Blog · 12 posts · 86 min end to end

Measured,
not explained.

Build logs and measurements from the desk that writes the software. Numbers we ran ourselves, including the ones that came out against us.

Build log09 Sept 20267 min

v0.15.0: 26 of Our 27 Backtests Were Identical — and the Optimiser Ranked Them Anyway

Our walk-forward offered three parameters to sweep. They were hard-coded — the settings of the example strategy, not of the strategy you loaded. Load a model that reads fifteen settings of its own and the optimiser sweeps two keys it never reads: 27 runs, 26 of them bit-identical, a "best" configuration chosen by floating-point rounding, and nothing on screen to say so. That is the worst failure mode a measuring instrument has — it returns a number. This is what we found, what we shipped to fix it, and the second thing v0.15.0 adds so a backtest can tell you when it is fooling itself.

Read it updated September 2026

Everything else