BL-D-001: The Falsification Chain
BIJECTIVE LABS
BL-D-001: The Falsification Chain
Status: RELEASED
Classification: PUBLIC
Revision: 1
Date: 2026-09-07
1. The problem the chain solves
A quantitative research programme fails in one of two ways. It either commits capital to an idea that was never real, or it spends months proving that something is not real when a cheaper test would have said so in an afternoon.
The second failure is the more common one and the less discussed. Building a strategy is expensive. Backtesting it properly is expensive. Running it in shadow for a month is expensive in the only resource a small programme cannot buy more of, which is calendar time. A group that reaches for the expensive test first will get through four or five hypotheses a year.
The chain exists to make the cheap tests happen first, and to make them mandatory. It orders the ways an idea can die by what each test costs to run, and it forbids skipping forward.
2. The eight layers
| Layer | Test | What kills it here | Cost to run |
|---|---|---|---|
| 0 | Cost floor | Plausible signal magnitude is smaller than the cheapest execution channel | Minutes |
| 1 | Market observation | The phenomenon does not recur | Hours |
| 2 | Mechanism memo | Nobody can be named who is structurally forced to lose | Minutes |
| 3 | Falsifiable hypothesis | The claim cannot be stated so that data could refute it | Hours |
| 4 | Data requirements | No available instrument can measure the premise | Hours to days |
| 5 | Statistical validation | Interval spans zero, loses to a matched null, or lives in one subperiod | Days |
| 6 | Strategy design | The exit is not bound to the mechanism | Weeks |
| 7 | Shadow | Net expectation is negative in the live market | Weeks of calendar |
| 8 | Live capital | Capital |
Two properties matter more than the list itself.
The layers are ordered by cost, not by intellectual importance. Layer 2 is the deepest question in the chain and it costs almost nothing to ask, which is why it sits near the top.
The chain is strictly ordered. A hypothesis does not reach layer 5 because somebody is confident about it. Most of the damage this programme did to itself in its first months came from starting at layer 6.
3. Layer 0, the cost floor
Before anything else, write down two numbers. The minimum round-trip cost of the cheapest execution channel available, and the plausible magnitude of the signal being proposed.
If the second is smaller than the first, stop. There is no test to run.
This layer was added after a hypothesis reached full statistical validation, produced a real and well-powered directional drift of 3.0 basis points at a 30-second horizon with a sign-permutation p-value of zero, and then died because the taker round trip was 9.0 basis points. The measurement was correct. The answer had been available on the first day for the price of two lines of arithmetic.
Layer 0 also produces the parameter that makes a verdict transferable. Because the cost floor is written down before the test, a later reader can substitute their own and see immediately whether the conclusion survives.
4. Layer 2, the mechanism memo
One sentence, mandatory before any hypothesis is written:
Who is forced to lose, and why does the structure force them?
Not who might be wrong. Not who is less sophisticated. Who is compelled, by margin, by mandate, by settlement timing, by the mechanics of the venue, to transact at a price they would not otherwise accept.
A thesis that cannot answer this is a pattern somebody noticed. Patterns are plentiful in financial data and most of them are the residue of sampling. A forced loser is a reason for the pattern to persist after you start trading against it.
This is the layer most often skipped, and skipping it is the single largest source of wasted work in the first programme. Eighteen of the thirty-seven settled verdicts in this record died on the mechanism itself, and twenty once the two cases first exposed by a data defect are re-tested and die on the same axis. A meaningful fraction of those would have failed a serious layer-2 challenge in five minutes.
5. Layer 4, and the difference between a null and an absence
A hypothesis reaching layer 4 must show that some available instrument can actually resolve its premise. Not that data exists, but that the data has the resolution, the coverage, and the timing to see the thing being claimed.
This matters because a measurement taken with the wrong instrument returns a null result that is indistinguishable from a falsification.
Two cases in this record make the point.
