THE BOT,
AND HOW IT EARNS ITS PLACE
What it does · How I test it · How I decide what runs live
The Algorithmic Engine & Model Fleet
Our battle-tested, non-custodial trading robots engineered for Gold (XAUUSD) and global liquidity. Runs 24/7 on a private cloud VPS directly inside your personal regulated broker account.
Pipeline | Tick to Position
Signal Pipeline
Plate 01 | Core Engine
The Feature Matrix
Five modules. Select one to inspect its spec sheet.
Level 7 Trail
Basket-aware trailing that raises the floor under winners in 7 steps. Avg% for market-relative management, points for hard precision.
Advanced Trailing
Seven steps, two formats. Gains protected without choking the move.
Level 7 Trail
Partial Closes
Trim on milestones. Edge compounds, exposure falls.
Dynamic Trimming
Risk Modules
Daily caps, equity guard, lot throttling on margin health.
Equity Guard
Time and News Filters
Session windows plus macro pause and resume. No chaos trading.
Session Sovereignty
Prop-Firm Tooling
Daily and total drawdown caps, stealth execution, multi-account mirror.
Evaluation Rails
Multi-Asset and Market
Forex, crypto, stocks, commodities. MT4, MT5, Rust terminals.
Cross-Platform
Plate 02 | Grid Doctrine
The Core Grid Engine
Non-predictive by design. Noise cannot be forecast; mean reversion, exhaustion and volatility expansion can be measured. The engine buys dips in small steps instead of one big bet.
01
Counter-Trend Edge
Absorbs expansion liquidity, profits on pullback and rotation.
02
Structured Progression
Spacing, ceilings and throttles cap exposure across 20 domains.
03
Basket Life-Cycle Exits
Trims and 7-stage trail drag break-even toward price.
Precision Grid Orchestration
Each entry recalculates the basket target, pulling the required exit closer to price. Capital spreads over 20 domains so no single swing concentrates margin.
Fill noise, slippage, spread bursts, tick delays - edge verified before live capital. Full protocol in Plate 05.
Plate 02A | The Test Ladder
How a Setfile Earns Its Place
A setfile is only a list of numbers until it has survived a test. Every configuration goes through the same two lanes. First it fights the other configurations in a 360-test mutation pool, hunting for more profit and less drawdown. Then the survivor is put through 1,500 verification tests — 500 periods of real gold history, every one of them run three times under three different conditions. A test only counts when the tester report and the terminal journal agree to the cent.
New configurations are put in a pool together and made to earn their place against each other, on the same ranges and under the same three conditions. The only thing that decides the winner is the numbers: more profit, smaller drawdown.
The winner gets the long test: the full historical pool, period after period, under all three conditions. This is not a friendly sample. It is the whole span, including the periods nobody would choose to show you.
Every period is tested three times, and the three answers must agree. The third condition is deliberately unfair — that is the whole point of having it.
A backtest that reports a number is not proof. The proof is the terminal's own journal, written while the test was running. A result only lands in the ledger when four things are in that journal and the arithmetic closes to the cent.
deposit 100000.00 + profit -407.28 = 99592.72
== journal line, cent-exact · mdd 428.24 from report
24 trades · test ran 16.5s on T3
This is the part nobody shows you. Choosing a setfile is not a click. It is thousands of machine hours, spread over days, and the answer only exists at the end of the last one.
Most of that calendar is queue, harvest and proof, not trading. The machine spends about a minute per test; the rest is the work of making sure the minute was real.
This is the current benchmark configuration measured across the full verification pool. I am showing the weakest condition first, because it is the one that decides whether it goes live.
The delay changes nothing. P1 and P2 land on the same answer, so the result does not depend on perfect execution timing.
The stress condition hurts. A quarter of the periods go red once entries are missed, the price shifts and exits are delayed. That is what a real broker does to you.
That is the number I read. Not the clean one. If P3 cannot hold, the configuration does not go live, no matter how good the first two look.
Before the first test I write the thresholds down: the maximum drawdown I will accept, the profit I require, the tolerance for a rerun. The decision is made against that page, never against a number I saw afterwards.
When a stress test blows the account, I do not adjust the stress until it survives. That is curve-fitting to your own bad case, and it produces a book that looks validated while teaching you nothing about the real risk.
There is no way to reach the live account without going through this page. Not because the gate is clever, but because it was written before the answer was known.
The algorithm is the easy part. This is the part that decides whether it touches money. Everything above was produced on the local Linux environment: the terminals, the local brain, the sandbox and the virtual machines that keep a failed experiment from touching anything real.
Plate 03 | Survival Systems
Risk, Time, Compliance
Built to survive margin pressure, news spikes, thin sessions and evaluation rules.
Trigger state persists through restarts and session transitions. The system never forgets.
Hard algorithmic limits enforcing prop firm drawdown rules and equity integrity in real time with stealth low-friction execution.
Counter-exposure coupling that dynamically flips net bias or scales opposite lots on volatility breach.
