BL-P1-005: Programme I, Decision Ledger

BIJECTIVE LABS

BIJECTIVE LABS . SEP 7, 2026

BL-P1-005: Programme I, Decision Ledger

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-07

Fifty-six dated entries, 2026-05-24 to 2026-07-24, drawn from the 48 decision documents of the period together with the verdict reports they reference.

How to read this ledger

Every structural decision in the first programme carries a dated document. The convention was adopted on the first pivot and held for the duration: a major change requires a record stating rationale, alternatives considered, and acceptance criteria.

This ledger is the index to those records. The Kill Ledger in BL-SD-000 records what each hypothesis measured. This one records what the programme decided and why.

The Kind column separates four things that are easy to conflate. A pivot changes direction. A protocol changes how work is done. A pre-registration locks a hypothesis before data. A verdict settles one.

1. May: origin, and the maker era

Date	Kind	Decision	Trigger
05-24	Pivot	Taker to maker execution	Breakeven win rate 122.2 pct on 356 fills, above unity
05-24	Fix	Signed net position tracking; signal age 500 to 200 ms	One instrument accumulated 308.76 USD unhedged through a reset-to-zero defect
05-24	Investigation	Instrument a mark-price anomaly rather than patch it	A single fill sized 2.18 times intent; implied price less than half the real one
05-24	Pivot	Start the maker engine immediately, decoupled from a planned refactor	The refactor had been claimed as a prerequisite; no empirical bottleneck supported the claim
05-24	Design	Resolve six implementation gaps before writing code	Pre-implementation review found unresolved design holes
05-24	Protocol	Lock six strategic parameters; adopt a staged launch protocol	Post-review consolidation
05-25	Design	Build a ten-account variant testing platform	Sub-accounts reframed from throughput multiplication to capital-isolated hypothesis testing
05-26	Reset	Clean-slate flatten; exchange-bootstrapped risk state; replace persistent quoting with trigger-based	Eight hours of running produced exposure 84 times the declared limit through two undetected defects
05-27	Protocol	Enforce declared and dynamic instrument sets as a union at the ingestion layer	Audit found 4 of 8 declared instruments were not actually streaming
05-28	Pivot	Trigger-based quoting to directional taker entry	823 triggers, zero fills in 18 hours; the spread events were not there at the required width
05-29	Design	Run three shadow variants in parallel	24 hours at 7,916 trades: take-profit rate 35 pct against a 60 pct requirement
05-30	Pivot	Halt all maker strategies; adopt taker with a wider edge and a mandatory empirical gate	28,119 cumulative shadow trades over 8.5 days; win magnitude below the fee floor at every horizon
05-30	Verdict	T0 empirical gate passed	4,800 cells screened; 4 combinations cleared a 55 pct win rate at a 100 bps target with n at or above 30
05-30	Protocol	Formalise a phase-gated roadmap with hard go and no-go criteria	Post-pivot planning; no deployment without a passed gate
05-31	Analysis	Score the full instrument universe on movement specification	3 of 8 declared instruments failed the movement filter
05-31	Pivot	Static instrument set to a dynamic rolling universe	26 hours of shadow produced 5 trades; the static pool was too efficient to generate events

2. June: the parallel machines, and the measurement error

Date	Kind	Decision	Trigger
06-01	Protocol	Capture an integrity manifest before infrastructure transfer	Pre-migration checkpoint; 662 trades at +1.38 bps expected value, 4 of 6 criteria passed
06-01	Design	Prepare a four-strategy contingency, activating only on failure	Fallback planning; never activated
06-02	Verdict	Do not activate live; filter and re-shadow	995 trades at -1.97 bps expected value; a single signal drove the entire net loss
06-02	Design	Implement a two-gate high-conviction confluence	Realised frequency 1,522 per day against a 5 to 50 design target
06-07	Discovery	Every verdict to date was computed gross	A question about whether the code rather than the strategy was at fault
06-12	Design	Separate research and live containers before any live activation	Shadow strategies were sharing the live container
06-14	Design	Test three approaches on data already streaming but unused	Challenge to whether the architecture was locked to one strategy; it was not
06-15	Protocol	Pre-register a universal evaluation standard	Prior conclusions had been drawn post hoc on small samples
06-17	Design	Build an open-interest accumulation strategy	Market thesis on post-correction accumulation footprint
06-17	Gate	Data verification caught a depth assumption before implementation	Design assumed 7 to 14 days of history; the field held approximately one hour
06-22	Retraction	Withdraw a rejection verdict as invalid	87 pct of 39,250 events traced to one microcap
06-22	Protocol	Canonical data substrate manifest; measure rather than assume	Prior strategies had assumed venue conditions instead of measuring them
06-23	Design	Generic per-strategy universe registry with denylist and liquidity floor	Direct consequence of the contamination finding
06-25	Verdict	Retire the cointegration strategy; record its inverse as an unproven lead	Walk-forward turned +180 bps into -55 to -68 at t at or below -4
06-27	Verdict	Retire cash and carry; do not build spot execution	13 of 14 net-positive instruments had no spot market; the executable one yielded 2.7 pct annualised
06-28	Reference	Publish the measured latency profile as a sizing reference	Need for a factual latency band in future strategy design

