BL-P1-000: Programme I, Overview
BIJECTIVE LABS
BL-P1-000: Programme I, Overview
Status: RELEASED
Classification: PUBLIC
Revision: 1
Date: 2026-09-07
Period: 2026-05-04 to 2026-07-24 Venues: Bybit perpetuals, Hyperliquid perpetuals Strategy families evaluated: 22 Decision documents: 48
1. The question the programme asked
Where in this market is there an inefficiency, and can we reach it?
That framing determined everything about how the programme ran. It sent the team looking for narrow, specific dislocations: a spread that widens predictably, a liquidation cascade that overshoots, a pair that reverts, an accumulation footprint that precedes a move. Each one is a small edge, and the working assumption was that a small edge captured reliably compounds.
The assumption is defensible. It is also the assumption that the second programme was created to abandon, and Volume III explains why.
2. The arc
Six phases, each ending in a documented pivot.
| Phase | Period | What was tried | How it ended |
|---|---|---|---|
| Origin | to 2026-05-24 | Taker execution on an order-flow signal | Breakeven win rate of 122.2 percent, above unity |
| Maker era | 05-24 to 05-30 | Passive quoting, three variants | 28,119 shadow trades; win magnitude below the fee floor at every horizon |
| Taker wider edge | 05-30 to 06-12 | Fewer trades, wider targets, empirical gate first | Marginal, then negative once costs were correctly counted |
| Parallel machines | 06-12 to 07-02 | Five independent strategies on one substrate | Four retired, one retracted verdict |
| Hyperliquid track | 07-02 to 07-18 | Identity-level data on a transparent venue | Three exit mechanisms tested and closed in five days |
| State evaluation | 07-24 | Assessment across both venues | 20 of 22 families dead, 2 open |
3. The three structural pivots
Taker to maker, 2026-05-24. Twenty-five hours and 356 fills produced a
breakeven win rate of 122.2 percent. The pivot was not a judgement call; it was
arithmetic. Documented in BL-IN-001.
Maker to taker with a wider edge, 2026-05-30. After eight and a half days and 28,119 shadow trades across three variants, median win magnitude of 1.68 to 3.99 basis points sat below the fee floor at every horizon tested from one second to thirty. The response was to invert the design: fewer trades, targets of 50 to 100 basis points, and a mandatory empirical gate before any deployment. This is the pivot that introduced phase gating to the programme.
Bybit to a parallel Hyperliquid track, 2026-07-02. Not an abandonment. The reasoning was structural: on a centralised venue the programme was competing against participants with better fee tiers and colocated execution, with no visibility into queue position. A venue with deterministic on-chain ordering and public position data offered a different class of question, one where the disadvantage was not structural.
4. The turning point
On 2026-06-07 the programme discovered that realised return had been computed as the difference between entry and exit mid-price, with no fee, no spread, no slippage, and no funding anywhere in the calculation.
Every verdict issued before that date was a gross number carrying a net label. The strategy that had been recorded as marginally positive at +1.38 basis points was negative at -5.6.
This is the hinge of the entire programme. Before it, the working theory was
that signal quality was the problem. After it, the question became whether any
signal at this magnitude could clear the cost structure, and that question has a
much shorter answer. Documented in BL-IN-002.
5. What the programme built while it was failing
The strategies did not survive. The apparatus did, and it is the part that transferred to the second programme intact.
| Built | Purpose | Introduced |
|---|---|---|
| Universe constitutional gate | Enforce that declared instruments are actually streamed | 05-27 |
| Parallel shadow portfolio | Test hypotheses concurrently without disturbing a running strategy | 05-29 |
| Phase-gated roadmap with hard criteria | No deployment without a passed empirical gate | 05-30 |
| Evaluation standard, pre-registered | Net from day one, correlation-aware, null control, subperiod | 06-15 |
| Data verification gate | Inspect a field live before building any consumer of it | 06-17 |
| Universe registry with denylist | Declared, versioned instrument sets per strategy | 06-23 |
| Walk-forward harness | Parameters fitted on trailing data only | 06-25 |
| Data integrity protocol | Eleven incidents, six root causes, eight binding rules | 07-06 |
| Multi-confirmation doctrine | Orthogonality verified by measurement, not asserted | 07-05 |
| In-repository instrument requirement | A verdict is valid only from a tested instrument | 07-02 onward |
Nine of the ten were written in direct response to a failure. The tenth, the parallel shadow portfolio, was written to protect a sample.
6. Result
| Verdict | Count |
|---|---|
| FALSIFIED | 21 |
| CLOSED | 5 |
| NOT PROVEN | 3 |
| OPEN | 3 |
| CONFIRMED as fact | 3 |
| WITHDRAWN | 1 |
Two families ended the programme open and event-limited: cascade exhaustion on Bybit, and post-exhaustion reversion entered passively on Hyperliquid.
They are the same mechanism. Reversion after forced liquidation flow exhausts, entered by resting rather than crossing. The programme found it independently on two venues, from two different research lines, and could not power either one to a conclusion before the calendar ran out.
That convergence is the most interesting unresolved result in the archive.
7. Honest assessment
The first programme spent roughly three months to reach the finding that short-horizon directional signals on these venues, at this cost structure, do not clear their execution cost.
Two thirds of that time was spent before the measurement error was found. A programme that had run the cost identity on day one would have reached the same conclusion considerably faster, and that is the single largest efficiency available to anyone repeating this work.
The other observation worth recording is that no capital was lost. Twenty-two families were evaluated, thirteen were built, two ran in shadow, and the one live venue touch produced zero fills. Whatever else the programme did poorly, it did not pay for its education with the balance.
8. Cross-references
BL-P1-001 through BL-P1-004 narrate the phases. BL-P1-005 is the full
decision ledger. BL-IN-001 through BL-IN-005 are the case studies from this
period. Volume III covers the successor programme.
Copyright (c) 2026 Bijective Labs. All rights reserved.