Systems state & virtual real-money criteria (NFA)

What each layer is for

Production scanner & active picks — Generates many candidate signals across crypto, forex, equities, and futures. Volume and diversity are high; not every signal is meant to be traded blindly.
Smart Picks (smart_picks_engine.py) — Ranks candidates using ml_composite (ML score, confidence, forward win-rate hints), plus filters (e.g. score floor, volume FOMO cap, validated score, non-crypto policy). Elite score is a quality input, not the primary ranker.
Cross-aggregation monitor — Combines multiple named systems with trust tiers. Raw predictions / legacy social JSON is BANNED from consensus. kol_consensus is WATCH until forward-validated; news-inferred vs true KOL weighting is scored differently in the execution path.
Virtual “real money” tracker — real_money_tracker.py simulates a small, stricter subset of live picks for measurement only (not your brokerage).
Sports betting stack — Separate pipelines (live-monitor/, sports APIs). Do not mix sports bankroll rules with crypto scanner output.

Stricter criteria used by the virtual real-money tracker

The tracker name means “track as if real size,” still paper / simulation. Selection rules in code (April 2026) include, among others:

Strategy allow-list — Only picks whose strategy key appears in PROVEN_STRATEGIES (includes high-score style flows such as hs_lb_None, selected copy-trader lanes, fear/greed contrarian, several crypto mean-reversion / regime models). The list is curated from historical closed-trade stats and can change.
Score floor — MIN_SCORE (currently 50): code comments note weaker historical win rate below that band.
Risk–reward cap — MAX_RR aligned with smart-picks RR policy (wide targets are deprioritized).
Confidence “dead zone” — Avoids a mid confidence band that historically underperformed in the project’s closed-pick sample.
Safety / conflict checks — Requires compatible safety tags and no direction conflicts; limits concurrent picks per symbol.
Staleness — Old picks are rejected for this virtual book (freshness matters in forward results).
Symbol blacklist — e.g. symbols with repeated losses in this virtual book are blocked from new entries.
Cooldowns — Weekday filters and consecutive-loss cooldowns reduce overtrading after drawdowns.

Full detail lives in the repo: alpha_engine/real_money_tracker.py. NFA — this is engineering documentation, not a recommendation to mirror settings with real cash.

Hedge-fund-quality policy targets (approved thresholds)

Multi-agent sign-off (April 4, 2026) documents institutional-style gates for decay, concentration, display tiers, retirement, and conflicts. Implementation rolls out in scoring and audit tooling; cite the source of truth:

docs/HEDGE_FUND_QUALITY_NEXT_STEPS.md

ID	Policy (summary)
A	Backtest vs forward decay hard-gate when sample is large enough
B	Concentration caps (per symbol / direction / duplicate portfolios)
C	Tier-1 display defaults to multi-system agreement; single-system labeled experimental
D	Strategy retirement after sustained poor forward win rate with enough closes
E	Portfolio circuit breaker on deep drawdown (pause new risk)
F	Stronger penalty when ML composite falls back without a real model score
G	Hard-reject true consensus conflicts (direction vs recommendation with high delta)

NFA — thresholds describe internal risk research targets, not personal position sizing advice.

Operational notes

Redis agent bus — Used for lock coordination and broadcasts (e.g. HF-P0, audit template, scraper outages).
Copy-trader ingestion — Hyperliquid leaderboard URL drift was flagged; failover to the official API shape is an ops priority (see bus log from dash-integrity).
Verified alpha — Roadmap item: per-pick cohort linkage on the audit UI (not only aggregate counts).