditions

Binding parameter: cancellation latency inside the adverse-selection window.

Reversal condition:  adverse_selection(tau*) < half_spread_proxy
                     for the horizon tau* at which the effect completes

The measurement places tau* at or before 10 seconds: adverse selection already exceeds the half-spread at that horizon on 5 of 6 instruments tested, and the control fails within one second, with the fraction of profitable fills falling to 35.4 percent inside the first second.

This envelope reacts on an order round trip of 5 to 6 ms and an event-driven floor of 10 to 20 ms, which sounds comfortably inside a 10-second window. It is not, once the full path is counted: market data freshness on this substrate carries a 99th-percentile tail of 2.6 seconds, and cancellation requires seeing the adverse move before acting on it. The gap between order latency and data latency, not order latency alone, is what a reader should substitute their own numbers into.

What would reopen it: a cancellation path, end to end from adverse signal to cancel acknowledgement, demonstrated to act inside the 10-second completion window on this instrument set, from a colocated or near-colocated position. That is a latency claim, testable in principle, and not one this programme is positioned to make.

What would not reopen it: a better fee tier. Gate 1 already shows most of the universe quoting above the fee floor; the fee floor was never the binding constraint after gate 2 ran.

11. Reference implementation

The spread census, BL-IS-011, and the markout instrument, BL-IS-006, are specified in Volume V. The two supporting libraries they rest on are published there as running code: the cost and spread primitive as 05_INSTRUMENTS/src/cost_model.py and the markout primitive as 05_INSTRUMENTS/src/markout.py. The sign-convention verification against the tick-direction field is preserved in the test suite rather than left as a one-time check, which is why reversing any of the four sign combinations fails loudly instead of quietly moving a verdict.

The maker engine this row's economics closed is published as 05_INSTRUMENTS/strategies/rust/src/maker_core.rs, with 29 tests: the five-step quote placement, the four-tier inventory skew, the rolling-percentile toxicity guard and the force-flatten triggers, each as a pure function. A reader on a better fee tier re-prices the spread census in section 7 against their own round trip and then has the quoting logic that would collect it.

12. Evidence register

13. Lessons extracted

This row was closed twice, and the sequence is the lesson, not either measurement alone.

The first closure, on seven instruments carrying a median spread of 1.31 bps, reached what turned out to be the correct conclusion by the wrong route. The sample had been drawn from a liquidation-event recorder, which by construction only captures instruments active enough to produce liquidations, which selects toward the most liquid and therefore the narrowest-spread names in the universe. The sample was not merely small; it was selected in the direction of the conclusion it produced.

Reopening it and measuring all 676 instruments returned a median of 5.26 bps, four times larger, and on the fee schedule alone put most of the universe on the profitable side of the maker round trip. Had the record stopped there, it would have recorded an opportunity. The second gate, adverse selection, closed the door for real: the price movement following a fill exceeds the half-spread being earned on every instrument tested, including a control chosen specifically because its answer was already known.

The instructive point for the archive is that a verdict which happens to be correct on inadequate evidence is not a verdict. Nobody would have known the first closure was reached on the wrong grounds if it had not been reopened, and it produced the second, sounder closure only because someone treated seven instruments as too small a sample to trust, regardless of which direction it pointed.

One further note belongs here because it changes how a reader should treat every number in section 7: the maker fee behind the entire gate-1 census was marked unverified at the time of measurement, absent from every config file in the repository, and reported as a grid rather than a confirmed schedule for exactly that reason. The conclusion held across the whole fee grid tested, but the grid, not a single confirmed number, is the honest artifact.

14. Cross-references

BL-P2-001 section 4, the door 3 narrative this dossier documents in full. BL-EC-003 for the adverse-selection economics in general, independent of this specific door. BL-D-008 section 6 for the cost-first doctrine's own treatment of this result. BL-D-005 for the adverse-selection axis definition. Ledger row D3.

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