XGB-GATE-20 Research Watchlist — ML Observational

XGB-GATE-20 is an observational ML risk-gate candidate selected from the Phase 2 feature-family grid. Historical research evidence is shown separately from forward watchlist evidence. Forward PnL starts at dashboard activation and is not backfilled. Paper-only; real orders disabled.

Drawdown is measured on the cumulative daily net-PnL path. No historical research rows appear in this table — the pinned PR-evidence aggregates never had a per-day series persisted, and none is fabricated here.

Historical PR #18 evidence above is a fixed, pinned research artifact — it is never recomputed or blended into the forward numbers. Forward PnL starts at zero at the dashboard-activation timestamp and accrues only from decisions made at or after that moment. There is no code path by which historical backtest PnL could leak into the forward total.

Model family: XGBoost · Feature set: side_filter_features · Use case: ML risk gate / side filter.

Feature-lineage review passed (2026-07-07): the two reviewed decision-time inputs, edge_q10 and model_prob_no, were both approved as safe decision-time prior features (no settlement-derived information, structurally excluded).

Full feature-family taxonomy and public-safe descriptions: Feature Families page. Exact feature formulas, full feature lists, hyperparameters, and thresholds beyond the public q20 label are not published.

XGB-GATE-20 is the single public watchlist candidate to emerge from the Phase 2 grid: 89,040 evaluated rows → 30 representative candidates → 7 nominal passes → 2 distinct underlying models → 1 public candidate. Near-duplicate calibration/threshold variants of the same underlying model were collapsed into a single public entry per distinct tradeable idea rather than shown as separate candidates. Full audit trail retained in the internal research artifacts for this run.

Promotion status: Research Watchlist only. Execution review would require a separate human approval process, longer forward evidence, risk-limit review, and no change to real_orders_enabled=false.