Sovereign Engine v23.9 is a decision-support workstation for a discretionary or semi-systematic trader who wants a machine to do the scanning and refuse the bad trades — not to press the buy button.

It runs as a daemon on your own machine (or a cheap VPS) and sweeps 32 instruments — 14 Tier-1 majors first, 18 Tier-2 names, plus a DXY/VIX/US10Y/SPX intermarket block — on 1m, 5m, 1h, 4h and 1d data. For each candidate, it builds the structural picture the way a smart-money trader would read it (market structure, order blocks, fair-value gaps, liquidity sweeps, displacement, premium/discount, OTE, Judas swing, Asian-range sweep, silver-bullet windows, volume profile, L2 liquidity absorption), then routes it through a regime router that picks one of three playbooks — trend expansion, mean reversion, or liquidity provider — each with its own minimum R: R and counter-trend rules.

Only then does grading happen: a 15-component weighted confluence score, a strict 5m sweep-or-MSS trigger model, and then the gates that actually earn the engine its keep — microstructure guard (spread percentile, quote stability, single-tick expansion, news proximity, volume sufficiency), a three-layer macro/event/risk veto stack, a correlation-cluster cap, and a precision gate that re-derives every claimed level and rejects any alert whose R:R or lifecycle isn't actually reachable.

Accepted setups come out as a finished ticket with position size and margin already computed from your broker's instrument spec file, risk scaled to grade (1.5% / 0.5% / 0.25%), and an R: R quoted after an estimated spread and slippage haircut. Alerts go to Telegram (with a FIFO outbox and a resilient worker so an API outage can't lose or spam them) and to a local dashboard on port 8080.

The part most signal engines skip: it grades itself. An append-only journal records every signal, and a resolver walks real 5-minute bars chronologically forward to score what each one actually did, with spread and slippage haircuts applied. Loss streaks trigger defensive mode, daily drawdown past −3R trips the circuit breaker, position sizing feeds a Kelly estimate, and component governance plus drift telemetry flag when a scoring component has stopped predicting anything. Shadow A/B evaluation keeps unvalidated experiments observable instead of live.

It never trades. There is no broker connector, no order-submission path, and no execution API anywhere in the package — that boundary is enforced by absence of code, not by a promise in the docs. The trader reviews the ticket and executes manually. Every provider key is optional: an unconfigured layer report.s UNAVAILABLE, AUTH_FAILED, PLAN_REQUIRED or STALE and is skipped, and when an independent source can't confirm a live price, the setup is rejected rather than graded. It fails closed, out loud, by design.