Release Radar
Changelog
Continuous improvements to your trading edge. Every feature, fix, and optimization — timestamped and documented.
40
Releases
247
Features
78
Days Active
Latest
January 25•v1.44.0
Enterprise Resilience Suite - Production Stability
Exchange Circuit Breaker—CLOSED/OPEN/HALF_OPEN states for exchange API protection. Per-endpoint tracking with configurable failure thresholds (default: 5 failures). Auto-reset after 60s cooldown. Thread-safe with async support.
Data Freshness Guard—Prevents trading on stale market data. Thresholds: orderbook (5s), price (3s), positions (5s), funding (300s), CVD (10s). Critical types (orderbook, price, positions) block trading if stale.
State Backup Manager—Redis → SQLite backup for crash recovery. Backs up positions, open_orders, last_reasoning, telemetry. Gzip compression for large payloads. Automatic restore on startup.
Risk Alerts System—VaR/drawdown monitoring with threshold alerts. Alert types: VAR_BREACH, DRAWDOWN_CRITICAL, CONSECUTIVE_LOSSES, POSITION_CONCENTRATION. Cooldowns prevent alert fatigue.
Exponential Backoff Retry—Handle transient API failures gracefully. Formula: min(base × 2^attempt + jitter, max_delay). Defaults: 3 retries, 1s base, 30s max. @retry_with_backoff decorator.
Market Data Validator—Sanity checks for anomalous data. Detects: price jumps >5%, zero/negative spreads, stale timestamps >30s, orderbook imbalance >95%, missing required fields.
January 25•v1.43.0
Institutional Analytics Suite - Risk Intelligence
Value at Risk (VaR)—Historical simulation VaR at 95%/99% confidence levels. Uses 47+ days of trade history. Conditional VaR (CVaR/Expected Shortfall) for tail risk measurement.
Drawdown Analysis—Real-time max drawdown tracking with underwater period detection. Peak-to-trough measurement with recovery tracking. Daily/weekly/monthly drawdown metrics.
Execution Quality Metrics—Trade-level slippage analysis, maker/taker ratio, fill rate tracking. Average execution cost per trade with fee breakdown.
Liquidity Depth Profile—Multi-level orderbook analysis at $10K, $50K, $100K, $500K depths. Bid/ask imbalance scoring with time-weighted averages.
Order Flow Analytics—CVD velocity tracking, aggressive buyer/seller ratio, large trade detection. Institutional flow signals with configurable thresholds.
API Endpoints—GET /api/v1/intelligence/risk-metrics, /liquidity, /order-flow. Real-time data served from telemetry with Redis caching.
January 24•v1.42.0
Regime-Adaptive System - Dynamic Market Intelligence
Unified Weighted Philosophy—Complete system overhaul based on "Uno solo sistema di pesi" (ONE unified weighted system). All hard blockers converted to weighted adjustments. NO parallel gate systems - everything flows through unified weighting.
5 Market Regimes with Economic Foundation—ACCUMULATION (smart money buying dips, favor LONG), DISTRIBUTION (smart money selling rallies, favor SHORT), TRENDING (clear momentum, follow direction), CHOPPING (no direction, mean-reversion only, high bar), CAPITULATION (extreme fear, DO NOT SHORT, wait for reversal).
Centralized Regime Parameters—New regime_params.py module with 20+ parameters per regime: edge_floor (0.08%-0.25%), conviction_adj (-30 to +5), size_mult (0.1-1.2x), cooldown_minutes (5-30), trailing_activation (0.2-0.6%), tp/sl_multipliers. All thresholds have economic rationale.
Direction-Aware Thresholds—Parameters vary by trade direction within each regime. ACCUMULATION: edge_floor_long=0.08% (easy), edge_floor_short=0.20% (hard). DISTRIBUTION: reversed. TRENDING: symmetric. CHOPPING: 0.25% both (very high bar).
LLM Regime Context Injection—When MRD confidence >= 50%, regime-specific guidance injected into Alpha prompts. Example: "REGIME: ACCUMULATION - Smart money is accumulating. Favor LONG trades. Buy dips. SHORT trades fighting smart money require higher conviction."
