A
ATLAS
The Architecture of Edge
A signal engine that produces positive-expectancy intraday trades on NQ & ES — by reading the tape institutions leave behind, weighing evidence the way a prop desk would, and refusing to fire when the conditions don't support edge.
COLD ▸ HOT · NOW ▸ LIVE
◬
78 modules
Engine surface
⧉
~100 fields/fire
Substrate density
⟁
711 fires resolved
Closed-loop sample
⬢
21 cards · 4 layers
This atlas
Layer 1
SUBSTRATE
what the engine sees
Layer 2
PILLARS
how it thinks
Layer 3
DECISION
how it commits
●
Substrate L1
what the engine sees — raw market reality, classified
7 CARDS
Order Flow
01
Delta, CVD, absorption — read the tape, not the chart.
  • Bar-level delta + session CVD with divergence detection
  • Four-state absorption FSM: build → confirm → exhaust → reset
  • 3-tier aggressor: exchange flag → side → tick-rule fallback
  • Distinguishes passive replenishment from active consumption
Prevents False breakouts that move price without supportive flow.
⟨engine-order-flow.js · NQEliteAbsorption⟩
4-state FSM
Liquidity
02
Sweeps, traps, and the levels institutions defend.
  • Stop-hunt taxonomy: sweep+reclaim vs sweep+continuation
  • Trap detection on failed breakouts (acceptance test inside the level)
  • Key-level frame: prior H/L, value area, opening-range extremes
  • Per-symbol pause halts fires when level structure breaks down
Prevents Chasing into liquidity pools engineered to absorb retail.
⟨stop-hunt-panel · pipeline sweep-state⟩
±1 tick dedup
Market Profile
03
Where the auction agreed, where it didn't.
  • Initial Balance tiering with ATR-relative thresholds
  • VPOC migration + value-area shifts as bias frame
  • Anchored VWAPs from session opens + key events
  • Prior-day H/L persisted across sessions (bridge-restart safe)
Prevents Mean-reversion trades against a developing-value migration.
⟨engine-ib-* · poc-tracker · vwap-context⟩
Dalton framework
Whale & OI
04
Large-trader footprints retail can't see.
  • Aggressor-streak bursts as tape-aggression signal
  • OI overnight-delta as institutional-positioning shift
  • Real DOM depth from dxFeed AGGREGATE Order events
  • Whale-defended pullback as Tier-3b setup confirmation
Prevents Counter-trading positions large players are actively defending.
⟨whale-tracker · oi-tracker · aggressor-streak⟩
dxFeed depth
Cross-Market
05
Four-symbol consensus on who's leading and how.
  • NQ ↔ ES ↔ YM ↔ RTY regime classification per bar
  • Leadership: tech-led · broad-led · smallcap-led · divergent
  • SMT divergence (Smart Money Technique lagging-leg)
  • Quiet detection: cross-symbol compression as first-class signal
Prevents Strong NQ continuation while ES rolls and YM diverges.
⟨engine-cross-index · noa-cross-market⟩
4-symbol fabric
Macro Context
06
What the macro layer is doing to your micro setup.
  • VIX / DXY daily slices with regime buckets
  • Event proximity flags: FOMC · NFP · CPI
  • News-risk gate suppresses fires inside event windows
  • DXY-direction conditioning — same setup, opposite verdict
Prevents Firing inside FOMC drift, or fighting the dollar's structural lean.
⟨engine-news-risk · macro_daily.json⟩
VIX · DXY · event
HTF · Volatility
07
The regime you're operating inside, not just this bar.
  • Daily ATR(14) → required-RR scaling (no fixed points)
  • Day regime: trend · chop · reversal × low · normal · high vol
  • Asia chop bypass — known low-volume carve-out
  • Daily-loss circuit breaker enforces session-level risk
Prevents Applying chop-regime tactics inside trend-regime tape.
⟨daily-risk-gate · ib-day-snapshot⟩
3×3 regime grid
◆
Pillars L2
how it thinks — evidence weighed the way a prop desk weighs it
6 CARDS
Institutional Confluence
08
How the engine weighs evidence — the canonical aggregator.
  • Multi-pillar score: flow + structure + location + context
  • Per-evaluator weights tuned per setup, not globally
  • Suppression formula penalizes minority pillars vs dominant ones
  • One source of truth — no shadow re-implementations allowed
Prevents Confluence inflation from double-counting correlated evaluators.
⟨rr-confluence · production-context⟩
minority × (1−d/140)
Setup Catalog
09
~36 setups · 8 families · each with a literature anchor.
  • Every family has academic + practitioner backing
  • Dedup at emission: 15 setups ≈ 10 distinct bets
  • Per-setup graduation: ≥30 universal / ≥20 with literature prior
  • Family-affinity reweighting documented, gated on data
Prevents Shipping correlated copies of the same setup under different names.
⟨setup-classification · docs/setup_theses.md⟩
8 families
Bias Arbitrator
10
CONTINUATION vs REVERSAL — and the expensive FALLBACK path.
  • Per-fire arbitration between continuation and reversal composites
  • FALLBACK costs ~1.6R vs clean alignment (measured)
  • Per-direction calibration — LONG and SHORT not symmetric
  • Bias style classified before risk plan, not after
Prevents Composite bias from contaminating clean directional signal.
⟨engine-bias-arbitrator.js⟩
Δ −1.6R fallback
Phase F · Risk Plan
11
The layer that overrides setup enthusiasm.
  • Gate 9: RR inflation cap — NQ 2.5 / ES 4.0
  • Gate 9b: Regime Fit — hard block in opposing trend
  • Daily-loss circuit breaker enforces session boundaries
  • Advisory in Hot Run · enforcement re-armed for Live
Prevents 5R "lottery ticket" sizing pulled from intuition not distribution.
