Phantom Auditor adversarial integrity

The embargo pitch, finished.

EdgeBuild-class tools stop at a train/test embargo. A leaky embargo still leaks: overlapping labels, best-of-N selection, settlement giveback and velocity-recycling all survive it. Phantom Auditor runs the full battery plus 9 phantom detectors that fired on our OWN live fleet (they caught a −8,510,669%/day cost-basis phantom we then fixed). If it survives us, it is real.

Combinatorial Purged Cross-Validation doesn't just hold out a test fold — it purges training samples whose labels overlap the test window and adds an embargo band after it, so no information leaks across the seam (López de Prado, AFML ch.7). Green = train, red = purged, amber = embargo, blue = test.

PBO reads off the combinatorial rank distribution; DSR haircuts the Sharpe for the number of trials and the return shape. Same gate that promotes our live quant fleet — you are buying our own referee.

Named per the literature (López de Prado/Bailey · Harvey-Liu · White · Hansen · Romano-Wolf · Politis-Romano · Bai-Perron · Benjamini-Hochberg-Yekutieli). A green dot = computed live on your ledger in the free check below; the rest ship in Quant/Desk.

Buy only what you trade — 7 packs @ $19/mo: Prediction-Market · Options-Greeks · Perps-Funding · Futures-Roll · Spot-Crypto · Equities · Execution-Microstructure.

Statistics tell you if the numbers are lucky. Phantom detectors tell you if the ledger itself is a lie. Nine detectors, hardened on 2.8M of our own legs:

• Near-settle look-ahead · Net-decoupled return · c0 / cost-basis corruption
• Field-drift · pre-liquidity fills · recycle / velocity-explosion · survivorship
• Signal-lag / point-in-time robustness — lag every input one bar; a real edge survives (the classic look-ahead killer).
• Corporate-action / split phantom — unadjusted splits/dividends/spin-offs manufacture phantom moves.

AuditZK is a free summary-stat calculator (no ledger). QuantConnect walls you in. mlfinlab is a library you wire yourself. We are the one-stop, ledger-in, adversarial shop — and we dogfood it on a live fleet.

Paste a backtest summary (and optionally per-trade returns). We run the live subset — DSR/PBO/PSR, Harvey-Liu haircut, MinBTL, White-RC/Hansen-SPA, permutation null, stationarity, and the microstructure checks — and return a family-grouped Go/No-Go.

(optional) per-trade returns — needed for the full battery (>= 20)

Run the battery