Gate 0 is the pipeline's first checkpoint and the only one that cannot be overridden by any downstream condition. It runs three independent checks in sequence. Gate 0 (Institutional Veto) scans for any reason the system has flagged as an institutional-level block — circuit breakers, exchange halts, or operator-set kill conditions. If any flag is set, the signal dies immediately with no appeal.

Gate 0a (Symbol Kill Switch) is a per-symbol manual toggle. During Phase 1 calibration, ES was locked in analysis-only mode — setups were computed and logged but never promoted to signal status. This allowed the engine to collect baseline data on ES before risking signal credibility. Gate 0b (Setup Quarantine) checks a static blacklist of setups that failed catastrophically in live data: session-ext-L/S lost -20.48R combined, vwap-bounce-L/S showed 0% win rate, ib-brk-S cost -4.42R. Quarantined setups continue computing (the engine still logs their confluence scores to the substrate for analysis), but they can never promote to SIGNAL state.

Gate 1 consults the market context analyzer to determine whether the market is open and in a valid, tradeable state. The severity of any block depends on the reason: a pre-market or post-close condition is a hard block (no point evaluating setups when the market is closed), while a data-quality degradation might only impose a soft penalty.

Gate 1.5 evaluates three separate confidence dimensions. 1.5a (State Confidence) requires the engine's own classification of the current market context to be at least 50% confident. Below that threshold, the engine doesn't trust its regime classification — it might be calling a "trending" market when it's actually rotating, and every downstream gate depends on that classification being at least plausible. 1.5b (Production Mode) penalizes when tradeMode = NONE — no sweep+trap pattern detected and no trend exception active. This means the engine has no high-confidence thesis about what the market is doing. 1.5c (Transition Risk) applies a proportional penalty when the market is transitioning between regimes (e.g., from range to trend), because regime boundaries are statistically where false signals cluster.

The Bias Arbitrator aggregates votes from five independent directional modules — each contributes a long score and a short score based on its own evidence. The arbitrator then applies a voting framework: if one side exceeds 55% and leads the other by at least 8%, a directional bias is declared. If both sides exceed 55% within 8% of each other, the state is CONFLICT — not uncertainty, but genuine contradictory evidence, which is worse for signal quality than having no bias at all.

The gate checks whether the proposed setup's direction aligns with the declared bias. A long setup in a SHORT bias environment gets a penalty. However, a critical exception exists: if there's an actionable sweep or absorption event, the bias check is skipped entirely. The logic is that liquidity events (stop hunts, institutional absorption at a level) override directional bias — a trapped short squeeze doesn't care what the trend says.

The CONTINUATION_FALLBACK state deserves attention: when bias direction matches but confidence is below the decisive threshold, the arbitrator falls back to a continuation assumption. Cohort attribution analysis found this fallback costs approximately -1.6R on average — it's a known weak spot that ships as a soft penalty rather than a hard block, awaiting more data to decide its fate.

Gate 2.6 (Market Story Fit) computes a 0-100 score measuring how well the proposed setup family matches the current auction narrative. A continuation setup in a trending market scores high; the same setup in a rotating, mean-reverting market scores low. Below 55, the signal takes a soft penalty. Below 35, with a raw confluence score below 60 and the market classified as CHOP_NO_EDGE, the gate escalates to a hard block — the thesis has no structural support.

Gate 2.7 (Chop Permissions) is the engine's most aggressive defensive gate. When the market story evaluator classifies the environment as SEVERE_NO_EDGE_CHOP — meaning no directional conviction, no institutional footprint, and no structural level nearby — the gate hard-blocks unless there's an actionable sweep or absorption event. Other chop states (MILD_CHOP, ROTATIONAL_CHOP) impose graduated score penalties but allow the signal to continue. The distinction matters: mild chop can produce legitimate mean-reversion setups at levels, but severe chop is a fee-burning machine.

Gate 3 (Regime Compatibility) enforces that continuation setups only fire in trend/expansion regimes, and reversion setups only fire in range/rotation regimes. In STRICT mode, a style mismatch is a hard block — a breakout setup in a ranging market simply cannot fire. In BALANCED mode, the mismatch imposes a soft penalty, allowing the setup to continue at a score disadvantage. Both modes hard-block all setups in low_participation regimes, because thin markets produce unreliable signals regardless of setup quality.

