BL-P2-002: The Gates and the Sealed Era

BIJECTIVE LABS

BIJECTIVE LABS . SEP 7, 2026

BL-P2-002: The Gates and the Sealed Era

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-07

1. Five gates, in order

Programme II ran its work through five sequential gates. Each one asks a different kind of question and each can stop the programme.

Gate	Question	Cost of failing here
G1	Does the substrate exist, complete and verifiable?	Days
G2	Can the phenomenon be detected at all, and at what threshold?	Days
G3	Does the effect exist in the research era?	One hypothesis from the budget
G4	Does each layer contribute beyond the ones before it?	Build cost
G5	Does it survive on data nobody has seen?	The seal, which is spent once

The ordering is by cost, in the same way the falsification chain is. G1 and G2 cannot produce an edge and can only produce a stop, which is why they run first.

2. G1: the substrate

Built and verified before any hypothesis was tested.

Property	Value
Rows	12,970,563
Size	314 MB
Instruments	154 perpetuals
Streams	4 per instrument
Retries during build	0

A gate that reports zero retries is worth as much as the row count. It means the substrate is not silently missing windows where the venue rate-limited the harvest, which is a common and hard-to-detect defect in this kind of build.

3. G2: can it be detected

G2 asks whether the instrument can see the phenomenon, and at what threshold. It computes no returns.

The first hypothesis needed a deleveraging event detector. At hourly resolution the proxy correlated with the target at rho = 0.067, which is a failure [E-0016]. At five-minute resolution it reached rho = 0.702 against a recomputed hourly baseline of 0.085, on 25 instruments [E-0018].

The gate passed at five minutes, and the amendment that recorded it did something worth copying: it permanently capped the proxy at WEAK.

The cap is a statement about what any downstream verdict can claim. A hypothesis resting on a proxy that correlates at 0.70 with the thing it stands for cannot be reported as a clean test of the thing itself, and writing the cap into the pre-registration meant no later reader could treat the proxy as validated.

This is the kind of discipline that costs nothing at the time and prevents an overclaim months later.

4. G3: does it exist

The first hypothesis reached G3 and died there [E-0019].

The claim was that crowded leverage precedes directional deleveraging: when positioning is one-sided, forward returns should skew against the crowd.

Measurement	Value
Skew, crowded long	+1.087
Skew, matched baseline	+0.412
Difference	+0.675, CI95 [-0.938, +2.020]
n, crowded long	904
Short mirror, difference	+1.204, CI95 [+0.108, +2.180]
n, short mirror	3,050

The sign is wrong. The prediction was a more negative skew and the measurement returned a more positive one. The mirror is worse: both tails fatten on both sides.

The correct reading is that crowding predicts a volatility expansion rather than a direction. That is a real finding, it is not tradeable directionally, and it became the seed of the volatility work in doors 5 and 6.

A hypothesis that dies at G3 and hands a usable observation to the next one is the funnel working.

5. G4, and why it was never reached as scoped

G4 was designed to ablate layers: build the full specification, then remove one layer at a time and measure what each contributes.

It was never run in that form, because no specification survived G3 and G5 with enough layers to ablate. The marginal-contribution work that did happen ran under the single-pass ladder in BL-D-007 section 5 instead.

Recording a gate that was designed and not used matters. The alternative, quietly renumbering the gates so the sequence looks complete, would misrepresent how the programme actually ran.

6. G5: the seal

The substrate was partitioned before any research began.

Era	Range	Purpose
Held-out A	Substrate floor to 2021-12-31	Sealed, never opened
Research	2022-01-01 to 2025-06-30	All development and iteration
Held-out B	2025-07-01 onward	Opened once at G5

R-11 in the programme's numbering governs the seal: opened once, against one pre-registered specification, with acceptance criteria fixed in advance, and no second attempt.

The single-shot property is the whole design. A held-out era opened twice is a research era with extra steps, because the second look is informed by the first.

7. What happened when it was opened

The hypothesis that reached G5 had a strong record. It had survived research-era validation, held under fragility conditioning, and beaten a permutation null by 5.83 sigma with three of three subperiods positive.

Four acceptance gates were fixed in advance [E-0037]:

Gate	Criterion	Result	Outcome
P1	Net interval excludes zero	+4.28, CI95 [-2.16, +10.72]	FAIL
P2	Calm-regime control not significantly positive	Control negative	PASS
P3	Liquidity ratio at or above 0.8	1.11x	PASS
P4	Both research and held-out positive	Research +8.24, held-out -8.99	FAIL

n = 826 across the combined eras.

P4 is the one that matters. The mechanism did not merely weaken out of sample. It inverted. Fragility amplification, which measured 8.8 times in the research era, reversed sign on data the programme had never seen.

The post-mortem quantifies the reversal [E-0036]. In the research era, the separation between fragile and calm gross returns was +4.17. In held-out B it was -0.32.

Axis: mechanism inverted out of sample. This is the only hypothesis in the entire archive to die on that axis, and it is the reason the seal exists.

8. What the seal cost, and what it bought

It bought the only finding in the programme that no earlier gate could have produced. Every gate before it passed. Conditioning passed, the null passed, the subperiods passed, the liquidity control passed. A programme without a sealed era would have carried this hypothesis forward as validated.

The cost follows from the design rather than from any error. The seal opens once, it was spent here, and it was spent correctly. The consequence is that the momentum family identified later in the programme has no clean held-out data remaining on this substrate to certify it against.

That is a statement about where the next measurement comes from, not a shortfall. The next hypothesis on this class needs either a different substrate or a forward shadow accumulating out-of-sample data in real time, and both are available.

9. The audit that followed

The G5 result was re-examined on 2026-08-03 against a variable nobody had controlled: the rebalance hour.

The instrument built for that audit was first required to reproduce the G5 result exactly, and it did, to the basis point on all four reported quantities. Only then was the specification re-run across all twenty-four hourly phases.

The kill held and strengthened. Zero of twenty-four phases were significant, the median was -3.54 basis points, and the phase used had ranked third of twenty-four. Any other anchor would have rejected the hypothesis more decisively.

BL-IN-006 is the case study.

10. Cross-references

BL-D-003 section 5 on the sealed era as a standard. BL-IN-006 the anchor audit. BL-P2-001 the doors this gate sequence served. Evidence E-0016, E-0018, E-0019, E-0036, E-0037.