BL-D-007: The Multi-Confirmation Doctrine

BIJECTIVE LABS

BIJECTIVE LABS . SEP 7, 2026

BL-D-007: The Multi-Confirmation Doctrine

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-07

1. The failure this doctrine was written for

A strategy was producing 1,522 trades per day against a design target of five to fifty, and losing on aggregate. The proposed fix was a confluence gate: require two conditions to fire together rather than one, on the reasoning that fewer, higher-conviction entries would carry more edge each.

The gate reduced frequency to 438 trades per day. It did not fix the strategy.

The post-mortem found two defects that had nothing to do with the idea of requiring confirmation, and everything to do with how this particular confirmation was built.

The two gates were not independent. Both were derived from the same order flow. One measured imbalance in that flow and the other measured its toxicity. Requiring both is close to requiring the first one twice, which reduces frequency without adding information.

The threshold was tuned to a frequency target. The gate parameters were chosen because they produced a trade count near the design range. That is fitting to a property of the sample rather than to a property of the market, and it is the definition of a curve fit.

Everything below follows from those two findings.

2. The nine rules

ID	Rule
1	Gates must be orthogonal by construction: derived from different underlying observables
2	Orthogonality is verified by measurement, not asserted. Co-firing is tested before any profit calculation
3	Each gate is binary and mechanical, with at most one free parameter
4	Each gate carries its own causal thesis: why this condition, mechanically, changes the outcome
5	Gates are added in a single-pass marginal-contribution ladder, never searched combinatorially
6	Trade frequency is an output of the specification, never an input to it
7	Every configuration considered is recorded, including the ones not taken
8	Each gate raises the required sample: an effective sample of at least 100 per gate
9	The exit is bound to the mechanism, not to a fixed distance

3. Orthogonality by measurement

Rule 2 is the one that does the work, and it is cheap to run.

Before computing any profit and loss, measure the association between gate firings across the sample. The statistic is the phi coefficient on the co-fire contingency table, which for binary gates is the Matthews correlation applied to the firing pattern rather than to outcomes.

phi(A, B)  computed on the 2x2 table of {A fires, A does not} x {B fires, B does not}

A pair above roughly 0.3 is redundant. Requiring both adds a frequency reduction and nothing else.

This test was run on a four-gate candidate in the second half of the first programme. Three of the four gates were mutually redundant above threshold. The specification was reduced to a three-gate ladder built from genuinely distinct observables, and the reduction happened before a single return was computed.

The cost of that test is minutes. The cost of not running it, on the earlier strategy, was a redesign that reduced trade count by a factor of 3.5 and changed nothing else.

4. Frequency is an output

Rule 6 is a discipline about what you are allowed to look at while choosing parameters.

A strategy specification implies a trade frequency. The specification comes first, the frequency follows, and if the resulting frequency is unusable then the specification is wrong at the level of the mechanism rather than at the level of the threshold.

Turning that around, choosing thresholds so that frequency lands in a desired range, is fitting to the sample. It will always succeed, because frequency is monotone in threshold, and it will teach you nothing.

The tell is easy to spot in retrospect and hard to spot at the time: a threshold with an oddly specific value and no mechanical justification. Rule 3 exists to make that tell visible by limiting each gate to one free parameter and requiring it to be mechanical.

5. The marginal-contribution ladder

Rule 5 forbids searching gate combinations. The permitted procedure is a single pass.

Start with the base signal and measure it. Add the gate with the strongest independent causal thesis and measure the marginal change. Add the next and measure again. Stop when a gate does not pay for the sample it costs.

One pass, one ordering, decided in advance and recorded.

Combinatorial search over gates is a multiple-testing problem that presents itself as engineering. With five candidate gates there are thirty-one non-empty subsets, and the best of thirty-one looks impressive whether or not anything is there. Rule 7 requires that every configuration considered is written down, which makes the count visible and makes rule 5 enforceable.

6. The sample cost of complexity

Rule 8 attaches a price to each gate.

Every gate reduces the number of events that reach the measurement. A four-gate specification on a substrate producing forty events per week does not have a statistical problem, it has an arithmetic one. The requirement of an effective sample of at least 100 per gate makes that arithmetic explicit before the specification is built rather than after the results come back underpowered.

The word effective is load-bearing. On a substrate where mean pairwise correlation runs at 0.737, one hundred events across fifteen instruments is not one hundred observations. BL-D-011 section 5 sets the collapse.

7. Mechanism-bound exits

Rule 9 governs the other end of the trade, and it is the rule the programme violated most consistently.

An exit at a fixed distance, whether a take-profit, a stop, or a timeout, is a free parameter attached to the part of the strategy that determines most of its return. If the entry thesis is that a forced participant is being unwound, the exit belongs where the unwind ends, not at a round number of basis points.

The evidence for this rule is unusually direct. A reversion hypothesis was retested with a causal volatility stop at four levels. Every level made the result worse than no stop at all, and two were significant after correction [E-0023]. At the tightest level the stop fired on 38.3 percent of trades and was correct on 82.1 percent of those firings, saving an average of 149 basis points each time. The 17.9 percent where it was wrong cost 862 basis points each, an asymmetry of 5.8 times, which is more than enough to erase the gains from being right four times out of five.

Widening the stop did not help. At three sigma it fired on only 7.3 percent of trades and made the result worse still, because the value it was rescuing at that distance was itself negative. Exiting at the point furthest from entry is exiting at the point closest to reversal.

A stop is a mechanism claim. It says the thesis is invalidated at this distance. If the thesis has no such distance, the stop is cutting the strategy rather than protecting it.

8. What the doctrine does not claim

Multi-confirmation is not a source of edge. Every rule here is a constraint on how confirmation may be constructed, and none of them make a weak signal strong.

The strategy that produced this doctrine died anyway. Two hypotheses built under it also died, on the mechanism axis, with the gates working exactly as specified.

The doctrine's value is that those deaths were clean. When a specification built under these rules fails, the failure is attributable to the thesis rather than to the construction, and the branch closes for a reason that will still be true next year.