Gate 3.5 (Session Playbook) restricts which setup families are approved for each trading session. Asia (slow, level-driven) permits key-mag, exhaust-rev, and vpoc-mig — setups that work with clear levels and minimal flow. London permits ib-reject and trap-rev — setups that exploit the London open's stop-hunting patterns. NY (institutional flow-driven) permits ib-brk, ib-ext, and flow-surge — setups that ride directional momentum. Setups outside their playbook are either blocked (STRICT) or downgraded (BALANCED). The playbook is calibrated from historical outcome data grouped by session, not from theory about what "should" work.

Gate 3.6a (Gap Authority) checks whether a significant overnight gap conflicts with the setup's direction. A long setup when there's a large bearish gap (price opened well below prior close) faces headwind — the gap acts as an overhead magnet pulling price back. This is a soft block, not hard, because a confirmed sweep+trap can override the gap's authority.

Gate 3.6b (Breakout State) penalizes setups that fight a clean, untrapped breakout. If price has broken out of a range in one direction and no trap has been confirmed, setups pointing the opposite way are swimming upstream.

Gate 3.7 (Tier-1 Dominance) is a priority arbitration mechanism: when a tier-1 setup (London sweep, failed auction — the highest-conviction patterns in the system) is active, lower-tier setups in the opposite direction are suppressed. Even lower-tier setups in the same direction yield — the tier-1 setup takes the slot. This prevents signal clutter and ensures the best signal gets the attention.

Gate 3.8 (Whale Defense) reads DOM (Depth of Market) clustering patterns. When institutional-sized orders cluster at resistance above the current price (defense-resistance) or at support below (defense-support), the engine detects "whale defense" — large players actively protecting a level. If their defense opposes the setup's direction (e.g., massive sell-side clustering at resistance while a long setup tries to fire), the setup takes a soft penalty. If whale defense aligns with the setup, it's logged as a research note but doesn't adjust the score — alignment is expected, not bonus-worthy.

Gate 3.9 (OI Conviction) cross-references open interest dynamics with the setup's direction. A long setup firing during a SHORT_COVERING environment gets penalized: the rally looks like buying, but OI falling says it's shorts closing, not new longs entering — the rally is structural unwind, not genuine demand. Similarly, a short setup during LONG_CAPITULATION faces a penalty because the selloff is exhaustion, not new bearish conviction. Counter-trend penalties apply when longing against REAL_DOWNTREND or shorting against REAL_UPTREND. All OI gates are soft penalties because institutional positioning can be wrong, and a strong enough setup-level signal can override positioning headwinds.

Gate 4 (Retirement) checks the setup's rolling expectancy across its last 12 closed trades. Below -0.12R expectancy with at least 6 completed trades — the setup is benched. It continues computing (for substrate observation) but can never fire. Between -0.05R and -0.12R it's flagged as "Degrading" with a -5 priority penalty. Gate 5 (Confluence Threshold) is the raw score check: the setup's weighted confluence must reach a dynamic threshold (~68% base, adjusted by volume regime and qualification profile). BALANCED mode gets a -4 offset, making the threshold more permissive during data collection.

Gate 6 (Critical Fails) catches single-point-of-failure conditions. If any individual evaluator with weight ≥18 fails (scores zero), that's a critical fail. One critical fail is logged as a warning. Two or more critical fails trigger a hard block — the thesis has multiple load-bearing pillars that collapsed. Gate 7 (Entry Confirmation) requires setups below the bypass threshold (~75-80%) to show confirming price action, order flow, or liquidity signals before promoting. Above the bypass threshold, the confluence itself is confirmation enough.

Gate 8 (Risk & R:R) validates the stop/target geometry. R:R is calculated from the planned entry, stop, and target prices. The floor is 3.0:1 — below that, the setup goes to WATCH regardless of confluence. R:R tiers map to signal states: below 3.0:1 → WATCH, 3.0-3.5:1 → ARMED, above 3.5:1 → eligible for SIGNAL or PREMIUM_SIGNAL. This ensures that even the highest-confluence setup doesn't fire if the math doesn't work.