Guidelines Middleware Activated—conviction_adjustment, edge_adjustment, and size_reduction_pct from Guidelines Engine now APPLIED (previously logged but ignored). Feature flags: guidelines_conviction_adjustment_enabled, guidelines_edge_adjustment_enabled, guidelines_size_adjustment_enabled.
Regime-Aware Cooldowns—Cooldown durations scale with regime: TRENDING=1x (quick re-entry), ACCUMULATION/DISTRIBUTION=1x (normal), CHOPPING=3x (avoid whipsaw), CAPITULATION=6x (minimal trading). Per-symbol, post-win, and post-loss cooldowns all scale.
Config Parameters—New flags in brain_config: regime_gate_enabled, guidelines_conviction_adjustment_enabled, guidelines_edge_adjustment_enabled, guidelines_size_adjustment_enabled, regime_context_injection_enabled, regime_aware_cooldown_enabled. All true by default.
January 21•v1.41.0
Boss Daily Filter + Flip Anti-Whipsaw Protection
Boss as Daily Bias Filter—Boss transformed from weighted signal (27%) to strategic daily filter. Updates every 6 hours with macro H4/D1 analysis. When trading CONTRA to Boss bias: -25 conviction penalty. When NEUTRAL: 50% position size. When ALIGNED: maximum size allowed. Reduces noise from short-term fluctuations.
Flip Anti-Whipsaw Protection (v58)—Post-mortem Trade #1096: BTC LONG→SHORT flip in 25 seconds caused $5.49 loss. Signal Fusion flip-flopped from +0.163 to -0.284 in 25s. Solution: flip_neutral_skip_threshold increased 0.12→0.20 (need stronger signal to trigger flip).
Flip Hysteresis System—New flip_hysteresis_threshold=0.35 - to flip an EXISTING position, opposite signal must exceed 0.35 (not just 0.20). Prevents rapid direction changes when signal oscillates around threshold. Zone between 0.20-0.35 = "no flip zone" for existing positions.
Critical Bug Fixes (3)—Fixed datetime import scope error in flip_handler.py (module-level conflict). Fixed flip_context undefined variable. Fixed ThesisMonitor wrong method name (check_thesis → check_thesis_validity).
Reprice Abort + Momentum Gate (v56)—Post-mortem Trade #1089: BTC LONG with 6 reprices chasing falling price. Solution: After N reprices (2-3), abort execution instead of continuing. If price moves against by momentum_threshold_pct (0.03-0.05%), abort immediately. No fallback to market order.
Config Parameters—New params: boss_daily_filter_enabled (true), boss_daily_update_interval_hours (6), boss_daily_contra_penalty (-25), boss_daily_neutral_size_mult (0.5), flip_neutral_skip_threshold (0.20), flip_hysteresis_threshold (0.35), maker_max_reprices (2-3), momentum_threshold_pct (0.03-0.05%).
January 21•v1.40.0
V11.2 Scientific Audit - 4 Proposte per +3% EV
Kill Switch Factor (PROPOSTA 1)—Trasforma la somma pesata in validazione a cascata. Fattori critici (Regime, CVD, Boss) ora MOLTIPLICANO per 0.5 invece di sommare. Un trade deve essere perfetto per passare. Risultato atteso: Win Rate +10%.
ATR-Based Dynamic Sizing (PROPOSTA 2)—Stop Loss dinamico basato su volatilità reale. Formula: SL% = (ATR_14 × multiplier / price) × 100. In alta volatilità → SL più largo → size più piccola. In bassa volatilità → SL più stretto → size più grande. Riduzione perdite da "wicks" del 30%.
Sniper Maker Execution (PROPOSTA 3)—In regime CHOPPING o RANGING, forza ordini POST_ONLY (maker) invece di market. Timeout 30s, cancella se non fillato. Risparmia spread + fee taker (~0.05% a trade). Su 50 trade = 2.5% PnL "gratis".
Asymmetric Free Ride Exit (PROPOSTA 4)—Exit R-based che migliora drasticamente R:R ratio. TP1 a 0.5R chiude 33% della posizione, sposta SL a break-even. Il 67% rimanente corre risk-free con trailing stop. Permette profitto anche con 40% Win Rate.