Plate 04 | Proof, Not Promises
Backtesting and Robustness Lab
The algorithm is the hypothesis. The machine is the answer. The current verification lane is 1,500 journal-proved tests on real gold history, every one of them checked against the terminal's own journal to the cent, and a set of stress tests written down before the run that decide whether a configuration is allowed anywhere near a live account.
This figure is the model fleet's method, not the setfile verification above. In the 1,500-test lane nothing is trained and nothing is rolled: one fixed setfile runs across 500 fixed periods under three fixed conditions. What is trained is the oracle that reads those results and decides what is worth testing next.
A tester report is a claim. The terminal journal is the proof. Every result that lands carries the run ID, the terminal, the period, the condition, the profit, the drawdown and the verdict — plus the path to the journal line it came from. The ledger is append-only: a wrong result gets corrected in the open, never quietly rewritten.
500 periods of real gold history, from 2004 to 2026 — full years, quarters, months and weeks. The pool includes the periods that hurt, because a configuration that only survives the parts you choose to show is not a strategy. It is a picture.
Instead of permuting every knob, I test the four failure modes that can genuinely kill a grid live: worse fills than the ideal tick price, missing entries when the broker requotes or disconnects, exits that do not realise their modelled result, and spread that suddenly widens. One axis at a time, so when something breaks I know exactly what broke it.
The thresholds are written down before the run, and the answer is read against them afterwards. The binding test is the combined bad week: worse fills and worse exits together. It has to hold the drawdown budget and clear the profit floor, the damage has to be proportional rather than a cliff, and a rerun with a different seed has to land in the same place. If a stress test blows the account, the configuration is discarded — it is never tuned until it survives.
The machine that reads the machine. Per-range surrogates predict profit, drawdown and profit factor before a burn; feature importance surfaces which of the EA's 42 logics drive outcomes; dead-knob detection kills wasted sweeps. Full fleet detail in Plate 04A. Bayesian optimization plus low-discrepancy mass-sweeps propose candidates; active learning ranks the next experiment by insight per compute dollar.
A backtest factory, not a laptop with a terminal open. Parallel terminals driven through a typed control protocol, orchestrator manifests, campaign chunking and clean resume after mid-flight stops. One driver per manifest, orchestrated like a data center - not hobby scripts.
The compounding asset. A structured knowledge base: full map of 42 logics across 20 groups, per-range learnings, distilled lessons queryable by agents, and champion doctrine. Explicitly built as training fuel for a future model - today it is the agents' shared memory; tomorrow it is the fine-tune.
Plate 04A | Model Fleet
The Models Trading The Algorithm
The EA is not a black box - it is a 42-logic machine. A second machine reads the first one. Four gradient-boosted tree model systems (a fleet that includes XGBoost) sit between the algorithm and raw market history. The strongest result so far is blunt: the models proved some parameters do nothing, and we removed them from the search space. This is not "AI trading." This is machine learning as infrastructure - the EA trades; the ML decides what to test next.
The per-range surrogate scored every knob against baseline. On any range where the model cannot separate a setting from baseline, the parameter is dead - it does nothing. Those knobs were cut from the optimization search space. Falsifiable, concrete, and the reason wasted simulation cycles drop.
XGBoost panel model: per logic, per market regime - who makes money where. The EA runs 42 independent logics across 20 engine groups. The map says which earn in a 2004-2015 squeeze, which die, which only work in modern tick regimes. Refreshed on every retrain.
Meta-labeling in the Lopez de Prado pattern: not "does this strategy make money" but "does this exact trade right now make money." Trained on 20,522 real extracted deals; emits win-probability and magnitude per closing deal. The actual technique big-fund quants use - built on retail hardware. Proof-of-concept stage, labeled honestly.
Classifier for drawdown cascades before they happen: family concentration, unrealized pain, regime condition -> P(cascade next N bars). Early warning most retail stacks do not have. Still in development; shown as lab work, not as a live gate.
Gradient-boosted trees from knob settings straight to profit, drawdown, and profit-factor per range - so simulation burns only on ideas that forecast well. Also the engine behind dead-knob detection. Agents call it daily through the MCP predict-range tool.
Plate 04B | Agentic Floor
The Agentic Pipeline
One laptop. That is the whole stack. The rest is the same cloud and local LLMs everyone can rent - the difference is a factory floor. Agents read the same bootstrap doc, follow the same laws, append to the same ledger, and ask the same knowledge base before acting. Cloud models reason; local models recall. The pattern is not "agent generates code." Agents are the researchers; the backtester is the lab; the ledger is the truth.
Agents form hypotheses, burn compute on evidence, reject their own intuition when the ledger disagrees, and hand institutional memory to successors in a format built for a future fine-tune. A fund has departments: research, backtesting, validation, domain expertise. This stack has lanes - the workflows are equally formal.
A custom MCP server gives every agent the same toolkit. Lessons are distilled from measured intervention outcomes - knock out logic X, watch range profit move - not vibes.
The 42-logic algorithm trades gold. The system around it is the actual edge. Machine-learning infrastructure and agent lanes close the department gap a one-person shop would otherwise carry alone - without pretending the lab models are production when they are not.
See the Quantum lab around the bot
The Quantum dashboard is the cockpit that watches this bot run — local LLM agent, IDE, and charting.