Placeholder copy. The structure is real; the words are not.
The problem
A backtest that looks good is the default outcome, not the signal. Lookahead bias, survivorship bias and an over-fitted parameter sweep all produce the same green equity curve.
The approach
Replay the data one bar at a time through the same interface a live strategy would see, so a strategy physically cannot read a price it would not have had.
The tradeoff
Bar-by-bar replay is roughly two orders of magnitude slower than a vectorised backtest. That rules out large parameter sweeps — which, given what sweeps do to out-of-sample results, is arguably the point.
What I would do differently
Record the slippage and fee model as part of the result. Comparing two runs made under different assumptions was a mistake it was too easy to make.