For external readers: This document is one of 10 curated case-study artefacts from Project Aṣe, a quantitative crypto-trading research project that produced a documented negative result with mechanism across four strategy attempts plus a literature audit, all under pre-committed criteria (Sharpe > 1.0 AND profit-factor > 1.4 OOS, called “§14” internally) that held without bar erosion across 19 ADRs. The full project repo is private (scope B of the Phase 5e shipping spec). Start with ./00-readme.md for the reading guide. Read order context: The methodology that emerged from the first two kills. SIA = Signal Independence Audit. Reference this for the four-screen framework. The substance of the document below is unchanged from the internal version; only this framing block and personal/identifying content scrubs are added.

ADR-0014: Signal Independence Audit (SIA) framework

Context

BMR’s Phase 3 §14 failure (ADR-0013) had a deeper finding: the failure was diagnosable from BMR’s design (the macro-EMA stack dominated the entry signal; the Bollinger parameters did not materially shift WHICH candles fire) but was only discovered after ~30 hours of Hetzner hyperopt + OOS + walk-forward compute. This is wasted compute on a thesis-level defect that could have been caught pre-compute with a small battery of pre-hyperopt screens.

The Signal Independence Audit (SIA) is the formalised “would BMR have died here?” check, applied prospectively to every new strategy thesis from Phase 4 onward.

Decision

The SIA is a hard gate before any hyperopt compute. A new strategy thesis cannot enter hyperopt until SIA passes. If SIA fails, the strategy is killed and a kill ADR is written. The SIA consists of four screens applied to the training window only (2021-01-01 → 2024-12-31), with the first ~97 days excluded as warm-up for any rolling-window indicators.

Screen 1 — Coverage test

Screen 2 — Lift test

Screen 3 — Entry-set stability (Jaccard)

Screen 4 — Information split

Statistical methodology

Brainstorm 2026-05-11 selected Option B: parametric headline thresholds for fast mechanical verdict (matching the existing §14 report style) + bootstrap 95% CIs as supporting evidence in the SIA report. Pure parametric (Option A) was rejected as too loose; strict hypothesis tests with Bonferroni correction (Option C) were rejected as too strict for the 4-year sample size.

Verdict logic

Condition Verdict Exit code Next action
All 4 screens PASS PASS 0 Proceed to hyperopt + OOS
Any screen FAIL, no inversion trigger FAIL 1 Kill ADR written; strategy phase closes
Any screen FAIL, inversion trigger active FAIL+INVERT 2 Phase-prime authorised under separate ADR

Inversion clause (Phase 4 application)

For Phase 4 specifically: the funding gate is locked at funding < 60th pct based on Option D of the design brainstorm (locked thesis: low funding = uncrowded positioning = fresher fuel for breakout). The brainstorm noted a ~50% prior on direction with mild lean against the locked direction. To buy insurance against the false-negative scenario at the cost of one pre-committed contingency:

Generality

The SIA framework is strategy-agnostic. For non-direction-gated strategies (no inversion clause needed), Screen 2’s inversion test simply isn’t run; Verdict logic still applies.

Consequences

References