Every analysis is stamped with the engine version that produced it, and history is never regraded — a journal entry shows the rules as they stood when the decision was made. The turning point in that history was a walk-forward replay that found the score did not rank outcomes out of sample.
The stamping contract
The engine version is recorded on every analysis, and past reads are never re-scored when the engine changes.
A retroactively improved track record is not a track record. If history were regraded, every published accuracy figure would silently describe a product that never actually made those calls.
The early eras
The first versions fitted per-market thresholds from real candles rather than using one global set, and added the crypto-only positioning factors.
A subsequent version made the higher timeframes count properly: the multi-timeframe weight was raised, counter-trend entries became a default rule, and a read with no readable higher timeframe was capped.
The walk-forward replay, and what it found
A large replay of resolved historical setups tested whether the confluence score ranked outcomes out of sample. It did not — discrimination sat near chance in every asset class, and roughly the same fraction of simulated setups reached target first almost regardless of score.
The result was published rather than buried, and it reshaped the product in three ways. The grade is presented as evidence agreement rather than as a probability. The calibrated confidence exists as a separate, checked number. And the declared factor weights stood, because no fitted alternative beat them out of sample.
The replay measures simulated fills rather than resolved live outcomes, and every surface that leans on it says so. A replay is not allowed to borrow the authority of a live track record.
Since then
Later versions introduced evidence-gated entry behaviour: a cell's entry mode flips to a resting limit only when a replay shows it clearing pre-registered criteria, and the criteria are fixed before the fit runs.
That machinery has also been extended with measurement-only candidates that ride along a plan without affecting it, so a future change can be evaluated on real data before it is ever allowed to alter what a trader sees.
Frequently asked
- Did the walk-forward backtest prove the grades work?
- It proved something more useful: out of sample the score did not rank outcomes. So the grade is presented as evidence agreement, and the calibrated confidence — anchored on the replay until live records clear their floors — is the number carrying a checked claim.
- Are my old journal entries re-scored when the engine changes?
- No. Every analysis is stamped with its engine version and history is never regraded. A retroactively improved track record would describe a product that never actually made those calls.
Updated Sep 1, 2026