⟨daily-risk-gate · l4-risk-advisor⟩
RR cap NQ 2.5 / ES 4.0
Phase G · Additive Setups
12
Four new families anchored in published research.
  • G1 NQ↔ES SMT lagging-leg (ICT + pairs-trading)
  • G2 First-5-Min Gap Halfback Fade (Edgeful + Dalton)
  • G3 Overnight Inventory Continuation (NY Fed sr917)
  • G3b London-Sweep-Then-NY Failed-Breakout
Prevents Setup proliferation untethered from published evidence.
⟨production-context setup definitions⟩
NY Fed sr917
Cadence & Pause
13
Knows when to be quiet.
  • Cadence gate throttles fire frequency per symbol & setup
  • Per-symbol pause halts emission without halting observation
  • Earned-silence: substrate keeps writing while UI stays still
  • Anti-overfire — a missed signal beats a noisy one
Prevents Spam-fires that destroy operator attention budget.
⟨cadence-gate · symbol-pause⟩
silence = information
▲
Decision L3
how it commits — filters, modes, plans, and the two learning loops
4 CARDS
L1 / L2 / L3 Filter Stack
14
Three filters, three rejection criteria, zero silent drops.
  • L1 · structural validity — entry, stop, risk-coherent
  • L2 · confluence threshold + pillar-minimum
  • L3 · hero-filter + dedup + per-symbol pause
  • Every rejection writes to substrate — full attribution trail
Prevents Invalid plans reaching trader UI · silent attrition that hides bugs.
⟨engine-pipeline.js · filter stack⟩
3 layers · 0 silent drops
EV Decision · Run-Mode Aware
15
Cold · Hot · Live — three modes, one engine, no confusion.
  • COLD · shadow collection, zero UI emission
  • HOT · real signals, manual decision, 3-stream capture
  • LIVE · auto-execution — only graduated setups go live
  • Mode flag gates substrate writes (silent dormancy was a real bug)
Prevents Live capital on unproven setups · mode-confusion data contamination.
⟨engine-ev-decision.js · EV_PHASE.mode⟩
3-mode pipeline
Multi-Leg Trade Plan
16
50% partial + 50% runner — two outcomes per fire.
  • buildStructuredPlan emits legs[] per fire
  • evCheckOutcomes walks legs independently
  • Trader exits captured separately from engine hypothetical
  • Two R streams: engine realizedR vs trader rResultActual
Prevents Collapsing partial+runner into a single R that hides variance.
⟨engine-trade-management.js · Step 4a⟩
50% + 50% legs
The Two Loops
17
Engine learns from itself and from the human delta.
  • Loop 1 · engine — outcome tracker walks every fire → actualR vs engineR
  • Loop 2 · trader — journal ✓/✗ + exits → take-rate per setup
  • Substrate captures ~100 fields/fire feeding both loops
  • Success metric: take-rate on PROFITABLE_CANDIDATEs, WoW
Prevents Learning the engine's mistakes while ignoring the operator's.
⟨ev-decision tracker · engine-journal-*⟩
2 loops · 1 substrate
⬡
Cross-Cutting Discipline L4
the wisdom that protects the system from itself
4 CARDS
Decision Clock · 3-Frame
18
HTF + Signal + Trigger — three timing frames, one decision.
  • HTF clock · daily regime, ATR, value migration
  • Signal clock · setup arming inside the HTF frame
  • Trigger clock · bar-close commitment with anti-flicker stabilizer
  • Current engine has trigger; Step 4b makes all three explicit
Prevents Bar-close decisions blind to higher-frame regime context.
⟨decision-clock · render-stabilizer⟩
3 clocks · 1 commit
Canonical-Source Principle
19
Wire to the canonical engine source, not the convenient one.
  • Every UI value reads from the same upstream the old renderer used
  • Memory-3 #7 — enforced architecturally, not by policy
  • Four documented bypass bugs caught during a 2026-05-16 swap
  • A blank panel beats a wrong one
Prevents UI drift where numbers look populated but mean the wrong thing.
⟨buildLiveEngineExplanation() · pattern⟩
4 bypass bugs caught
Discipline Lock
20
No tuning mid-measurement window.
  • Lock through ≥2026-06-14 (≥2 weeks post-bridge resume)
  • Exception: correctness-only fixes — broken data, definitional bugs
  • Not gate tuning · not new setups · not classifier refactors
  • Confounded weeks waste the only thing you can't manufacture: time
Prevents PF-0.92 lesson repeat — calibrating before distribution speaks.
⟨ROADMAP · DISCIPLINE LOCK⟩
until 2026-06-14
Falsification Mandate
21
Bad evidence = removal. Not "tune harder."
  • Frank's rule, codified architecturally for every surgery candidate
  • Each candidate carries an explicit kill condition
  • 4-week post-bridge cohort fails CIs → candidate dropped, not rescued
  • Applies to NOA layer too — macro accuracy gate, take-rate gate
Prevents Sunk-cost preservation of features the data has already killed.
⟨ROADMAP §LATER S1–S12 · noa doctrine⟩
12 kill conditions

Read this carefully. The engine's job is not to be right about every fire. It is to fire only when the institutional reading of the tape, the regime above it, and the discipline around it all agree — and to learn faster than the market changes. Every card above answers one of three questions: what does it see, how does it think, how does it commit. The cross-cutting cards answer a fourth — how does it refuse to lie to itself.

HOT RUN · PHASE A
BRIDGE OFF · awaiting dxFeed
ENGINE LOCKED · until 2026-06-14
78 modules · ~100 fields/fire
--:--:-- UTC