3. July: Hyperliquid, and the close

Date	Kind	Decision	Trigger
07-02	Design	Five Hyperliquid strategy designs and a verification runbook	Structural disadvantages on centralised venues at this capital scale
07-03	Verdict	Premium characterisation passed; cohort persistence failed	Half-life 3.9 s median with ADF at or below 1e-10; rank correlation reversed sign across windows
07-05	Protocol	Establish the multi-confirmation doctrine, nine binding rules	Post-mortem of a confluence gate built from two non-orthogonal signals
07-05	Audit	Independent audit of all prior venue work	Found coverage below gate and an acceptance instrument that was invalid
07-06	Design	Post-mortem 15 Bybit strategies; scope three narrow successors	Explicit decision not to abandon the venue; forensic re-derivation of design principles
07-06	Protocol	Data integrity protocol: 11 incidents, 6 root causes, 8 binding rules	Repeated implementation defects producing confidently wrong decisions
07-08	Plan	Set a runway deadline and sequence the remaining gates	Session handoff
07-09	Re-ground	Replace a falsified identity premise with mechanical classification	Persistence premise measured at approximately zero
07-09	Pre-registration	Gate orthogonality by measurement	Doctrine requires orthogonality tested before any profit calculation
07-09	Pre-registration	Cascade event definition and calibration	Frequency gate for a target sample
07-10	Verdict	Reversion after cascade exhaustion, falsified	Net -5.97 bps, fails the null at p=0.56; recovery fraction negative
07-10	Verdict	Adverse selection differs by aggressor identity, confirmed	Delta 0.906 bps, CI95 [0.757, 0.983], permutation p=0, n=4.51M
07-10	Verdict	Momentum following the cascade, falsified	Gross negative at both entry points before any fee
07-10	Verdict	Monetising the identity gap through quoting, falsified	Every arm gross negative; mid-caps worse than majors
07-14	Verdict	Identity bursts, directional: signal real, economics closed	Drift +3.04 bps at p=0; net -5.96 bps against a 9 bps round trip
07-14	Verdict	Whale state-change drift, closed on cadence	Drift -0.60 bps at p=0.723, observed at 92.8 s median detection latency
07-14	Verdict	Passive post-exhaustion reversion: event-limited, not falsified	Interval spans zero; beats a matched null at p=0.0102
07-14	Register	Consolidate all venue research assets	Mandate to build the next round from recorded learnings
07-14	Handoff	Venue track paused, not closed	Three exit mechanisms falsified and audited
07-16	Incident	Migrate a deprecated liquidation topic; fix inverted side semantics	A recorder produced zero files in 42 hours; the feed had been dead for more than nine days
07-17	Characterisation	Session structure in spread and depth	Test a timing hypothesis and a concentration lever
07-17	Verdict	Session structure in reversion quality, closed	No session pair with non-overlapping intervals; pooled result spans zero
07-18	Gate	Cascade premise on the second venue: underpowered and coverage-limited	7 events at 29 pct post-event coverage; the recorder design could not see the answer
07-24	Assessment	State evaluation across both venues	20 of 22 families dead, 2 open and underpowered, sharing one mechanism

4. What the ledger shows

Three patterns are visible in the dates that are not visible in any individual document.

Pivots cluster at the start. Five of the six structural pivots happened in the first eight days. That is not indecision; it is a programme discovering that its cost structure ruled out an entire class of strategy, one variant at a time.

Protocols cluster in the middle. Every one of the durable process artifacts, the evaluation standard, the universe registry, the walk-forward harness, the data integrity protocol, was written between 06-15 and 07-06. That window follows the measurement discovery on 06-07 and precedes the concentrated verdict run in July.

Verdicts cluster at the end. Eleven of the twenty-four July entries are verdicts, and eight of those fall in a five-day stretch from 07-09 to 07-14. Once the apparatus was built, hypotheses could be resolved at a rate the earlier programme could not approach. Three exit mechanisms were tested and closed in that stretch, on a venue the programme had been working for less than two weeks.

That last observation is the return on the middle period. The process artifacts cost roughly three weeks to build and then resolved four hypotheses in five days.