Phase 3 — Alpha Expansion Plan

Ranked by expected IR contribution × feasibility × diversification to existing book (gold/silver beta, trend cluster, COT cluster).

Tier 1 — Wire existing code (days)

Rank	Initiative	IR est.	Effort	Owner	Acceptance
1	Futures curve feed for GC/SI (`GC=F`, `SI=F` or Databento)	High (unblocks 3 carry sleeves)	2–3d	Eng	`gc_term_structure` fires in paper 5 sessions
2	Options IV surface (GLD/SLV chain or CME)	High (4 vol sleeves)	3–5d	Eng + data	`vol_risk_premium_pm` not `_empty(missing_iv)`
3	Consolidate trend sleeves → single gold trend + allocator weight	Med (reduce corr)	1d	PM + Eng	Correlation cap in aggregator tests
4	Merge vix_haven + gold_vix_haven	Med	0.5d	PM	One strategy_id in registry
5	Wire `regime_classifier_pm` or delete	Med	1d	Eng	Regime labels in telemetry match classifier

Rank	Initiative	IR est.	Effort	Notes
6	Gold–oil ratio MR	0.4–0.7 Sharpe (est.)	1–2d	BNO/USO plumbing exists
7	GDX/GDXJ rotation	0.5	2–3d	Miners beta explicit
8	LBMA fix-to-fix overnight	0.4	1d	Data partially present
9	Gold/copper macro tilt (CPER)	0.3–0.5	2d	Diversifies rates channel
10	CME gold options OI/skew	High if real	1w+	Datamine budget

Initiative	Why
Revive ML loop (`dead_ml_loop/`) with shadow-only promotion	Meta-label on 46 sleeves — only after per-sleeve OOS
IC-weighted aggregator live calibration	Already in `StrategyICTracker` — prove it moves weights
Cross-asset cap in ensemble	Stop 4× gold trend in risk-on

No new strategy code shipped in this audit window. Highest ROI prototype: futures curve + IV data providers (unblocks 7 existing files).

Until GAP-002 closes, do not add new LIVE sleeves — add to RESEARCH with shadow PnL in Redis only.