Gate 9 (R:R Inflation Cap) addresses a structural bias in the engine's trade planner: while historical resolved trades show a median R:R of ~1.5:1, the planner consistently produces 4.5:1 median projections. The cap would hard-block any signal where the projected R:R exceeds 2.5:1 for NQ or 4.0:1 for ES. Currently advisory-only: the engine logs what it would have blocked, but doesn't enforce, pending calibration data proving that the cap improves outcomes rather than just filtering high-conviction setups.

Gate 9b (Regime Fit Gate) layers a higher-timeframe regime check on top of Gate 3's style compatibility. It reads the daily regime bias and applies two sub-checks: a direction violation (shorting in a trend_up day) would be a hard block; a style violation (breakout setup in a range day) would be a soft block. Like Gate 9, it's advisory-only pending data proof. Gate 10 (Session Router) is permanently disabled. Out-of-sample validation on 840 historical trades showed it blocked profitable families more often than unprofitable ones — it destroyed net value. Kept in the codebase as an architectural placeholder and a reminder that "intuitive" filters can fail empirical validation.

After every gate has executed, the pipeline resolves the signal into one of six discrete states. SIGNAL means every gate passed and R:R is confirmed — this is the only state that can fire. ARMED means confluence is strong and most gates passed, but the signal is waiting for one remaining condition (typically R:R confirmation or entry timing). Armed signals are near-miss candidates: if the pending condition resolves, the signal promotes automatically on the next evaluation cycle.

WATCH is the passive observation state — the setup is being evaluated but isn't close enough to actionability. BLOCKED means at least one hard gate failed — structural rejection with no workaround. CONFLICT is a distinct state from BLOCKED: it means the directional evidence is genuinely split, and the engine refuses to pick a side. PAUSED is operator-initiated — the trader manually paused this symbol via the dashboard's per-symbol pause button. All six states, along with the specific gate that determined them, are written to the substrate for every evaluation cycle. The trader sees all of this on the dashboard and can act on contextual nuance the gate system can't encode.

Every setup in the engine — regardless of family — scores against the same 100-point budget divided into six categories. The budget is fixed but the distribution is not: each setup type allocates different weights to different conditions within each category, reflecting what matters for that specific thesis. An IB breakout weights trigger conditions heavily; an exhaustion reversal weights flow divergence conditions.

Location receives the largest allocation (25 points) because where price sits relative to key structural levels is the single strongest predictor of whether a setup succeeds. A perfect trigger at a meaningless level is noise; a mediocre trigger at a critical level is a trade. Regime (20) and Trigger (20) share the next tier — regime ensures the market environment supports the thesis, while the trigger is the specific event that creates the opportunity. Flow (15) confirms the thesis with real-time order flow evidence. R:R (10) scores the mathematical quality of the stop/target geometry. Risk Filters (10) are binary safety checks — session timing, stop-hunt clearance, book stability — that don't generate edge but prevent easily avoidable losses.

The engine operates in one of two qualification profiles that control selectivity across the entire pipeline. STRICT is sniper mode: one setup per symbol, minimum 2.2:1 R:R, no score offset, and regime/playbook mismatches are hard blocks. BALANCED is scouting mode: two setups per symbol can compete, 1.8:1 R:R floor, -4 score offset (lowering the effective confluence threshold), and regime/playbook mismatches impose soft penalties rather than hard blocks.

Phase A (current) runs BALANCED to maximize data collection. The engine intentionally fires more signals to build a statistical sample for each setup type. Once enough data accumulates to reliably distinguish setup quality, the profile will shift to STRICT — fewer signals, higher average quality, lower noise. The -4 score offset in BALANCED mode means a setup that needs 68 in STRICT only needs 64 in BALANCED. This isn't a quality compromise — it's deliberate observational permissiveness.

The Structural Breakout family trades the initial balance (IB — the range formed in the first hour of RTH) as a structural reference. IB Breakout (ib-brk) fires when price decisively exits the IB range. Its heaviest weight is ib_one_sided (22) — was the IB dominated by one side? A tight, one-sided IB that breaks out is institutional intent; a wide, balanced IB that breaks randomly is noise. IB Extension (ib-ext) fires after the breakout holds and extends. It shifts weight from the trigger to flow confirmation: delta_aligned (20) and mkt_order_tempo (14) together require proven, sustained flow in the breakout direction.

