Six layers. One direction. Auditable end to end.
The platform is a unidirectional pipeline. Each layer can fall back to deterministic logic if the layer above it is unavailable. Nothing places an order without passing the containment layer.
01DiscoveryUniverse funnel across equities, ETFs, bonds, forex, crypto
02ConvictionStrategy signal plus AI scoring and news context
03AllocationSizing, displacement and cluster caps
04ExecutionVenue selection, fee and slippage routing
05ContainmentRisk engine veto, the final authority
06AuditFills ledger, decision heritage, deterministic logs
Strategy families
StrategyStatusDetail
quantitativeliveMean reversion, breakout, factor sleevesnews_drivenliveHeadline scoring, embargo and freshness checksmacroliveRegime classifier, rate and curve sensitivityvolumeliveUnusual flow, order book imbalancevolatility_regimegatedConfidence gated by AI classifierfunding_rate_arbitragepaperCross-venue, paper onlyThe AI pipeline
AI is local-first. It feeds sizing and conviction. It never places orders. The execution path is rule based and auditable.
Stage 1
Rules
Deterministic filters. Always run. No model needed.
Stage 2
FinBERT
Local sentiment scoring on news and filings.
Stage 3
Local LLM ensemble
Qwen + Llama. Runs on operator hardware. Voted.
Stage 4
Premium fallback
Optional. Only invoked under explicit policy.
AI is advisory. It does not place orders.
Model outputs feed sizing, conviction, and, under explicit policy, signal-level vetoes. The order path itself is deterministic. If every model is offline, the platform still operates on rules.