BL-P2-004: Programme II, Decision Ledger

BIJECTIVE LABS

BIJECTIVE LABS . SEP 7, 2026

BL-P2-004: Programme II, Decision Ledger

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-07

Forty-nine dated entries across nine days, 2026-07-26 to 2026-08-03, drawn from the 10 pre-registration documents of the programme and the 28 verdict reports they produced.

How to read this ledger

Programme II ran for nine days and resolved eight hypotheses, six economic mechanisms, and one systematic defect touching nineteen prior pre-registrations.

The compression is the point. Programme I took three months to close 22 families, and most of that time went into building the apparatus. Programme II inherited the apparatus and spent its time on measurement.

Quota marks entries that consumed one of the twelve declared hypothesis slots. Eight of the forty-nine entries did. The rest are characterisation, cost-floor checks, audits, and closures that cost nothing from the budget, which is the intended shape.

1. Opening

Date	Kind	Entry
07-26	Charter	Programme opens; research contract and decision framework declared
07-26	Protocol	Framework Alpha: eight-layer chain, five-angle audit, axis of death, dead is not underpowered
07-26	Protocol	Hypothesis budget capped at 12; held-out eras sealed
07-26	Position	Layered conditioning rejected as a starting point: conditioning multiplies edge, it cannot create it

The opening position was itself revised two days later, and the revision is worth recording. The original argument was that if unconditional expectation is zero, no conditioning produces a positive one. That is true arithmetically and misleading in practice, because a zero unconditional expectation can mean uniformly no edge or it can mean offsetting conditional edges. The programme corrected its own framing before testing anything on it.

2. Substrate and the first hypothesis

Date	Kind	Entry	Quota
07-27	Gate	G1 substrate built: 12,970,563 rows, 154 instruments, 4 streams, 0 retries
07-27	Gate	G2 characterisation: event configuration selected, no returns computed
07-27	Pre-registration	H-P1, crowded leverage precedes directional deleveraging	Quota
07-27a	Amendment	Detection threshold revised
07-27b	Amendment	Proxy validated at 5-minute resolution and permanently capped at WEAK
07-28	Verdict	H-P1 falsified at G3: skew difference +0.675, CI95 [-0.938, +2.020], sign opposite to prediction; short mirror fattens both tails

The amendment on 07-27b is the entry to read. It passed the gate and simultaneously limited what any downstream verdict could claim from it, by writing the cap into the pre-registration where a later reader would have to encounter it.

3. The doors, measured

Date	Kind	Entry	Quota
07-28	Cost floor	Door 5: variance risk premium measured, BTC +8.10 volatility points at t=22.0
07-28	Simulation	Door 5: hedged overlay +1.51 pct at t=2.43; unhedged +0.40 at t=0.25
07-28	Calibration	Door 6: regime gate trained and evaluated, n=2,356,200
07-28	Characterisation	Information decay map by horizon; no returns computed
07-28	Pre-registration	H-V1, fragility-conditioned volatility premium	Quota
07-28	Verdict	Small-capital assessment: no mechanism satisfies edge, cost and capital simultaneously
07-28	Position	Alpha redefined: a configuration of condition, signal and execution with positive net expectation, verified out of sample, fitting our capital

Four doors were measured in a single day, and none of them consumed a hypothesis slot. Cost-floor work at layer 0 is deliberately outside the budget, because charging the budget for arithmetic would discourage the cheapest test in the programme.

4. Conditioned reversion, and the retraction

Date	Kind	Entry	Quota
07-29	Pre-registration	H-R1, conditioned reversion at 24 hours	Quota
07-30	Verdict	H-R1 closed on cost: mean -8.73 at a 22 bps round trip, CI95 [-27.00, +9.60], n=11,252 collapsed
07-30	Diagnostic	Trade-level concentration check: top-ten instrument share 0.0799
07-30	Retraction	First closure of the conditioned-reversion branch withdrawn as invalid
07-30	Rule	R-10: a kill computed only on unconditional statistics is invalid
07-30	Rule	R-11: an information coefficient may characterise, never issue a verdict
07-30	Pre-registration	H-R2, the same reversion with a causal stop	Quota
07-30	Verdict	H-R2 falsified: every stop level worse than no stop; two significant after correction
07-30	Closure	Conditioned-reversion branch closed from four independent directions
07-30	Diagnostic	Exit geometry: 0 of the tested bracket cells beat the no-bracket baseline
07-30	Audit	Independent audit of the programme to date

The retraction is the most instructive entry in this volume.

