How ATLAS turns a market snapshot into a decision.

Mechanism

The layer reads the account block of the snapshot and asks binary questions first: is trading permitted on this account, is the Expert Advisor allowed to trade, are balance and equity real positive numbers? Any hard failure ends the heartbeat immediately — there is no point analysing a market the account cannot touch.

Beyond the binary checks it computes basic health ratios — equity against balance, margin usage against equity — and emits soft warnings when they are elevated. Soft warnings do not stop the heartbeat; they travel forward in the context so later layers can trade more defensively.

Mechanism

Balance is a headline; free margin is the truth. This layer derives what capital is actually deployable right now: it compares free margin to equity, computes the current margin level, and checks both against configured floors. A large account with heavy open exposure can still fail here.

The output is a set of capital metrics — usable capital, margin headroom, utilisation — that the risk layer and the sizing layers consume later. When headroom is thin but above the hard floor, the layer degrades posture instead of blocking outright.

Mechanism

Brokers rename instruments freely — XAUUSD, XAUUSD.m, GOLD. The layer normalises the incoming name against the configured allow-list, so the rest of the system reasons about one canonical instrument.

It then verifies the contract is complete enough to trade legally: tick size and tick value, digits, minimum and maximum volume, volume step, and stop-distance rules. Every later price and size computation depends on these fields, so a missing or absurd value is treated as a hard failure, not a guess.

Mechanism

Three independent checks, each fatal on its own. Freshness: the quote's timestamp must be within the configured staleness window — a decision made on an old price is a decision about a market that no longer exists. Depth: each required timeframe must carry enough bars for the analytics that follow; a trend measure over too little history is noise. Integrity: bars must be chronologically ordered, and every bar must obey basic geometry — high at or above open, close, and low; low at or below all of them.

Mechanism

Two profiles built side by side. The friction profile expresses the live spread relative to the instrument's own recent norm — a spread that is normal for gold is catastrophic for a major FX pair, so absolute numbers are never compared across instruments.

The movement profile summarises recent range behaviour (an ATR-style measure across timeframes) and classifies the moment: compressed, normal, expanding, or shock. These labels are not decisions; they are context. A shock label, for example, later makes the execution-cost particle price friction more punitively and pushes risk posture defensive.

Mechanism

ATLAS reads the market on several frames at once — from minutes to hours. This layer confirms the frames are mutually usable: each required frame is present, each is current enough, and their latest bars describe the same moment in market time. A frame that lags the others is marked unusable rather than silently trusted.

The result is a timeframe map that tells tactics which views may drive a decision and which may only provide background context this heartbeat.

Mechanism

A regime classifier. Using directional strength and range behaviour it labels the environment — trending, ranging, volatile-choppy, or quiet — because the same signal means opposite things in different regimes: a break of a level is an entry in a trend and a trap in a range.

The label is deliberately coarse. Its job is not to predict; it is to tell the tactic families which of them should even bother proposing scenarios, and to let scoring reward regime-appropriate ideas over regime-blind ones.

Mechanism

Around scheduled economic events, technical evidence loses meaning — price moves on the announcement, not the chart. This layer evaluates proximity to configured sensitive windows and emits one of three postures: clear, caution, or blackout.

Caution does not block trading; it reduces conviction — the news particle later scales directional bias down during caution windows. Blackout is different: inside a blackout window the news particle holds veto power and no new trade opens, however good the technical picture looks.

Mechanism

The layer combines the stored risk state — realised losses today, open exposure, recent outcome streaks — with policy limits to answer one question: how much new risk may this heartbeat add? The answer is a budget, not a boolean.

Alongside the budget it emits a posture: aggressive, normal, defensive, or stand-down. Posture is a ladder: each rung raises the quality bar a candidate trade must clear (minimum reward-to-risk, minimum confidence). After a losing streak the system does not switch off — it demands progressively better setups before it will act again. Daily-loss or exposure breaches are the hard edge: budget goes to zero and the heartbeat ends in WAIT.

Mechanism

Sizing happens twice in ATLAS, deliberately. This layer does the first half: from the risk budget and the instrument's tick economics it derives a base volume — the size at which a stop-out costs exactly the budgeted fraction of capital — then clamps it to the broker's minimum, step, and maximum volume rules.

The result is an envelope: a floor, a ceiling, and a policy base inside it. The final number is chosen later by the lot-decision particle, after trade quality is known — a mediocre-but-acceptable setup gets sized below base; a clean one gets the base, never above the envelope.

