BL-P1-003: The Parallel Machines

BIJECTIVE LABS

BIJECTIVE LABS . SEP 7, 2026

BL-P1-003: The Parallel Machines

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-07

Period: 2026-06-12 to 2026-07-02 Strategies run concurrently: five Outcome: four retired, one verdict retracted; the evaluation apparatus built

1. The architectural question

A challenge was put to the team: is the system locked into one strategy?

The answer, on inspection, was no. Data ingestion published to a shared substrate, universes combined additively, and nothing in the hot path assumed a single consumer. A new strategy could be added as an isolated sidecar reading the same state, placing no orders, without touching anything already running.

That property turned out to be the most valuable thing the first programme built, and it was not designed for the purpose. It was a consequence of separating ingestion from strategy, and it converted the research rate from one hypothesis at a time to several.

The constraint that came with it was written down immediately: new work runs as an additive parallel shadow and never disturbs a running strategy. A machine accumulating a clean forward record is an asset that a container restart destroys.

2. What was already streaming and unused

The first use of the capability was an inventory. The system was recording open interest, funding rates, and full order book depth that no strategy consumed. Three hypotheses were built against them.

Hypothesis	Method	Result
Cross-sectional residual reversion	Walk-forward, 45 instruments	0 of 9 configurations positive; net -96 to -558 bps
Cross-sectional momentum	Walk-forward, K=5, full span	0 of 4 significant; the earlier +338 bps at p=0.001 lived in the final third
Cross-sectional funding carry	Walk-forward, 26 instruments across 287 timestamps	Net -345 bps, t=-3.10, CI95 [-566, -124]

All three rejected within days of being proposed.

The funding carry result is worth reading closely, because the decomposition matters more than the verdict. The carry itself was real and positive at +15.2 basis points. It was buried under -344.9 basis points of price movement, because the position was not beta-neutral. High-funding instruments are high-funding for a reason, and that reason moves the price.

The momentum result is the one that produced a standing rule. It had reported +338 basis points at a p-value of 0.001, which is the kind of number that ends scrutiny rather than starting it. Split into three sequential subperiods, it lived entirely in the last one. R-04 exists because of it.

3. The evaluation standard

On 2026-06-15, before any of the running machines reached a verdict, the programme pre-registered how they would be judged.

Net from day one. Correlation-aware intervals from a block bootstrap on collapsed portfolio returns. Sizing through the live risk path rather than equal weights. Subperiod stability across at least three blocks. Power declared before the test. A matched null control. No cherry-picking of instruments or windows.

Writing the standard before the results arrived is the whole point. Three machines were mid-flight and none of their outcomes were known, so no threshold could be set to accommodate a number somebody wanted.

BL-D-004 is that standard in its current form. Nothing in it has been loosened since.

4. Five machines, five outcomes

Cointegration. Pairwise mean reversion on cointegrated instruments. Backtest +180 basis points at 80 percent, walk-forward -55 to -68 at t at or below -4, live -25.5. Retired 06-25. The hedge ratio had been fitted on the window it was then traded through. Full case study in BL-IN-003.

Open-interest accumulation. Long-only, keyed on open interest building while price bases. Retired after 107 forward-shadow trades. It went long into a nine-day decline at 89 longs to one short, and a funding term in its cost model returned zero throughout.

The useful part of this machine happened before it ran. Its design assumed the venue exposed seven to fourteen days of open-interest history. A verification check pulled the field live and found approximately one hour. The redesign cost nothing because nothing had been built, and R-22 exists because of it.

Cash and carry. Delta-neutral funding harvest. Retired 06-27, and not on the signal. Of fourteen instruments carrying a net-positive carry after costs, thirteen had no spot market on the venue, so the carry was measurable and not executable. The one that was executable yielded 2.7 percent annualised against a bar of roughly 10.

This is the first verdict in the archive closed on venue access rather than on the market, and it is classified ENVELOPE-BOUND for that reason.

Accumulation and distribution. A footprint strategy built on order-flow divergence, open-interest build, funding crowding, and book absorption, gated by regime and symmetric long and short. The long arm returned -27.1 basis points on 152 trades with an interval of [-59.1, +2.3] and a win rate of 34.2 percent. The short arm reached five trades and was recorded as inconclusive.

Dislocation. Tick-exact fade of forced-flow displacement. Rejected, then retracted. BL-IN-004.

5. What the retraction changed

The dislocation retraction happened on 2026-06-22 and it altered how the programme treated its own conclusions.

The verdict being challenged was a rejection. Nobody had an interest in defending it, which is precisely what makes the episode useful: the discipline was applied to a conclusion nobody was attached to, and it still found an error large enough to invalidate it.

Two artifacts followed within a day. A canonical data substrate manifest, with a standing instruction to measure venue conditions rather than assume them. And a generic per-strategy universe registry, with declared instrument sets, a liquidity floor, and a hardcoded denylist behind the automated filter.

The second one closed a defect that had been live since the rolling-universe pivot three weeks earlier.

6. What this period cost and returned

Five strategies built, five closed or retracted, no capital committed.

Against that: the evaluation standard, the data verification gate, the universe registry, the walk-forward harness, and the non-disturbance invariant. Every one of them was carried into the second programme unchanged, and four of the five were written in direct response to something that had just gone wrong.

The programme's research rate roughly tripled in the month that followed. That is the return on this period, and it did not come from any of the five machines.

7. Cross-references

BL-IN-003 cointegration. BL-IN-004 the retraction. BL-D-004 the evaluation standard. BL-D-006 section 4 on the verification gate. Ledger rows XSEC-REV, XSEC-MOM, XSEC-CAR, COINT, OI-ACC, CARRY, ACCDIST, DISLOC, LEADLAG.