BL-SD-009: High-Conviction Two-Gate Confluence

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-07

Venue: Bybit perpetuals Programme: I Verdict: FALSIFIED Axis of death: Method Transferability: UNIVERSAL

1. Identification

This dossier carries an irregular gap between pre-registration and verdict. The deployment memo of 2026-06-02 stated acceptance criteria for a 24 to 72 hour shadow window on live Tokyo infrastructure. The record does not show that window completing to term; the only accumulated evidence is a 23-minute pilot read. The verdict that closes this hypothesis was reached five weeks later, not by that window's accumulated statistics, but by a retrospective audit of the specification itself, published as the multi-confirmation doctrine and reproduced in BL-D-007. Both facts are stated plainly because the axis of death, Method, only makes sense once the reader understands that this hypothesis was not out-argued by more data. It was out-argued by an audit of how its own two gates were built.

2. Economic thesis

An order-flow imbalance burst that coincides with elevated trade toxicity marks a moment where an aggressive, informed participant is sweeping the book faster than the passive side can reprice. Requiring both signals to fire together, rather than either alone, was intended to raise conviction that a detected event is a genuine informed sweep and not an isolated noise print, trading fewer times per day at a larger edge per trade in place of the prior instrument's high-frequency, low-conviction posture.

Who is forced to lose. The resting order on the side being swept. A maker quoting opposite an informed aggressor is filled at a price that is about to move away from it, and the fill has already happened by the time that becomes visible. This is the same forced counterparty named in the predecessor instrument, BL-SD-007 and BL-SD-008; the confluence redesign changed the selection filter around that counterparty, not the counterparty itself.

3. Falsifiable claim

Reproduced from the 2026-06-02 deployment memo, the acceptance window this hypothesis was pre-registered against:

Over a 24 to 72 hour Tokyo shadow window, high-conviction confluence (order-flow imbalance at or above 4 standard deviations, jointly with trade toxicity at or above 0.85) sustains a win rate at or above 50 percent, a mean expected value at or above +5 basis points per trade, and a trade frequency at or below 500 per day, with no single symbol contributing more than 20 percent of volume and no directional bias between long and short.

A second, later decision rule supersedes the first and is the one that actually settled this hypothesis, stated in BL-D-007 section 1 and reproduced here:

Two gates drawn from the same underlying order-flow observable do not constitute independent confirmation. Where a confluence specification's threshold was chosen to land trade frequency within a target range rather than derived from a stated mechanism, the specification is invalid regardless of its measured profitability.

This second rule is a methodological falsification criterion rather than a statistical one, and it is why the verdict below is reached without a completed accumulation of the first window's acceptance data.

4. Mathematical specification

Notation per BL-D-011.
 
Gates, both computed from the same trailing order-flow tape at t:
 
  ofi_z_t   =  order-flow imbalance across book levels 1 to 5, normalised,
               resolved from venue state at or before t
 
  vpin_t    =  volume-synchronised probability of informed trading,
               resolved from venue state at or before t
 
Confluence signal:
 
  g_t       =  1{ |ofi_z_t| >= theta_ofi }  AND  1{ vpin_t >= theta_vpin }
 
  theta_ofi   = 4.0
  theta_vpin  = 0.85
 
Entry, fire at t when g_t = 1 and all of:
 
  cooldown elapsed  >= 1,800 s since the last entry on this symbol
  open_positions(t) < 3
 
  direction  d_t = sign(ofi_z_t)
  order type: Market IOC (taker)
 
Exit, path-dependent, first touch wins, unchanged from the underlying
instrument's bracket (BL-SD-008 section 4):
 
  tp_price  =  entry_mid + d_t * entry_mid * 100 / 10000
  sl_price  =  entry_mid - d_t * entry_mid * 50  / 10000
  timeout: 7,200 s
 
Sizing:
 
  notional_usd = 5.0
 
Cost, per outcome branch, inherited from the same underlying instrument:
 
  c_rt(tp_hit)              =  6.0 (taker entry) + 1.0 (maker exit) = 7.0 bps
  c_rt(sl_hit or timeout)   =  6.0 (taker) + 6.0 (taker)            = 12.0 bps

Causality holds for the entry decision by construction: ofi_z_t and vpin_t are both computed from trailing venue state at or before t, the cooldown check reads only the timestamp of the prior exit on that symbol, and open_positions(t) counts only positions already opened. Every term entering g_t is measurable with respect to F_t.

Causality does not extend to the choice of theta_ofi and theta_vpin themselves, and this is the defect the verdict turns on. BL-D-011 section 3 states that a parameter is part of the signal and must be measurable with respect to F_t in the same sense the signal is. These two thresholds were instead selected, per the deployment memo's own account, by observing which values produced a trade count near the target range on the accumulated Opsi D sample, then holding them fixed. The parameter was fit to a property of the sample that generated it, which is the condition BL-D-007 rule 6 prohibits.

5. Operating envelope

6. Instrument

The taker wider-edge strategy implementation, unchanged in code from BL-SD-008. The confluence behaviour required no code change: the existing signal-firing logic already applies its threshold checks as a conjunction, so setting both the order-flow-imbalance threshold and the toxicity minimum on a single signal specification in the strategy registry configuration produces an AND gate without a new code path. Three prior signal specifications, ofi_3sigma, vpin_extreme, and spread_spike, the same three carried over from BL-SD-008, were disabled rather than removed.

The absence of a prospective orthogonality test is itself part of the finding. BL-D-007 rule 2 exists precisely because this hypothesis, and its predecessors in the same family, deployed a multi-signal specification without first measuring whether the signals carried independent information.