Mechanism

Exits are proposed from two independent geometries and reconciled. Structure-based: the stop belongs beyond the level that invalidates the idea — under the swing low for a long, above the swept high for a fade. Volatility-based: the stop must sit outside normal noise, expressed as a multiple of current range, or ordinary fluctuation will destroy a correct idea.

Targets are proposed the same way, toward the next meaningful structure, and every candidate is snapped to the instrument's tick grid and checked against the broker's minimum stop distance. These are candidates — the stop-loss and take-profit particles make the final selection for whichever scenario wins.

Mechanism

The final aggregation. It re-reads every prior layer's verdict, applies session and cool-down rules (minimum spacing between trades, limits on concurrent exposure in one direction), and produces a go / no-go for scenario generation — sometimes constrained: eligible, but long-only, because existing exposure already leans short.

When the verdict is go, it collapses all twelve layers' outputs into a single immutable Prepared Data Packet — the only input the decision engine is allowed to read. DTEA never touches the raw snapshot; it sees the market exclusively through validated, structured preparation.

Mechanism

Reads the posture computed in L08 and converts it to decision-time effect. Clear contributes nothing. Caution applies a configured haircut to directional conviction — the committee may still trade, but with reduced confidence, which in turn shrinks size through the quality-aware lot decision. Blackout raises the veto flag: whatever the other particles conclude, no trade opens inside the window.

Mechanism

Three short-horizon measures. Momentum: net directional progress over a recent bar window — is price actually going somewhere, or oscillating? Pressure: over an even shorter window, are closes finishing near their highs (buyers in control) or near their lows? Proximity: is the current price sitting within a configured distance of a known structural level — a place where behaviour tends to change?

The three combine into a directional lean with a confidence. It is deliberately reactive rather than predictive: its job is to describe the tape as it is right now.

Mechanism

A pattern reader with strict definitions. An engulfing bar only counts when its body exceeds the previous body by a configured ratio — near-misses are noise. A pin bar requires the rejection wick to dominate the bar's total range. Inside bars signal compression; a short streak of same-direction closes signals momentum continuation.

Patterns are read on a designated primary timeframe so the same shape is not double-counted across frames, and each detected pattern carries a direction and a strength into the evidence record.

Mechanism

Builds a map of where the market has repeatedly reacted. Swing highs and lows over a lookback window are clustered into zones — a level only qualifies when price has touched it a minimum number of times within tolerance. Equal highs or lows within a tight band are flagged specially, because clustered stops live behind them.

The particle then classifies live interaction with those zones: a genuine break requires a close beyond the level with margin; a break that closes back inside within a few bars is reclassified as a failed break. A spike through a level that immediately rejects with a dominant wick is flagged as a suspected liquidity sweep — a trap. Suspected sweeps penalise the score of ideas trading in the sweep's direction rather than hard-blocking them, keeping the system flexible where evidence is circumstantial.

Mechanism

On each timeframe it evaluates a stack of three moving averages — fast, intermediate, slow. Fully ordered stacks (fast above intermediate above slow, price above all) read as established trends; tangled stacks read as no-trend. A directional-strength measure (ADX-style) must exceed a threshold before any trend label is granted at all — direction without strength is drift, not trend.

Per-frame verdicts blend into one bias, and the particle also watches for exhaustion: when price has stretched multiple average-ranges from its mean, or printed a long unbroken streak of one-direction closes, trend confidence is reduced — the observation that a move is mature is different from the belief it will continue.

Mechanism

Compares current range against the instrument's own baseline over a lookback period. Current range well above baseline is expansion; well below is contraction; an extreme spike is a shock. Each state changes downstream behaviour: expansion widens stop geometry so noise cannot reach it, contraction warns that breakout tactics may be firing on air, and shock flips the system risk-off — simulation gets harsher, cost estimates inflate, posture tightens.

Mechanism

The committee's most influential member, and the most unusual: it evaluates not the market but the transaction. It estimates the full round-trip cost of a candidate — spread plus expected slippage, inflated by a multiplier during volatility shock — and expresses that cost in the only unit that matters: as a fraction of the trade's risk unit and of its expected edge.

Two veto lines follow. If cost exceeds a configured fraction of one risk unit, or would consume more than a configured share of the setup's expected edge, the particle vetoes — the setup may be genuinely good and still untradeable at this moment's friction. This single mechanism kills more marginal trades than any other, which is precisely its job: a strategy that ignores cost is profitable only in backtests.

Mechanism

Carries the L09 posture into decision time and enforces the ladder there. Each posture rung defines the minimum reward-to-risk and minimum confidence a candidate must clear: normal conditions accept solid setups; defensive conditions demand better ones; stand-down demands setups so exceptional they almost never occur — which is the intent. A configured loss streak escalates posture automatically, and a strong regime alignment can grant a modest posture boost in the other direction.