Scientific Audit Analysis—Analisi ultimi 50 trade: Win Rate 40%, R:R 1.21, Kelly Fraction -0.0967 (negativo!). Diagnosi: "Death by a Thousand Cuts" - sistema operava in modo sub-ottimale. V11.2 inverte la rotta.
Config Parameters—Nuovi params: atr_based_sizing_enabled, atr_sizing_multiplier (2.0), atr_sizing_min_sl_pct (0.5), atr_sizing_max_sl_pct (5.0), sniper_maker_enabled, sniper_maker_regimes, asymmetric_free_ride_enabled, asymmetric_first_tp_r_multiple (0.5).
January 19•v1.39.0
Plan V3: Fee-First Architecture - Execution Consistency Revolution
Fee-First Execution Router—New ExecutionRouter module determines TIF (Alo/Ioc) based on market regime. CHOPPING = Alo (60s timeout), TRENDING = Ioc (immediate), ACCUMULATION/DISTRIBUTION = Alo (30s). Addresses root cause: 82% of losses came from fee leakage ($145.97 fees on $31.27 gross = -$114.65 net).
Golden Rule: close_tif == open_tif—Execution consistency enforced at system level. If position opens with Alo (maker), it MUST close with Alo. Prevents toxic combo (open=Alo + close=Ioc) that averaged -0.107%/trade across 159 trades.
Regime Scoring (Non-Blocking)—Replaced hard gates with weighted scoring. Regime orchestrator now returns regime_score (0-1), conviction_adjustment, and size_multiplier instead of blocking. Only capital protection can hard-block.
TransitionOpportunity Detection—Instead of blocking during unstable regimes, system now detects BREAKOUT/REVERSAL opportunities. Evaluates vol_expansion, cvd_alignment, orderbook_strength, and momentum_score during transitions.
RSI Regime Inversion—Context-aware RSI interpretation: TRENDING + RSI>70 = strength (boost LONG), CHOPPING + RSI>70 = overbought (boost SHORT). Eliminates false signals from regime-agnostic RSI thresholds.
Config Parameters—New v3 params: execution_router_enabled, regime_first_enabled, rsi_regime_inversion_enabled, transition_opportunity_enabled. All deprecated hard_block flags marked with warnings.
January 19•v1.38.0
Critical Entry Protection - Extreme RSI Block + Telemetry Fixes
Extreme RSI Hard Block—New safety gate that blocks SHORT when RSI < 20 (extreme oversold) and LONG when RSI > 80 (extreme overbought). Prevents trading against extreme RSI conditions where mean reversion probability is statistically high.
Signal Fusion Telemetry—Fixed signal_fusion data capture in open_context (V52 telemetry). All trades now record full signal fusion breakdown including direction scores, indicator alignment, and market regime at entry for comprehensive post-trade analysis.
FLIP Trade Logging—FLIP trades now capture complete open_context with signal_fusion, unified_gate_result, RSI, orderbook imbalance, and market state at entry. Enables full traceability for position reversal decisions.
Configurable RSI Thresholds—New dashboard-configurable params: extreme_rsi_gate_enabled (default true), extreme_rsi_short_block (default 20), extreme_rsi_long_block (default 80). FLIP trades can bypass this gate to catch reversals at extremes.
January 18•v1.37.0
Position Protection System - Close Hysteresis + Min Hold Time
Close Hysteresis—3-cycle grace period when Signal Fusion direction flips. CLOSE blocked for 15s to allow temporary noise to settle. Exceptions: emergency loss (>1.0% ROE) or position already in profit. Prevents premature exits at -0.12% to -0.15% ROE.
Min Hold Time Gatekeeper—Universal protection: ALL CLOSE requests blocked unless position age >= 5 minutes. Applies to AI-initiated closes with reason=OTHER that were bypassing thesis protection. Emergency loss or profit positions bypass this check.
