Measuring Backtest-Live Fidelity: A Pre-Registered Framework and a Ten-Fold Activity Divergence

Systematic trading research almost universally assumes that a backtest reproduces the behaviour of the live system it models, and almost never measures it. We present a pre-registered framework for measuring backtest-live fidelity directly, and report a concrete failure found with it. Running an event-driven backtest engine and a forward, unattended paper-trading deployment of the same signal pipeline over the same 66-day window and the same seven FX pairs, the backtest took 2 trades where the deployed system took 20 — an activity ratio of 0.10 against a pre-registered acceptance band of [0.5, 2.0]. Over the same window, per-trade returns agreed to within 0.0002 percentage points (means -0.05333% deployed vs -0.05314% backtest) and the aggregate return gap of 0.047 pp fell well inside its 0.225 pp tolerance. The combination is the contribution. A backtest that prices trades correctly but selects them differently will produce plausible and wrong Sharpe, drawdown and turnover-cost figures — and roughly one-tenth the cost drag — and no returns-based diagnostic detects it. We argue that activity parity deserves first-class status alongside return agreement in any fidelity assessment. We also report two verdict states the framework needs in order not to overstate evidence: NO_BACKTEST_TRADES, distinguishing absence of evidence from disagreement, and UNDECIDABLE_UNDERPOWERED, excluding tests that lack the power to decide. Each was added after the harness would otherwise have emitted a false failure. A methods section documents three occasions in one project where a statistical test was asked to decide on a sample that could not decide, and the precondition that prevents it. Scope and limitations. The comparison leg is a broker demo account: real-time prices and real signal code, but broker-simulated fills and no real capital. It is a forward deployment, NOT a real-money control, so the fidelity gap most specific to real trading (simulated fill vs market fill) is outside this study. One signal pipeline, 20 deployed trades, FX only. The divergence's cause is identified as a quantified lead — a rolling-confidence rotation gate suppressing at roughly three times the deployed per-cycle rate — not established. No systematic search for prior fidelity frameworks was made, so the instrument and the case are claimed, not priority.