Those failure modes are exactly what the credential is built to catch — and the literature confirms their prevalence: models reaching 100% accuracy in-sample score roughly 50% out of sample (sources: mizan.market/evidence). Today the engine enforces point-in-time data, real after-cost returns, a forward-committed (un-backfillable) window, and the full López de Prado significance suite. As of 2026-07-21 all three of his statistical corrections are live and re-verifiable, each minted as a real STARK: the Deflated Sharpe test (selection bias — penalizes a Sharpe for the number of trials behind it), PBO (probability of backtest overfitting, held to ≤0.5), and CPCV (purged, embargoed combinatorial cross-validation for leakage). We found the bugs we retracted in our own work — that discipline is the product, not a liability. The retraction ledger is the receipt that we hunt them rather than hide them. Two honest limits still hold and we say so: the trial count N is forward-only (it can't recover experiments run before the ledger opened), and high-frequency strategies are refused outright — bar data cannot honestly verify them. Serial correlation itself is handled: the engine deflates by an AR(1)-corrected effective sample size, and since v11 the Harvey–Liu/Hansen suite (PSR, SPA, minimum backtest length) runs alongside the López de Prado corrections.
Every number above is on the record: every credential · verify one yourself · the vocabulary · all questions.