BL-P1-002: Taker With A Wider Edge

BIJECTIVE LABS

BIJECTIVE LABS . SEP 7, 2026

BL-P1-002: Taker With A Wider Edge

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-07

Period: 2026-05-30 to 2026-06-12 Design change: fewer trades, wider targets, an empirical gate before deployment Outcome: marginal on gross measurement, negative once costs were restored

1. The inversion

The maker era closed because win magnitude sat below the fee floor. There are only two ways out of that inequality: lower the floor, which had just been ruled out at this capital scale, or raise the magnitude.

Raising the magnitude means trading less and holding longer. The design target inverted accordingly:

Parameter	Maker era	Taker with a wider edge
Trades per day	thousands	5 to 50
Take-profit target	2 to 10 bps	50 to 100 bps
Holding horizon	seconds	hours
Execution	passive	crossing, accepted as a cost

Crossing the spread was no longer something to avoid. At a 100 basis point target, an 11 basis point round trip is 11 percent of the target rather than several times it.

2. The gate that came with it

The more consequential change was procedural. This pivot introduced the rule that no strategy deploys without passing a pre-declared empirical gate.

The gate screened 4,800 cells across signal, instrument, hour of day, and horizon. Four combinations cleared a 55 percent win rate at a 100 basis point target with at least 30 triggers each, and the best of them measured 56.4 to 59.1 percent on samples of 44 to 140.

That is a thin result and it was treated as one. It passed the gate, which was the question being asked, and the gate was a floor rather than an endorsement.

Two things about this screen are worth noting in hindsight, because both became rules later. It searched across hour of day, which is a phase parameter of exactly the kind that R-09 now requires to be reported across its full range. And it was computed on gross returns, because the cost identity had not yet been restored to the measurement path.

3. The universe was too efficient

The first deployment ran the selected combination on the declared eight-instrument set for 26 hours and produced five trades.

The instruments had been chosen for liquidity. Liquid instruments are, almost by definition, the ones where dislocations of 100 basis points are rare. The strategy had been designed for a market condition its own universe was selected to exclude.

A full scan of the venue quantified it:

Statistic	Value
Instruments enumerated	570
Passing the movement filter	192
Top-20 candidates, event frequency	1.4 times the declared set
Top-20 candidates, median move	2.5 times the declared set

The programme pivoted to a rolling universe on a 24-hour rotation, with instrument-agnostic signals so that nothing was hardcoded per instrument.

This solved the sample problem and created the one that surfaced three weeks later. A filter that ranks on movement prefers instruments that move for reasons unrelated to any tradeable mechanism, and that is the direct ancestor of the contamination documented in BL-IN-004.

4. What the rolling universe measured

Window	n	Expected value	Assessment
First, approximately 15 hours	662	+1.38 bps	4 of 6 criteria passed, marginal
Second, 23 hours	995	-1.97 bps	Not profitable as deployed

Decomposing the second window produced the useful finding. One of the three signals accounted for the entire net loss; the other two were approximately flat. Excluding it and restricting to instruments that had actually performed produced a subset at +6.33 basis points.

The recommendation was config-only: disable the failing signal, whitelist the performing instruments, re-shadow. No code change, no redesign.

That recommendation was correct in method and would have been wrong in substance, because all four of those numbers were gross.

5. The confluence detour

Frequency was still running far above the design target: 1,522 trades per day against 5 to 50. The response was a two-gate confluence requiring both an order-flow threshold and a toxicity threshold to fire together.

It reduced frequency to 438 per day and changed nothing else.

The post-mortem, written five weeks later, identified why. Both gates derived from the same underlying order flow, so requiring both was close to requiring one of them twice. And the thresholds had been chosen to land the frequency in the design range, which fits to a property of the sample rather than of the market.

That analysis produced the nine rules in BL-D-007. The detour cost roughly two weeks and bought a doctrine that governed every gate the programme built afterwards.

6. How the phase ended

It did not end with a verdict. It ended on 2026-06-07 when the measurement error was found, and every number in this chapter was revealed to be gross.

The corrected figure for the marginal pass was -5.6 basis points. The strategy had never been marginal.

BL-IN-002 is the case study. What matters for this narrative is the sequencing: the programme spent two weeks refining a strategy whose reported performance was an artifact of an absent cost term, and the refinements were individually sound. A correct process applied to an incorrect measurement produces confident, well documented, wrong decisions.

7. What survived

The phase-gated roadmap survived and became permanent. So did the rolling universe mechanism, in a corrected form with a denylist and a liquidity floor behind it.

The strategy did not. It appears in the Kill Ledger as T1 and T1.5, both closed on cost, both classified ENVELOPE-BOUND.

8. Cross-references

BL-IN-002 for the measurement error. BL-IN-004 for what the movement filter produced. BL-D-007 for the doctrine from the confluence detour. Ledger rows T1, T1.5, OPSI-H.