Mechanism

Builds a fingerprint of the current situation — regime, structure posture, volatility state, setup family — and queries the LMPI corpus for the most similar past situations and how they resolved. A strong historical lean nudges confidence up or down; a thin or ambiguous sample contributes nothing.

Its defining property is humility. Memory influence is capped by design, and every failure mode — empty corpus, query error, timeout — degrades to silence. A memory problem can never invent a trade, block a trade, or crash a heartbeat.

Mechanism

Trend-follow proposes in the direction of established bias, preferring entries on pullbacks toward value rather than chases into extension, with stops beyond the structure that would invalidate the trend thesis. Breakout watches the levels the structure particle mapped and proposes on qualified breaks — closes beyond the level with margin, ideally with expansion behind them — while the sweep detection actively filters the traps. Mean-reversion proposes fades of overextension back toward equilibrium, and only when the regime label says ranging; fading a genuine trend is how reversion strategies die.

Each family checks the regime first and stays silent when conditions don't fit. A quiet heartbeat with zero scenarios is a normal, correct outcome — it simply resolves to WAIT.

Mechanism

Each candidate's exit geometry is examined against modelled price behaviour drawn from current volatility: how plausibly does price reach the target before the stop, and how does that balance shift when conditions are less friendly than assumed — wider spread, more adverse noise, a volatility step-up? The output per candidate is a robustness picture, not a prediction.

The specific failure this stage hunts is fragility: ideas whose arithmetic only works when everything goes right. A candidate that survives its stress variants earns the right to be scored; one that collapses under mild degradation is marked down or discarded here, cheaply, instead of in the account.

Mechanism

Every surviving candidate is scored against the full evidence set using the committee weights: alignment with trend bias, quality of the structural context, pattern support, momentum agreement, volatility fit, simulation robustness, memory lean — minus the friction penalty, which is weighted hardest. Evidence that contradicts the candidate subtracts; nothing is clamped to look better than it is.

The result is a ranked list in which each score is reconstructible: the audit record keeps every component, so a reviewer can see exactly why one candidate scored 0.62 and another 0.31. Candidates below the quality threshold fall away here — often all of them, which is WAIT working as designed.

Mechanism

The top-ranked candidate faces its final examination: score above threshold, confidence above the current posture's bar, reward-to-risk above the posture's minimum, and no standing veto from news, risk, or cost. Pass everything and the system commits — direction, thesis, confidence, and a quality record capturing uncertainty and how strongly the committee agreed.

Fail anything and the answer is WAIT, with the true cause preserved: "vetoed on execution cost" is never rewritten into something more flattering. Over many heartbeats these truthful causes become the system's most valuable diagnostic — the record of every trade it chose not to take, and why.

Mechanism

Takes the L10 envelope and applies decision-time knowledge: confidence, posture, and any news haircut scale the base volume down within the envelope — never up beyond it. The result snaps to the broker's volume step. If scaling lands below the broker's minimum volume, the answer is zero, and zero is final: ATLAS never rounds a sub-minimum size up to "make the trade happen."

Mechanism

From the L11 candidates, the stop-loss particle selects the protective level matching the winning scenario's invalidation logic, validates it on the correct side of entry, and verifies broker minimum-distance rules. The take-profit particle pairs it with a target, and the pair's final reward-to-risk is checked once more against the posture floor — geometry that drifted during selection cannot slip through.

An invalid or unverifiable stop is not patched with a default distance. It forces stand-down: a trade whose protection cannot be trusted is not protected, and does not open.

Mechanism

Assembles the wire instruction: direction, volume, entry, stop, target, slippage limit, a unique decision identity, and an expiry. The identity makes execution idempotent — a re-delivered instruction can never open a second position. The expiry makes it perishable — an instruction that arrives late executes as nothing, because the market it was written for is gone.

Its critical invariant is synchronisation: if any wire field cannot be produced validly, both the internal decision and the wire instruction become WAIT together. The system never records "we decided to buy" alongside a wire that says wait.

Mechanism

The server's pre-trade check re-verifies the finished instruction against hard policy caps — maximum volume, exposure, stop distances — as a fresh reading, independent of the pipeline that built it. Unsafe instructions are not quietly rewritten to fit; they stand down, loudly and on record.

Then the terminal applies its own, separately configured limits before touching the broker. This redundancy is deliberate: even a compromised or buggy server cannot push the terminal past its local caps. A refusal at either guard travels back as a structured report, so the server always learns its instruction did not execute — and why.