The branch had been closed on unconditional statistics while a conditioned variant, named in the hypothesis's own economic premise, had never been tested. The team withdrew its own closure, tested the variant, and then closed the branch again with evidence from four directions: the base result at realistic cost, the gross result at zero cost on the strongest conditioning cells, the stopped variant across four stop levels, and a look-ahead upper bound proven causally unreachable.

R-10 was written the same day so that the failure mode is structural rather than remembered.

The independent audit found the programme's verdicts held and simultaneously found a volatility figure used to close two doors that was too small by a factor of 2.6, a sizing convention that flipped a sign between conventions, and a risk-gate lift overstated at 1.52 against a corrected 1.49. All six mandated corrections were executed within two days.

5. Cross-sectional work and the cost reality

Date	Kind	Entry	Quota
08-01	Pre-registration	H-P2, positioning drift	Quota
08-01	Verdict	H-P2 falsified: mirror test failed, forward return positive on both sides
08-01	Pre-registration	H-X1, cross-sectional residual reversion	Quota
08-01	Verdict	H-X1: net +1.66, CI95 [-3.17, +6.49]; later withdrawn on procedure
08-01	Pre-registration	H-X2, the same conditioned on fragility	Quota
08-01	Verdict	H-X2 not proven on power; mechanism confirmed; null beaten at 5.83 sigma
08-01	Re-measurement	Door 3 spread census corrected: 676 instruments, median 5.26, replacing a 7-instrument estimate
08-01	Verdict	Door 3 closed on adverse selection: markout exceeds half-spread within 10 s on 5 of 6 candidates
08-01	Verdict	Execution cost: taker 15.02 bps measured, hybrid 8.85 to 12.17, against 5.50 required

The execution verdict carries a finding that closed a lever the programme had expected to use. Extending the passive wait budget makes total cost worse, not better, because unfilled names are the ones moving away: measured drift on unfilled orders rose from +11.7 basis points at 60 seconds to +102.2 at one hour.

6. The seal

Date	Kind	Entry	Quota
08-02	Closure	H-X3: twelve weighting variants tested, none beat the pre-registered specification
08-02	Characterisation	Information channel map across nine, then seventeen cells
08-02	Pre-registration	H-X4, the fragility family with weighted legs, for the held-out gate	Quota
08-02	Verdict	G5 opened once. Gates 1 and 4 failed; research +8.24 against held-out -8.99. Family dead on mechanism inversion
08-02	Post-mortem	Research-era separation +4.17 against held-out -0.32

7. The closing audit

Date	Kind	Entry
08-03	Diagnostic	A 39.75 bps per week discrepancy bisected to a single variable: the rebalance hour
08-03	Validation	New hour-resolution loader reproduces the G5 result exactly on all four quantities
08-03	Audit	Rebalance anchor uncontrolled across 19 pre-registrations. G5 spec across 24 phases: 0 significant, median -3.54, phase used ranked 3 of 24
08-03	Correction	Momentum headline was the best of 24 undisclosed phases; median +8.72 at t approximately 1.4; annualised revised from 7.3 to about 4.5 pct
08-03	Rule	R-09: periodic rebalance measurements must report anchor sensitivity across phases
08-03	Verdict	Carry closed at class level: research t=+17.72 falling to +13.08 held out; net +4.52 to -0.51
08-03	Map	Substrate and knowledge map: six classes, status of each

8. What the nine days produced

Output	Count
Hypotheses tested	8 of a 12 budget
Economic doors measured to a conclusion	6
Numbered rules added	R-09, R-10, R-11
Verdicts retracted by their own authors	1
Published figures revised downward	1
Capital committed	none

Four budget slots were left unused. The programme closed with them rather than spending them, on the grounds that a hypothesis proposed to use up a budget is not a hypothesis.