Breakout Retest (brk-ret) is the family's highest-conviction variant. It fires when a breakout level is retested and holds. The weight profile flips entirely: sweep_reclaim (20) and key_level_near (18) replace ib_one_sided — the thesis is no longer "the breakout happened" but "the breakout survived its first challenge." Absorption (14) confirms that aggressive selling at the retest level was absorbed by passive buyers. The not_lunch filter (8) exists because lunch-hour retests frequently fail due to thin liquidity, not structural weakness.

Mean reversion setups bet that price has moved too far from fair value and will snap back. The family shares two dominant signals: delta divergence (flow disagreeing with price direction) and absorption (price stopping despite aggressive hitting). These two conditions together answer the question: "Is the move running out of fuel?"

VWAP Deviation Snap (vwap-dev) has the family's most concentrated weight: vwap_extreme at 24 points. The entire thesis is "price extended ≥2σ from VWAP — mean reversion probability is high." Without extreme VWAP deviation, the setup doesn't exist. VWAP Bounce (vwap-bnc) similarly loads 24 on vwap_bounce_zone — the trade IS the bounce at VWAP. IB Fade (ib-fade) combines divergence (20) with one-sided IB (18) — the IB pushed hard one way, but flow says the push is exhausted. Value Area Fade (va-fade) anchors on naked VPOC (18) and key levels (14) — fading from a value area boundary that hasn't been visited yet is a high-quality mean-reversion thesis because the unvisited VPOC acts as a magnet.

This family trades failed moves — situations where a directional push exhausted itself, trapping participants on the wrong side. The thesis is contrarian: the losers' forced exits fuel the reversal. Exhaustion Reversal (exhaust) carries the heaviest single condition weight in the entire engine: delta_divergence at 26. The signal is unambiguous: aggressive flow pushed price to a level (key_level_near at 20), but the flow is dying (divergence) while passive absorption (18) holds the level. Three independent lines of evidence converge on one conclusion: the move is over.

Trap & Reverse (trap-rev) leads with sweep_reclaim (20) — the failed breakout IS the trade. Price pushed past a level, swept stops, and reclaimed. The trapped participants are now underwater, and their forced exits become your fuel. IB Rejection (ib-rej) combines one-sided IB (18) with divergence (20) — the IB pushed hard one way but the rejection needs both: the setup (the push) and the proof (the flow reversal). Session_match (6) gives a small bonus because IB rejections are most reliable when they happen during the approved session for the setup family.

This family trades institutional flow events — situations where the order book tells a clear directional story. Flow Surge (flow-surge) is the engine's most aggressive setup, requiring overwhelming unidirectional flow: delta_aligned at 24 + mkt_order_tempo at 20. Together, these two conditions require both the magnitude (massive delta in one direction) AND the participation (heavy institutional trade count). DOM alignment (14) adds a third confirmation: the depth-of-market book structure should support the direction. This is a momentum trade — ride the institutional wave.

Key Level Magnet (key-mag) trades the pull toward an unvisited structural level. Its heaviest weight is naked_vpoc (22) — the unvisited Volume Point of Control is the magnet itself, a price level where significant volume transacted previously but the current session hasn't reached yet. Key_level_near (18) and absorption (16) add structural and flow confirmation. Volume Migration Follow (vol-mig) tracks when the VPOC physically migrates — it shifts (20) toward a new level, and the engine follows with delta confirmation (18) and cross-market alignment (14). This is a trend-following variant: the auction itself is voting on the new fair value.

Phase G setups are specialist plays defined by unique structural triggers that don't exist in the core families. They carry the highest single-condition weights in the system (25-30 points) because each setup is its trigger — without the specific condition, the setup literally doesn't exist.

Gap Halfback Fade (gap-fade) loads 30 on its trigger (gap_inside_range) and 25 on first_5min — together these two conditions comprise 55% of the setup's total budget. The thesis: when the market opens with a gap that falls inside yesterday's range, and the first 5 minutes show reversal, the gap will fill at least 50% (halfback). London Sweep (ldn-sweep) at 28 triggers on the London session's characteristic stop-hunting pattern: price pushes above/below the Asian range to sweep stops, then reverses into the NY open. Overnight Continuation (ovn-cont) at 28 uses overnight VWAP (at_ovn_vwap, 18) as its anchor — the thesis is that the overnight direction established by Asia/London will continue into NY. SMT Divergence (smt-lag) at 25 fires when NQ and ES diverge structurally — one makes a new high/low while the other doesn't confirm, suggesting the leader is trapping participants.

