BL-P1-001: The Maker Era

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-07

Period: 2026-05-24 to 2026-05-30, six days Shadow trades accumulated: 28,119 Variants tested: three, plus two quoting designs Outcome: the entire execution class closed

1. Why the programme went passive

The taker line closed on arithmetic: a required win rate of 122.2 percent against a ceiling of 100. The obvious response is to attack the term that made it impossible, which is the fee.

Taker round trip   0.12 pct     breakeven win rate  1.222
Maker round trip   0.02 pct     breakeven win rate  0.667

A 67 percent requirement against an expected 70 percent leaves margin. The signal stack was retained and repurposed as quote-skew and toxicity inputs.

That reasoning is correct as far as it goes. What it assumes, and what the next six days tested, is that a passive order will be filled on terms resembling the ones the arithmetic used.

2. Persistent quoting, and what eight hours showed

The first design quoted continuously at the top of book.

Eight hours in, position exposure reached 4,198 USD against a declared limit of 50. Two defects were responsible: no reconciliation against exchange state at startup, so the system began blind to positions it already held; and a force-flatten routine that could not see those legacy positions and therefore never acted on them.

The remedy was a full reset. Flatten everything, bootstrap risk state from the exchange rather than from local memory, and add a fifteen-minute resynchronisation loop.

The more important finding was structural. Persistent quoting at the top of book was diagnosed as negative expectation by construction on this venue, because participants running bulk order amendment hold queue priority that a straightforward quoting loop cannot contest. The strategy was not being filled because it was right. It was being filled because it was last in the queue when the price moved.

3. Trigger-based quoting, and 823 triggers

If continuous quoting loses to queue position, the alternative is to quote only when the spread is wide enough that queue position stops mattering. Hibernate until the market dislocates, then post.

Eighteen hours live. 823 triggers. Zero fills.

The measurement that explains it:

The design required a widening that the market produced 0.2 percent of the time, and in that window the quote still had to be reached before the spread closed.

This is the cleanest example in the archive of a backtest and a market answering different questions. The historical data showed the spread widening. It did not show whether a quote posted at that moment would be reached, and nobody had asked it to.

4. Three variants in parallel

With both passive designs closed, the programme built three shadow variants and ran them concurrently rather than sequentially. This is the point at which the parallel shadow portfolio became standard practice, and it was adopted for a practical reason: sequential testing at one variant per day was slower than the rate at which hypotheses were being generated.

The take-profit rate on the first variant was 35 percent against a design requirement of 60. The second measured a genuine gross edge of 0.64 basis points and lost 6.36 to costs. The third had the sign wrong: informed-flow exhaustion did not revert, and all eight instruments showed stop rates above 30 percent.

The fourth is the only one that did not close cleanly. It produced the strongest gross result in the set and turned positive only under an optimistic maker-to-maker fee assumption. It was recorded as NOT PROVEN on power, n=58, and the idea resurfaced later as a tick-exact dislocation strategy.

5. The verdict on the class

On 2026-05-30, after eight and a half days and 28,119 cumulative shadow trades, all maker-class strategies were halted.

The reasoning is one line and it generalises:

Median win magnitude of 1.68 to 3.99 basis points sits below the fee floor of 7 to 12 basis points at every horizon tested from one second to thirty.

Not at some horizons. At every horizon tested. The signal was measurable and it was the wrong size, and no amount of threshold tuning changes the size of a signal.

The distance to a better fee tier was also measured rather than assumed: qualification required 5,000,000 USD of thirty-day volume, roughly 333 times what the declared capital could generate at maximum sustainable throughput. The tier was treated as fixed.

6. What was kept

The code was not deleted. The maker engine, the quoting loop, the inventory caps, and the toxicity filter remain in the codebase, disabled by configuration, because the verdict was about economics rather than about correctness.

Three practices came out of this period and outlived it: exchange-bootstrapped risk state rather than local memory, parallel shadow variants rather than sequential testing, and the habit of measuring the market condition a design depends on before building the design.

7. Cross-references

BL-IN-001 for the arithmetic that started the pivot. BL-D-008 section 2 for the doctrine this period produced. Ledger rows VAC, DA-B, PS-A, PS-B, PS-C.

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