7. Results

The frequency reduction from 1,522 to 438 trades per day is the headline result quoted in BL-D-007's own account of this hypothesis, and it is the number that motivated the design in the first place. It is real. It is also still an order of magnitude above the design target it was built to hit, and the specification's threshold had been tuned specifically to approach that target rather than derived from a mechanism, which is the fact that resolves the verdict below independent of whatever the completed accumulation would have shown.

8. Verdict, axis of death, and transferability

Verdict: FALSIFIED. Axis: Method. Transferability: UNIVERSAL.

This hypothesis is not falsified by an interval that fails to clear a threshold. It is falsified because the specification that produced its results cannot support the inference the acceptance criteria would have drawn from them, regardless of what those results turned out to be. Two findings drive this, both stated in BL-D-007 section 1 and both applicable here without qualification.

The two gates are not independent. Order-flow imbalance and trade toxicity are both transformations of the same underlying trade tape; a burst of one-sided aggressive flow moves both simultaneously and mechanically. Requiring both to fire is close to requiring one of them twice at a tighter effective threshold. The observed frequency reduction, 1,522 to 438 trades per day, is consistent with exactly that: a stricter single-variable filter, not two variables adding information.

The threshold was tuned to a frequency target. theta_ofi = 4.0 and theta_vpin = 0.85 were the values that, on the accumulated Opsi D sample, produced a trade count approaching the design range. That is fitting a parameter to a property of the sample that generated it. A specification selected this way cannot be trusted to generalise, because its defining constants were never derived from anything outside the data used to choose them.

Applying BL-D-002 section 4: the question of whether this hypothesis would show a positive expectation under an arbitrarily favourable cost or capital envelope does not resolve the finding, because the defect is prior to that question. A specification whose thresholds were curve-fit to the sample cannot be rescued by any envelope substitution, which places this verdict at UNIVERSAL rather than ENVELOPE-BOUND. No cost tier, no capital scale, and no execution channel repairs a threshold that was never derived from a mechanism in the first place.

9. Cost and capacity analysis

Predecessor frequency (Opsi D)        1,522 trades/day
Confluence frequency (Opsi H, pilot)    438 trades/day, projected
Target frequency (2026-05-30 pivot)     5 to 50 trades/day
Overshoot factor                        approximately 9x to 88x

Cost is not the axis this hypothesis died on, and the branch-conditional cost model in section 4 is reported for completeness rather than because it drove the verdict. Capacity was not measured; frequency at 438 per day was well within the venue's and the infrastructure's throughput, and the constraint that closed this hypothesis was specification validity, not volume.

10. Envelope sensitivity and reopening conditions

No binding envelope parameter exists to name, because the finding is a defect in how the specification's own free parameters were chosen, not a magnitude shortfall against cost, capital, or latency.

Reversal condition:  none computable.

No cost substitution changes whether theta_ofi and theta_vpin were derived from a mechanism or from a frequency target; that fact does not depend on the operating envelope at all. This is the same reasoning applied in BL-SD-013 and BL-SD-012 for their own sign-reversal findings, extended here to a construction-validity finding rather than a magnitude finding.

What would reopen it: a redesign satisfying BL-D-007's nine rules in full, specifically an orthogonality measurement between candidate gates completed before any profit figure is examined, gates drawn from genuinely distinct observables rather than transformations of the same trade tape, and thresholds derived from a stated per-gate causal thesis rather than from a target frequency. This is precisely the successor doctrine's own flagship design, built on Hyperliquid rather than this venue, and is out of scope for this dossier.

What would not reopen it: accumulating the originally planned 24 to 72 hour window on the unmodified specification. Additional data cannot repair a parameter that was fit to a different sample; it would only extend the same fit to a longer one.

11. Reference implementation

The taker wider-edge strategy implementation, unchanged from BL-SD-008 and published with it as 05_INSTRUMENTS/strategies/rust/src/taker_wider_edge.rs, with 15 tests. The confluence configuration itself lives entirely in the strategy registry configuration and carries no code of its own; it is documented here as a configuration artifact rather than as a separate instrument, and nothing was written for this row because there was nothing to write.

That is worth stating rather than leaving as an absence. This row produced 438 trades a day against a design target of 5 to 50, and it did so through a configuration change to code that was already correct. A verdict reached by misconfiguring a working instrument is a different kind of finding from one reached by a defective one, and the published module is how a reader can tell which of the two this was. The exit design in 05_INSTRUMENTS/strategies/python/src/oi_accumulation.py carries the lesson forward: it has no fixed basis-point stop anywhere, and its module docstring names this row as the reason.

12. Evidence register

13. Lessons extracted

This hypothesis produced BL-D-007 in full, the multi-confirmation doctrine's nine rules. It is the single largest lesson-to-hypothesis ratio in this record: one pilot of seven trades, never completed to its own acceptance window, generated the construction rules that every later multi-confirmation specification in this programme, on either venue, was required to satisfy.

The specific defects it produced, restated as rules: BL-D-007 rule 1 and 2, orthogonality by construction and by measurement, came directly from the finding that order-flow imbalance and trade toxicity are not independent observables. BL-D-007 rule 6, frequency as an output rather than a target, came directly from the finding that theta_ofi and theta_vpin were chosen to hit a frequency band. Both are cited in BL-D-009 by reference through BL-D-007 rather than carrying their own numbered entry in the rules register, because they govern the construction of a specification rather than the inference drawn from one.

14. Cross-references

BL-SD-008 for the predecessor this design responded to, and for its own section 14 note naming this hypothesis as the successor its toxic-signal finding motivated. BL-D-007 in full, the doctrine this hypothesis produced. BL-D-002 section 4 for the envelope-substitution test applied in section 8. BL-D-011 section 3 for the parameter-causality constraint this hypothesis's thresholds violated. Ledger row OPSI-H.

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