BL-IN-004: The Contaminated Universe
BIJECTIVE LABS
BL-IN-004: The Contaminated Universe
Status: RELEASED
Classification: PUBLIC
Revision: 1
Date: 2026-09-07
Incident date: 2026-06-22 Class: K4, unchecked sample concentration Rule produced: R-19, and the principle that universe hygiene is signal quality
1. What was believed
Earlier the same day, a dislocation strategy had been marked REJECTED.
The verdict looked well founded. It rested on approximately 39,250 events, which by any normal standard is a rich sample, comfortably clearing the internal gate of at least 200 events per bucket. The container was healthy. The statistical process was correct.
2. The trigger
A demand to verify sample cleanliness and system normality before trusting any verdict, including one that had already been issued.
The verdict being challenged was a rejection. There was no result to protect and no incentive to look. That is what makes this incident useful: the discipline was applied to a conclusion nobody was attached to.
3. The investigation
System health first. The shadow container was healthy, zero errors, 307,000 loops, zero bad prices, deduplication working, and gross reversion measured at +6.4 basis points with the correct sign.
Then decompose the 39,250-event sample by instrument.
4. The finding
| Quantity | Value |
|---|---|
| Total events | 39,250 |
| Events from a single instrument | 34,217, or 87 percent |
| Genuinely clean events remaining | approximately 5,034 |
| Truly liquid events remaining | approximately 1,263 |
The dominant instrument was a low-grade microcap already on the programme's own forbidden list.
The root cause sits in universe selection rather than in the detector. The dislocation universe was drawn from streamed instruments selected on 24-hour volume and a count of 100 basis point moves. That rule actively prefers pump microcaps, because a pump microcap spikes constantly and therefore scores highly on move count.
The selector was not malfunctioning. It was doing exactly what it was written to do, and what it was written to do was select the instruments a live strategy would refuse to trade.
5. What the corrected verdict said
Re-run on the clean subset and on the liquid-only subset, with a live-exact cost model, the picture changes shape.
The edge is net positive only at zero slippage, which requires a zero-latency fill and is physically unavailable. At any realistic slippage it is strongly negative.
But the per-bucket sample collapses:
| Subset | Per-bucket n |
|---|---|
| Clean | 393, 129, 46 |
| Liquid only | 104, 17, 3 |
Against a gate of at least 200 per bucket, most of those buckets do not qualify to carry a conclusion in either direction.
The corrected verdict is therefore inconclusive on clean data. Not a clean rejection, and not a validation. The original rejection was recorded as invalid and retracted.
6. Why this is the harder kind of error
The rejection was directionally the same as the corrected finding: neither says the strategy works. It would have been easy to leave it standing.
Leaving it standing would have cost something specific. A REJECTED verdict
closes a branch. It tells the next researcher the question is settled, it
discourages re-testing, and it gets cited later as established. The honest
answer, that the clean sample was too small to conclude anything, keeps the
branch open and names what would be needed to close it.
That is the difference between a verdict and a result, and it is why FALSIFIED
and NOT PROVEN are separate terminal states in this record.
The measured facts also survived the retraction and are worth keeping: forced-flow reversion is real, it is monotone in liquidity, and it is latency-sensitive to the point that break-even requires entry slippage below roughly seven percent of the displacement being captured. The thesis was not dead. It was drowned in junk and constrained by speed.
7. What it cost
No capital. This was shadow and backtest analysis throughout.
The cost was a verdict that would have been wrong in the record, and the engineering hours spent producing a measurement on a sample that was 87 percent noise.
8. The fix
The document recommends rather than executes: exclude low-grade instruments from the dislocation universe at the source, so that the strategy logs clean liquid events going forward, then accumulate to a genuine clean-liquid sample of at least 200 per bucket.
A generic per-strategy universe registry followed, with declared instrument sets, a liquidity floor, and a denylist behind the automated filter.
9. The rules
Five durable lessons were recorded. All five are quoted in substance below, because each one is a distinct claim.
Universe hygiene is signal quality. Every strategy must exclude low-grade instruments at the source, not downstream.
Decompose by instrument and regime before any verdict. A clean statistical process on dirty data gives a confident wrong answer.
Forced-flow reversion is real and liquidity-monotone. The thesis is not dead; it is drowned by junk and sensitive to latency.
Entry slippage is the binding constraint for any speed edge. Break-even requires entry slippage below roughly seven percent of the displacement.
A healthy container is not a valid edge. Verify trade-production rate and sample composition, not just liveness.
The second became R-19. The fifth generalises into the data-integrity
protocol's distinction between a system that is running and a system that is
measuring.
10. What the source does not settle
The record does not state who issued the original rejection or the interval between issue and retraction; both appear on the same day without timestamps. It also frames the universe fix as a recommendation, and does not confirm within that document whether the selector was patched as a direct follow-up.
11. Cross-references
BL-D-006 class K4. BL-D-009 rule R-19. BL-D-004 section 5 on verdict
vocabulary. Ledger row DISLOC, status WITHDRAWN.
Copyright (c) 2026 Bijective Labs. All rights reserved.