BL-P2-003: The Knowledge Map
BIJECTIVE LABS
BL-P2-003: The Knowledge Map
Status: RELEASED
Classification: PUBLIC
Revision: 1
Date: 2026-09-07
The most directly reusable document in this volume. It says which classes of edge on these venues are closed, which are closed only for this operator, and which were never opened.
1. Why organise by class rather than by strategy
Thirty strategy families is a list. Six classes of information is a map.
A class is a kind of thing the market can tell you. Direction, magnitude, the price of risk, the cost of liquidity, speed, and identity. Every strategy either monetises one of them or it does not have a thesis.
Organising by class has a property the strategy list does not: closing a class closes every strategy that would have used it, including the ones nobody thought of. Twenty families dying individually is twenty data points. Twenty families dying on one class is a conclusion.
2. The six classes
2.1 Direction
What it claims. That some observable predicts the sign of the next move.
Status: closed, and confirmed out of sample.
Seventeen of nineteen pre-registrations across both programmes tested a directional claim. Twenty of twenty-two strategy families in the first programme did. All of them are falsified.
The confirmation is the part worth reading. The information coefficient of the directional signals, measured on the research era, ran at t = -5.10. On held-out data it ran at t = -1.42 [E-0052]. The signal weakens out of sample toward nothing, in the direction and by roughly the amount you would expect if the research-era measurement were largely fitting.
This is the single largest finding of both programmes. It is stated plainly because it cost four months to establish and it will save a reader the same four months.
Classification: UNIVERSAL, with the caveat below.
The caveat is real and is not a hedge. Direction was tested at horizons of four
hours and above, on a 154-instrument perpetual universe, using positioning,
flow, funding, and cross-sectional residual signals. It was not tested below four
hours, because R-13 closed that range on cost, and it was not tested on data
classes this substrate does not carry. A reader with sub-second execution is
asking a different question.
2.2 Magnitude
What it claims. That some observable predicts the size of the next move, without predicting its sign.
Status: real, validated out of sample twice, unmonetisable here, in production defensively.
This is the one class that survived everything thrown at it.
| Measurement | Research era | Held-out |
|---|---|---|
| Fragility to expansion, lift | t = +10.25 | t = +4.72 |
| Brier skill, regime gate | +0.01251 | +0.01053 |
Weakened out of sample, as everything does, and still there. The gate blocks 20 percent of periods and avoids 28.4 percent of volatility expansions, a lift of 1.49 times over the base rate [E-0012].
The direct monetisation is an options position, which requires a desk this envelope does not reach. Door 5a measured the premium at t = 22.0 and closed on capital; door 5b tried to invert it and was falsified.
So the class is used the only way it can be used here: to decline. It is the single component from either programme running in production.
Classification: ENVELOPE-BOUND on venue access. An operator with options
access is looking at a validated signal with no monetisation route tested.
2.3 Carry
What it claims. That funding rates paid between perpetual and spot are a harvestable risk premium.
Status: closed at class level.
Pooled across the research era, a beta-neutral carry construction returns roughly +14.2 percent annualised at a Sharpe near 1.0. That is a real result and it looks like a business.
Out of sample it is not [E-0003].
| Measurement | Research | Held-out B |
|---|---|---|
| t-statistic | +17.72 | +13.08 |
| Net return | +4.52 bps | -0.51 bps |
The t-statistic barely moves and the net return goes to zero. That combination is diagnostic: the signal is still there, and what it was compensating for has caught up with it.
The mechanical reading is that carry is inseparable from the volatility it pays for. You are not harvesting a premium, you are being paid to hold a risk, and over a long enough sample the risk arrives.
The first programme had already closed this class from a different direction, finding that thirteen of fourteen net-positive instruments had no spot market to execute against.
Classification: UNIVERSAL.
2.4 Spread
What it claims. That providing liquidity earns the spread.
Status: falsified on adverse selection.
