Featured Research ProgramES · NQ Futures

Atlas ActionEdge Project — Trade Automation

Discretionary edge, engineered into a system that can be audited. ActionEdge converts the Atlas indicator methodology into a governed, evidence-scored automation stack for S&P 500 and Nasdaq-100 futures — one in which every decision is typed, hashed, journaled, and reversible.

Member Access

ActionEdge Mission Control header showing gear, scenario and leader badges plus MAC, ES and NQ node heartbeats — Fig. 1 — Mission Control. Live 2-way heartbeats from the Mac control node and the ES/NQ execution VPS, the validated release each node is running, and the current Gear · Scenario · Leader verdict.

Why It Matters

Most retail automation fails in one of two ways: it hard-codes a signal and trades it blindly, or it wraps a black-box model that no one can interrogate after a loss. ActionEdge rejects both. The project's premise is that a trading edge is only durable when its reasoningis as observable as its P&L.

The stack begins where AQS research already lives — the Atlas indicator suite running on NinjaTrader 8 Renko charts for ES and NQ. Trend, regime, momentum, market-profile, and session-level telemetry is exported as frozen, schema-validated JSON and streamed to a scoring layer that produces a single, explainable verdict: gear, scenario, and leader. A confluence-lead-lag read on ES and NQ is treated differently from a sticky or conflicting one, and the system says which it saw and why.

An expanded Decision Journal row showing the 13-gate ladder for ES and NQ, the cross-market arbiter output, and the Sim101 fill — Fig. 2 — One decision, fully opened. Thirteen ordered gates per instrument, the ES×NQ arbiter (β, zε, ρ, lead/lag, dual-CVD), and the resulting Sim101 fill: buy 2 @ 5878, stop 5865, target 5906, exit at target for +2.8R.

Execution is deliberately two-key. The NT8 strategy consumes decision IDs exactly once, applies its own vetoes, and routes through predefined ATM templates. A decision the scoring layer likes but the execution node distrusts does not trade. This asymmetry — permission is earned, refusal is free — is the core risk philosophy of the project.

The Learning Loop

Every decision, its inputs, and its eventual outcome land in a versioned DuckDB/Parquet lakehouse. Nightly calibration re-estimates expectancy, re-fits regime models, and proposes policy adjustments as drafts. Drafts are diffed against the active policy, bounds-checked, and content-hashed. Nothing is promoted without a human reading the diff. The loop learns; it does not self-authorize.

The Decision Journal

Every observation the engine makes becomes a row: time, instrument, gear, ES×NQ scenario, setup, the S / C / E score triple, the verdict, the first gate that blocked it, and — for fills — the entry and realised R. Stand-offs and vetoes are recorded with the same rigour as trades.

Sim101 · fixture data

Decision Journal title bar with a cumulative-R equity path and per-trade win/loss markers drawn as a watermark — Fig. 5 — Journal header. The cumulative-R path of every filled trade is drawn behind the title; green markers are winners, red are losers.

Decision Journal rows 1–16 listing NQ and ES decisions with verdicts FILLED, VETOED and STAND_OFF and realised R multiples — Fig. 6 — Journal, first window. FILLED rows show entry mode, fill price and realised R (+1.8R, −2.1R, +2.8R…); VETOED and STAND_OFF rows name the exact gate or desync that stopped them.

Decision Journal rows 17–34 — Fig. 7 — Journal, second window. Note gear cycling G0→G3 as conviction builds and decays.

Decision Journal rows 35–52 — Fig. 8 — Journal, third window. ES and NQ alternate leadership; the arbiter records which side led.

Decision Journal filtered to FILLED verdicts only, showing the trade list with entry prices and R outcomes — Fig. 9 — Filtered to fills only. The trade list as a desk would read it: instrument, setup, entry, and R — with the blocking-gate column still present so near-misses stay visible.

Governed Handles, Not Hidden Knobs

Every tunable parameter is exposed with its bounds, its JSON path, and its lock state. The funnel is staged so a change to a Stage-2 veto is never confused with a Stage-5 sizing tweak, and the policy diff shows exactly what a draft would change before anyone can save it.

Funnel handles across Stages 1–5 with bounded numeric inputs, locked fields, and the signal-by-regime weight matrix — Fig. 10 — Funnel handles, Stages 1–5. Bounded inputs (e.g. ADX floor 35 within [25…45]), locked hard rules, and the indicator-weight matrix that must sum to 1.000 per regime.

Control plane strip — Fig. 11 — Control plane. Operator actions are explicit and logged.

Policy diff panel comparing baseline against draft for arbiter, cadence, expectancy, gates, kelly, risk, rollback and weights — Fig. 12 — Policy · Active vs Draft. Every path of the policy document is diffed and content-hashed. Saving writes only a draft; promotion is human-only.

LLM hand-off brief panel with copy and download buttons — Fig. 13 — LLM hand-off brief. Selected journal rows, policy state and node health export as a Markdown/JSON bundle for review by an analyst or a model.

Three-Node Control Plane

A Mac research and control node, an AWS broker/overseer tier, and dedicated ES and NQ execution nodes exchange typed, hash-verified telemetry over a persistent MQTT fabric with last-will detection. If any node goes dark, the system knows before a trader does.

Gated Scoring & the Gear Ladder

Every ES/NQ observation passes through hard gates, normalization, a composite conviction score, and an expectancy model before sizing is even considered. Permission advances through four gears — Dormant, Monitoring, Armed, In-flight — and can only step up when evidence agrees.

Regime-Aware Intelligence

Hidden Markov and Bayesian online change-point models classify the session regime and flag structural breaks. Lead/lag confluence between ES and NQ is scored explicitly, so the engine distinguishes a confirmed rotation from noise.

Human-Only Promotion

Policy changes are drafted, diffed, hashed, and staged — never silently activated. Promotion to a live policy requires a multi-node activation quorum and human sign-off, with an automatic rollback breaker as a non-negotiable release gate.

Back to summary