A recorder captured order book snapshots at liquidation events. That is correct for characterising the event and wrong for measuring what follows it, because post-event price paths were only observable where a second event happened to follow. Coverage came out at 29 percent. The hypothesis was reported as coverage-limited rather than dead, and it remains open.
A wallet scanner observed position changes at a 93-second median latency. A downstream hypothesis measured no directional drift after those changes. The honest reading is not that the information was absent. It is that we could not observe the event early enough to know, and the verdict is recorded as bound to cadence rather than as a falsification.
Both cases are cheap to detect at layer 4 and expensive to detect at layer 5, where they arrive dressed as results.
6. Layer 5, and what validation means here
Statistical validation in this programme is not a t-test on a series of trades.
The evaluation standard in BL-D-004 sets the requirements in full. Four of them
are load-bearing.
Net from day one. Every figure is computed after fees, spread, slippage, and funding accrual where the position spans a settlement. Gross is never a verdict.
Correlation-aware inference. Returns are collapsed to a per-timestamp portfolio before any interval is computed, because mean pairwise correlation across the major perpetuals on this substrate is 0.737, which leaves an effective independent sample of 1.33 per timestamp. Per-leg t-statistics on this data are not conservative, they are wrong.
A matched null. The result must beat a control built to share everything with the strategy except the claimed edge. An early time-shuffle null in this programme manufactured an apparent 2,200 basis points of edge from nothing, which is how the requirement for a matched rather than a naive null was learned.
Subperiod stability. The sample is cut into at least three sequential blocks and the edge must survive in each. This test alone killed a momentum result that had reported 338 basis points at a p-value of 0.001 and lived entirely in the final third of its sample.
7. Layer 7, and why layer 8 stayed empty
Shadow means the full strategy running against the live market in real time, computing net expectation from the first day, and placing no orders.
It is the most expensive layer in the chain because it can only be paid for in calendar time. It is also the only layer that tests the assumptions the previous six could not: that the signal survives in real-time data with real gaps, that the fills the backtest assumed are reachable, that the strategy behaves when the feed does something unexpected.
Two hypotheses in this programme reached layer 7. Neither passed it.
Layer 8 was never reached. The single occasion when the system touched a live venue was an 18-hour passive-quoting run that produced 823 triggers and zero fills, and therefore no capital exposure at all.
That outcome is what the chain is for. It is not a record of hesitation. A funnel whose whole purpose is to reach a verdict before capital is at risk has done its job precisely when the last layer stays empty, and the useful question is not why nothing was deployed but what it cost to find that out. The answer is four months and no capital.
8. The funnel, measured
| Stage | Count |
|---|---|
| Entered the chain | 51 recorded entries |
| Settled at or before statistical validation | 37 |
| Built as running strategy code | 13 |
| Reached shadow | 2 |
| Reached live capital | 0 |
The shape of that funnel is the argument. Thirty-seven hypotheses were resolved at or before the statistical layer, which is where resolution is cheap. Thirteen were built. Two were run. None consumed capital.
A programme that inverted this ordering, building first and validating after, would have produced the same verdicts at perhaps twenty times the cost, and would have produced far fewer of them.
9. What the chain does not do
The chain does not tell you which hypotheses to propose. It filters, and a filter applied to a poor stream of ideas returns a poor result quickly rather than a poor result slowly. Layer 2 is the closest thing here to a generator, and it is still only a test.
The chain also does not protect against a mechanism that holds through every
layer and then reverses out of sample. One hypothesis in this record did exactly
that, surviving conditioning, surviving a permutation null at 5.83 sigma, and
then inverting on sealed held-out data. That failure mode is documented in
BL-IN-006 and it is the reason the held-out era discipline in BL-D-003 exists
in the form it does.
10. Cross-references
The pre-registration standard that governs layer 3 is BL-D-003. The evaluation
standard governing layer 5 is BL-D-004. The classification applied to verdicts
coming out of the chain is BL-D-005. Cost-floor practice at layer 0 is
BL-D-008. The gap between layers 5, 7, and 8 is treated in BL-D-010.
Copyright (c) 2026 Bijective Labs. All rights reserved.