BL-D-005: The Axis of Death Taxonomy
BIJECTIVE LABS
BL-D-005: The Axis of Death Taxonomy
Status: RELEASED
Classification: PUBLIC
Revision: 1
Date: 2026-09-07
1. Why a dead hypothesis needs a cause of death
"It did not work" is not a research finding. It is the absence of one.
When a hypothesis dies, the useful output is not the fact of the death but the axis it died on, because the axis determines what else the death implies. A hypothesis that died because its mechanism does not exist has closed an entire branch. A hypothesis that died because a round trip cost eleven basis points has closed nothing except its own configuration.
Without the axis, falsification spreads. A team that kills three reversion hypotheses on cost will start to believe reversion does not exist, and will stop testing a class that was never actually tested. Recording the axis is what stops a verdict from contaminating ground it never touched.
2. The axes
| Axis | The hypothesis died because |
|---|---|
| Idea | The proposed mechanism is not present in the data |
| Cost | The mechanism is present and smaller than the execution cost |
| Data | The instrument could not measure the premise |
| Cadence | The measurement was too slow to observe the phenomenon |
| Power | The sample was too small to distinguish the effect from zero |
| Condition | The result exists only in one regime or one stretch of the sample |
| Method | The test itself was constructed wrongly |
| Capital | The mechanism requires more working capital than was available |
| Venue | The mechanism is priced in an instrument that could not be reached |
| Adverse selection | Passive execution is systematically picked off faster than it earns |
| Mechanism inverted | The effect held in research and reversed sign out of sample |
Two axes carry a warning label.
Power is not a death. A hypothesis on the power axis produces NOT PROVEN and
stays open. It appears in this taxonomy because teams routinely record it as a
kill, and doing so destroys the difference between knowing something is false
and not knowing anything.
Method is a death of the test rather than of the idea. The hypothesis returns
to the queue with a corrected instrument, and the record says so.
3. The split that matters to a reader
Every settled verdict carries one of two transferability classifications on top of its axis.
UNIVERSAL. The mechanism itself failed. No change to fees, latency, cadence, capital, or venue rescues it. These verdicts carry to any operator on any infrastructure.
ENVELOPE-BOUND. The mechanism held and a declared parameter of the operating envelope blocked it. A reader running different parameters should expect a different answer.
The mapping from axis to classification is nearly mechanical:
| Axis | Classification |
|---|---|
| Idea, Condition, Method, Mechanism inverted | UNIVERSAL |
| Cost, Cadence, Capital, Venue, Adverse selection | ENVELOPE-BOUND |
| Data | Either, depending on whether a better instrument exists |
Data is the ambiguous one and it is resolved case by case. A hypothesis that
failed because the recorder sampled at one-minute resolution is envelope-bound,
since tick data exists and somebody has it. A hypothesis that failed because the
venue does not publish the field at all is universal until the venue changes.
4. Worked classifications
Three cases show the taxonomy doing real work.
A signal that measured real and lost to fees. Aggregated informed-flow bursts
produced +3.0 basis points of drift at a 30-second horizon on 486 events, with a
sign-permutation p-value of zero. Net of a 9.0 basis point taker round trip, the
result was -6.0 with an interval entirely below zero. Axis: cost. Classification:
ENVELOPE-BOUND. The reversal point is a round trip below 3.0 basis points,
which is reachable on this venue quoting passively at tier zero. Nothing about
this finding tells an operator with a maker rebate that the signal is absent.
A signal whose premise was wrong in both directions. A positioning-drift
hypothesis predicted directional movement following accumulation. Forward return
was positive following accumulation at +1.4 and also positive following release
at +6.7. The mirror test failed: the conditioning variable does not separate
outcomes, so there is no direction to trade. Axis: idea. Classification:
UNIVERSAL. No fee schedule fixes a premise that is symmetric.
A signal we could not see in time. Whale position state changes produced no
measurable drift at a 300-second horizon. The detection latency of the scanner
that observed those changes was 92.8 seconds at the median. Axis: cadence.
Classification: ENVELOPE-BOUND. The honest statement is that the observation
arrived too late to test the claim, not that the claim is false, and an operator polling faster is testing a different question than we did.
That third case is the one most often misfiled. It looks like a falsification, it reads like a falsification, and recording it as one would have quietly closed a branch nobody had actually examined.
5. A hypothesis that has not died carries no death axis
The label follows the verdict, and the record separates the two cases in its document headers.
A FALSIFIED or CLOSED row carries an axis of death. Something ended it,
and the axis names what.
An OPEN or NOT PROVEN row carries a blocking axis instead. Nothing ended
it. Something is in the way, usually sample or coverage, and the hypothesis
stays live behind it.
A CONFIRMED row carries neither, because it is a fact about a venue rather
than a hypothesis about a strategy. Nor does a VALIDATED row, for the
symmetric reason: a component that passed acceptance and is running has nothing
in its way and nothing that ended it.
This looks like a labelling detail and it is the same distinction R-06 turns
on. Writing "axis of death: power" onto a hypothesis that is merely underpowered
is how an unresolved question starts reading as a settled one, and a later
researcher inherits a closed branch that nobody actually closed.
Eleven dossiers in this record carried the wrong label until an audit caught them. None of their verdicts was wrong; the field above each verdict was contradicting it.
6. Recording the axis
The axis is assigned at the moment the verdict is issued, by the same five-angle audit that validates the verdict itself, and it is argued rather than asserted. A dossier that names an axis without justifying it against the alternatives does not pass review.
Where two axes are genuinely in play, both are recorded in order. Two verdicts in
this record carry Data, then Idea: a defect in the instrument exposed the
result first, and the hypothesis then died on its merits once the defect was
fixed. Collapsing that to a single axis would misrepresent both the failure and
the fix.
7. The distribution, and what it says
Across the thirty-seven settled verdicts in this record:
| Axis | Count |
|---|---|
| Idea | 18 |
| Cost | 7 |
| Capital or venue | 3 |
| Data, then Idea | 2 |
| Method | 2 |
| Data resolution | 1 |
| Cadence | 1 |
| Condition | 1 |
| Adverse selection | 1 |
| Mechanism inverted | 1 |
Eighteen deaths on the idea, twenty counting the two that resolved there after a data defect was fixed. The dominant failure mode over four months was not execution and not infrastructure. It was that the proposed mechanism was not present in the market.
That is worth stating plainly because it contradicts the intuition most engineering-led research teams start with, which is that better infrastructure produces better results. On this evidence, better infrastructure would have changed thirteen verdicts out of thirty-seven and left the other twenty-four exactly where they are.
Copyright (c) 2026 Bijective Labs. All rights reserved.