The universe quotes wide enough. Median 5.26 basis points against a maker round trip well below it, with 62.1 percent of instruments carrying two ticks of room or more [E-0051].
The fills are the problem. Markout at 60 seconds on the control instrument runs to -0.53 basis points against a half-spread proxy of 0.14, across 9,352,722 trades [E-0002], and adverse selection exceeds the half-spread within ten seconds on five of six candidates.
The distinction matters: this is not a fee problem. It is that a resting order is filled preferentially by counterparties who know something, and that cost scales with how attractive the quote is.
Classification: ENVELOPE-BOUND on queue position and reaction speed.
2.5 Speed
What it claims. That reacting faster than other participants is itself an edge.
Status: excluded by design, not tested.
Order round trip on this envelope is 5 to 6 milliseconds on a persistent connection, from a cloud instance rather than a colocated one. Participants competing on speed operate below that by a wide margin.
The programme did not attempt a latency-sensitive strategy and lose. It declared the class out of scope at layer 0 and spent its budget elsewhere. That is the correct handling of a class where the operator is structurally disadvantaged, and it is recorded here so that a reader does not mistake absence for falsification.
Classification: not applicable. Nothing was measured.
2.6 Identity
What it claims. That knowing who is transacting changes what you should do.
Status: real, measured precisely, sub-cost.
The measurement is one of the cleanest in the archive. Adverse selection differs between large and retail aggressors by 0.906 basis points, CI95 [0.757, 0.983], permutation p = 0, on 4.51 million fills, persisting 10 to 30 seconds.
Every attempt to monetise it failed. Selective quoting returned gross negative across all instruments. A directional burst strategy measured real drift at +3.04 basis points and lost 6.0 net to a 9 basis point round trip, with the identity component contributing only +0.71 of the drift and an interval spanning zero.
The class is real and the effect is smaller than the cost of acting on it.
Classification: ENVELOPE-BOUND on cost.
3. The substrate the map was built on
| Source | Depth | Content |
|---|---|---|
| Bybit historical, 154 instruments | approximately 4.8 years | Klines, funding, open interest; anonymous |
| Hyperliquid | 2026-07-03 to 2026-07-24 | Wallet-level positions with entry and liquidation prices |
| Total | 8.20 GB |
The two sit at opposite ends of a trade between depth and richness, and neither answers the other's questions. Every finding in this map names which substrate carries it.
The identity class exists only because of the second one, and it is the only class the first substrate could not have produced at all.
4. What was never opened
Cascade exhaustion, entered passively. The two open leads from Programme I,
found independently on both venues, are the same mechanism. Neither reached the
power to settle and both are event-limited by the market's own rate of producing
cascades. Held at RESTRICTED.
Options and dealer positioning. The monetisation route for the magnitude class. Not a research question; a question about whether to acquire access.
Sub-four-hour direction. Closed by R-13 on cost, not by measurement. A
different cost structure reopens it.
5. Held-out data remaining
None on the Bybit substrate. Held-out B was opened once at G5 and is spent. Held-out A remains sealed but sits before 2022 and is not representative of the current market structure.
A new hypothesis on this substrate has three options: accept that it cannot be certified out of sample, run a forward shadow and accumulate genuinely unseen data in real time, or move to a different substrate.
The programme's own recommendation was the second, and it is a resource decision rather than a research one.
6. How to use this map
If you are deciding what to test on these venues, work from the classes rather than from strategy names.
Direction is closed and the closure is confirmed out of sample. Do not spend a budget re-testing it at horizons above four hours on this instrument set.
Magnitude is real and needs an options desk. If you have one, this is the most promising unexplored route in the record.
Carry is closed. Spread is closed unless your queue position is materially better than a non-colocated cloud instance. Identity is real and sub-cost unless your round trip is below roughly 3 basis points. Speed was never opened and is unlikely to be worth opening from a comparable position.
Copyright (c) 2026 Bijective Labs. All rights reserved.