Six specific mistakes void a gold strategy test without producing an obviously broken result: changing the rules mid-sample, letting unrecorded discretionary skips hide likely losing trades, testing at a size larger or smaller than the intended live account, ignoring rollover spread widening and scheduled US data windows, treating demo fills as equivalent to live fills, and mixing multiple distinct setups into one combined sample.
Each mistake changes what is actually being measured while leaving the reported statistics looking plausible, which is what makes them dangerous: a corrupted sample does not look corrupted, it looks like a normal equity curve describing a strategy that was never actually run as tested.
The corrections are procedural rather than analytical — freeze the rules in writing before the first trade, log every triggered setup including skips, size the test to the intended live account exactly, tag every trade by session and macro proximity, discount demo fills near scheduled news, and keep a separate book per named setup — and are cheapest to apply before a sample starts rather than mid-way through one.