The quarantine list is the engine's graveyard for setups that failed in live data. Each entry earned its place through objective performance collapse: session-ext-L/S lost a combined -20.48R (catastrophic), vwap-bounce-L/S showed 0% win rate across a meaningful sample (structurally broken), ib-brk-S lost -4.42R (short-side IB breakouts fail systematically). Quarantined setups are not deleted — they continue computing and writing to the substrate. If market conditions shift and the setup rehabilitates in the data, it can be unbenched. The distinction between quarantine and retirement is permanence: retired setups auto-bench at -0.12R expectancy and can auto-rehabilitate; quarantined setups require manual operator intervention to restore.

The session playbook maps each trading session to its approved setup families. Asia (pre-London, low participation) permits only level-based plays: key-mag, exhaust-rev, vpoc-mig. Breakout setups are banned — thin Asia liquidity produces false breakouts. London permits trap-based plays: ib-reject, trap-rev — because the London open characteristically sweeps Asia stops. NY gets the full momentum arsenal: ib-brk, ib-ext, flow-surge — because NY has the institutional flow to sustain breakouts.

The engine maintains a real-time ranked list of liquidity targets — price levels where stop orders are likely to cluster. Each target receives a composite score combining three factors: type base score, proximity adjustment, and session context bonus. The type hierarchy reflects institutional behavior: session extremes (Asia/London/OR highs and lows) score highest (34) because they're the most visible stop-cluster locations — virtually every retail trader places stops outside them. Prior-day extremes (PDH/PDL at 32) serve a similar role at daily scale.

The proximity formula adds up to 44 points based on how close the target is: max(0, 44 − distTicks × 0.55). This means a target at the current price gets +44, a target 40 ticks away gets +22, and a target 80 ticks away gets effectively zero. The decay rate of 0.55 per tick was calibrated to match the typical effective range of institutional sweep operations — stops too far away aren't getting swept in the current move. A session match bonus of +8 rewards targets that belong to the current session (Asia high targeted during Asia session) because same-session extremes are the freshest, most-watched levels. Range targets score only 6 base with an additional -20 penalty because intraday ranges are weak references — everyone sees them, but they lack the institutional significance of session extremes.

The order flow state machine is the engine's real-time interpretation of what the market is doing right now. Every bar, it evaluates 9 possible states in a fixed priority order — the first state whose conditions are met wins, and no further states are evaluated. This priority-first design prevents ambiguity: when multiple states could apply, the highest-conviction interpretation takes precedence.

Stale (15) checks first as a circuit breaker — if the latest flow data exceeds the age threshold, every other state is meaningless. Trap Flow (82) evaluates second and carries the highest confidence in the system: confirmed stop hunt with reclaim means institutional behavior — this is the most reliable signal the engine produces. Initiative Buy/Sell (78) require the tightest triple-confirmation: meaningful delta magnitude + CVD confirmation + efficient price travel in the same direction. Missing any one of the three drops the state.

CVD Divergence (72) catches a specific failure: price makes a new extreme but cumulative volume delta doesn't confirm — the "who's buying?" gap. Absorption (68) detects price stopping despite aggressive hitting, especially near a structural level. Exhaustion (62) is the "dying move" state: price is still traveling but delta has lost its meaningful threshold, and either volume or CVD is diverging. Delta Not Confirmed (48) captures the awkward middle: delta is meaningful but price/CVD don't agree. Noisy/Unusable (35) is the default — no conditions met, no interpretation possible.

Directional intent scoring aggregates five independent modules — each evaluates its own evidence and produces a long score and a short score. The scores are weighted and combined into a final directional call. The thresholds are intentionally stricter than the Bias Arbitrator (Gate 2.5): the side minimum is 62% (vs. 55%), and the edge requirement is 10% (vs. 8%). This dual-threshold design prevents two failure modes: declaring direction when evidence is thin (minimum check) and declaring direction when both sides have nearly equal evidence (edge check).

