BL-A-004: Glossary

BIJECTIVE LABS

BIJECTIVE LABS . SEP 7, 2026

BL-A-004: Glossary

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-07

Terms this record uses in a specific sense. Ordinary market vocabulary is not repeated here; what follows is the vocabulary a reader could otherwise reasonably interpret a different way.

Verdict vocabulary

FALSIFIED. A pre-registered kill criterion was met, with the statistical power to support it. The hypothesis is dead.

CLOSED. The mechanism measured real and a declared envelope parameter blocked it. Not the same as falsified, and the distinction is the point of the classification system.

NOT PROVEN. The interval spans the threshold, or declared power was not achieved. Nothing was concluded. This is a terminal state for a hypothesis and it is never recorded as a kill.

OPEN. Under test, or awaiting a measurement that has not been taken.

WITHDRAWN. A verdict that was issued and then retracted. The document stays in place with the retraction attached.

CONFIRMED. A fact about a venue rather than a hypothesis about a strategy. Carries no axis of death and no transferability.

VALIDATED. A component that passed its acceptance criteria and reached production. One row in this record holds it.

Classification vocabulary

UNIVERSAL. The mechanism itself failed. No envelope substitution rescues it, and the result carries to any operator on any infrastructure.

ENVELOPE-BOUND. The mechanism held and a declared envelope parameter blocked it. A reader with different parameters should expect a different answer, and the dossier states the value at which the verdict reverses.

PUBLIC, CLIENT, RESTRICTED. Distribution classes. Restricted material appears in the Kill Ledger by identifier and status with its substance withheld and the withholding stated openly.

Method vocabulary

Operating envelope. The seven declared parameters a measurement was taken under: cost, latency, capital scale, venue access, fee progression, measurement cadence, and the statistical envelope. Declared in BL-FM-002.

Axis of death. What actually killed a hypothesis, recorded so that a falsification on one axis does not spread to axes never tested. The taxonomy is in BL-D-005.

Reversal point. For an envelope-bound verdict, the parameter value at which the verdict flips, with the arithmetic shown. Mandatory in dossier section 10.

Mirror test. Running the same measurement on the opposite condition. If a conditioning variable predicts direction, its inverse should predict the opposite. When both sides produce the same sign, the variable is not separating outcomes.

Matched null. A control sharing timing, turnover, cost and exposure with the strategy, differing only in the claimed association. A null that changes anything else measures the wrong thing.

Layer 0. The cost-floor pre-check. Compare plausible signal magnitude against the cheapest available round trip before writing a hypothesis.

The chain. The eight-layer falsification sequence in BL-D-001, ordered by what each test costs to run rather than by intellectual importance.

The doors. The six economic mechanisms enumerated and measured in Programme II. Closing a door closes every strategy that would have gone through it.

Sealed era. A partition of the substrate that research may not touch, opened once against one pre-registered specification.

Shadow. The complete strategy running against the live market in real time, computing net from the first day, placing no orders.

Zero-order guarantee. A shadow process structurally incapable of placing an order, enforced by absent wiring rather than by a configuration flag.

Non-disturbance invariant. New research runs as an additive parallel shadow and never restarts, recreates or attaches to a running strategy.

Measurement vocabulary

Net. After fees, spread, slippage, and funding accrual where the position spans a settlement. Every figure in this record is net unless explicitly labelled gross.

Round trip. Total cost of entering and exiting, both legs, in basis points. Written c_rt.

Collapse. Aggregating per-instrument returns at a timestamp into one portfolio return before any inference. Required on this substrate because mean pairwise correlation is 0.737.

Effective sample. The number of independent observations a correlated panel actually carries. On the fifteen major perpetuals it is 1.33 per timestamp against a nominal 15.

Markout. Post-fill price movement in the direction that hurts the passive side, reported against the half-spread earned rather than alone.

Adverse selection. The cost of being filled preferentially by counterparties who know something. Distinguished from other costs because it scales with how attractive your quote is, so it cannot be escaped by quoting better.

Anchor. The phase of a periodic parameter, most often the hour at which a periodic rebalance starts. A parameter fixed by convenience rather than chosen.

Anchor sensitivity. The distribution of a result across every phase of its period. Reporting one phase without disclosing the others is unacknowledged multiple testing.

Information coefficient. Rank correlation between signal and forward return. Permitted for characterisation and forbidden in any decision path, because converting it to expected profit over-predicted by 5.6, 13.7 and 7.3 times on three occasions.

Event-limited. A hypothesis whose sample is bounded by the market's own rate of producing the events it needs, rather than by recording time or effort.

Coverage-limited. A hypothesis the recorder could not observe the answer to. Distinct from a null result, and recorded as such.

Programme vocabulary

Pre-registration. The metric, sample, method, cost model, power requirement and kill thresholds, written and locked before any data is seen.

Amendment. A dated document revising a pre-registration. May correct an instrument or tighten a threshold. May never loosen a threshold after data has been seen, and may never change a horizon.

Hypothesis budget. A cap on how many hypotheses a programme may test in total, declared before it begins. Programme II capped at twelve and used eight.

Five-angle audit. Adversarial review across calculation, cost, data, sample and mechanism, run by default before every verdict rather than on request.

Evidence identifier. [E-nnnn], resolving through BL-A-001 to a machine-readable artifact under _evidence/. No released number stands without one.

Rules Register. The standing decision rules in BL-D-009, each carrying the specific failure that produced it.

Two distinctions worth reading twice

Falsified against not proven. The first says the hypothesis is wrong. The second says the measurement could not tell. Conflating them closes branches nobody tested, and the programme did exactly that once before the rule was written down.

Universal against envelope-bound. The first is a statement about the market. The second is a statement about the operator. A record that does not separate them tells a reader nothing they can use on their own infrastructure.