Beyond the binary long/short call, the system infers a directional mode — how the market is expressing its direction. ACCEPTANCE means price is being accepted in the direction (trending smoothly). PULLBACK means the direction holds but price is retracing (potential entry opportunity). TRAP_REVERSAL means a failed move trapped participants on the wrong side (highest conviction). The mode tells the setup system what kind of trade to look for: a pullback in an uptrend calls for continuation setups; a trap reversal calls for reversal setups.

The confidence floor at 60% flags marginal directional reads. A call that passes the math (62% side, 10% edge) but has low confidence (below 60%) is technically valid but unreliable — the engine logs it as a research note rather than a conviction signal. Confidence ranges from 45% to 75%, with the floor at 60% reflecting the point where cohort attribution data shows directional calls become predictive of setup outcomes.

Absorption detection answers a critical question: is someone passively absorbing aggressive flow at a level? When large limit orders silently eat incoming market orders without letting price move, it's invisible on the price chart but detectable in order flow — high volume + no price movement = passive absorption. The engine models this as a progressive state machine rather than a binary flag.

ACTIVE_ABSORPTION (58) is the initial detection: order flow shows absorption, but it's just a fact — someone is sitting on a level. It becomes interesting at TRAP_BUILDING (74) when structural evidence accumulates: a reclaim, a failed breakout, or proximity to a planned target suggests the absorption will hold. CONTINUATION_WARNING (66) is the default middle state — absorption is present but there's no reversal evidence yet, and the market might push through.

The highest states require convergence. REVERSAL_CONFIRMED at 86 needs reclaim or failed breakout PLUS opposite structure or hold — the full picture. At 84, it needs divergence plus at least one structural confirmation. TWO_REVERSAL_WARNINGS (80) fires on CVD divergence alone without structural confirmation — it's close to reversal conviction but lacks the physical proof. INVALIDATED (18) kills the thesis entirely: price accepted through the absorbed level. The passive defender lost — game over.

Raw absorption detection is noisy — in a fast market, the absorption side can appear to flip on every tick as aggressive flow alternates between bid and ask. The stability tracker is a low-pass filter that prevents meaningless side changes from polluting the engine's absorption state.

The tracker maintains a 60-second history window of scored observations. Each observation carries the quality score from the absorption detection module, and votes are weighted by score — a high-quality observation (e.g., REVERSAL_CONFIRMED at 86) outweighs several low-quality ones (ACTIVE_ABSORPTION at 58). To flip the declared side, the new side must satisfy three independent conditions simultaneously: (1) maintain its lead for at least 5 ticks, (2) hold for at least 10 seconds, and (3) achieve ≥60% weighted dominance in the recent vote window.

A fade grace period of 10 seconds holds the last declared side even when no new observations arrive — this prevents the tracker from going blank during brief pauses in flow. The reversal banner lasts 30 seconds: when a genuine side flip is confirmed, the engine displays a reversal alert on the dashboard for 30 seconds to ensure the trader notices the change. Observations below score 50 are dropped entirely — low-quality noise shouldn't influence the stability calculation.

Market order tempo is a per-bar measurement of institutional participation intensity. The engine counts both the number of trades (order frequency) and the total contracts (order size) in each bar, comparing them against fixed baselines derived from average RTH bar statistics: NQ baseline is 900 trades / 12,000 contracts per bar; ES baseline is 600 trades / 8,000 contracts.

Below 60% of baseline (WEAK), the market is dominated by retail noise — small orders, no institutional footprint. The engine responds by raising the confluence threshold by +4 points, demanding more evidence before firing. Above 150% of baseline (STRONG), institutions are actively participating. The engine lowers the threshold by -3 to -5 points because the heavy participation itself is a form of confirmation — you don't need as much structural evidence when the order book is shouting a direction.

The tempo check is separate from the volume regime (P2) and works at a different timescale. Volume regime uses daily percentile rank (macro context); tempo measures bar-by-bar participation (micro context). A STRONG tempo bar in a LOW volume regime means "today is quiet overall, but right now someone big showed up" — that's a relevant signal for the current bar's evaluation.