Three Min Hold Mechanisms—System now has 3 distinct protections: (1) min_hold_minutes (20min) - cooldown between TRADES for anti-churn; (2) thesis_exit_min_hold_minutes (5min) - thesis-triggered CLOSE only; (3) close_min_hold_minutes (5min) - universal CLOSE gatekeeper.
Adaptive Shark 5-Scenario System—Market classification into 5 scenarios: CONSOLIDATION (range-bound, fade extremes), BREAKOUT (volume spike, ride momentum), TRENDING (sustained move, trail stops), REVERSAL (exhaustion signals, counter-trade), NEWS (macro events, reduce size). Each scenario has specific entry/exit rules.
Emergency Loss Bypass—Both Close Hysteresis and Min Hold Time allow immediate exit if position loses >1.0% ROE (configurable via close_min_hold_emergency_loss_pct). Capital preservation takes priority over position protection.
Profit Position Bypass—Positions in profit (ROE > 0) can close immediately without waiting for min hold time. Allows locking in gains while protecting underwater positions from noise-induced exits.
Decision Validator Integration—New _check_min_hold_time() method in decision_validator.py runs for ALL CLOSE actions. Returns ValidationResult with reason=MIN_HOLD_TIME when blocking, providing full audit trail.
Config Parameters—New dashboard-configurable params: close_min_hold_enabled (default true), close_min_hold_minutes (default 5.0), close_min_hold_emergency_loss_pct (default 1.0). All stored in risk_presets.json and loaded via config/loader.py.
January 17•v1.36.0
EV Ledger + Chasing Limit Execution + V18 Council of Experts
EV Ledger System—Complete calibration tracking for predicted vs realized edge. 5230+ lines of enterprise code tracking every trade's expected_edge, realized_edge, calibration_ratio, and strategy performance. SQLite + API endpoints for deep analytics.
Calibration API Endpoints—GET /api/v1/ev/calibration/summary (overall calibration health), GET /api/v1/ev/report/daily (daily breakdown), GET /api/v1/ev/status (system status). Enterprise-grade EV calibration analytics.
EV Snapshot Capture—Every trade captures full market state at entry: signal_fusion_score, unified_gate_score, boss_conviction, orderbook_imbalance, cvd, funding_rate, fear_greed. Enables post-trade analysis.
Chasing Limit Execution—New execution mode: places ALO (Add Liquidity Only) at best bid/ask, chases price up to 5s if not filled, falls back to IOC market. Saves taker fees while ensuring execution.
Cancel+Place Safety—modify_order() intentionally returns error to force safer cancel+place flow in execution_manager. Prevents partial fill race conditions on order modifications.
V18 "Council of Experts" Prompt—Multi-persona AI architecture: Sniper (microstructure + execution timing), Strategist (trend + regime), Guardian (risk + position), CIO (synthesis + decision). Each expert analyzes then synthesizes into final recommendation.
Structured Analysis Output—Alpha must output <analysis></analysis> block before JSON decision. Explicit reasoning traces for each expert perspective. Full explainability for every trading decision.
EV Calibration Metrics—Tracks mean_predicted_edge, mean_realized_edge, calibration_ratio (target: 0.8-1.2), confidence_intervals, and bias detection. Identifies systematic over/under-estimation.
January 15•v1.35.0
Unified Gates V7.1 - TRAP Gate + CVD Velocity (Bear Trap Prevention)
CVD Velocity Factor (12% weight)—Measures RATE OF CHANGE in CVD (USD/minute), not just the level. When CVD velocity >= 5M/min (extreme buying), SHORT score = 0.0 → BLOCKED. Case study: CVD -2.5M→+25M in 5 min would have blocked the $9 bear trap loss.
TRAP Gate V2 (12% weight)—Process-based trap detection with 3 analyzers: (1) Absorption - high volume but price stalls; (2) Momentum Exhaustion - slowing price velocity; (3) Orderbook Shift - sudden imbalance change. Each analyzer returns 0-1 score.
15 Balanced Factors—Expanded from 6 to 15 weighted factors. Primary: TRAP (12%), CVD Velocity (12%), Orderbook (12%), Whale (9%), CVD Level (8%). Secondary: RSI, MRD, Funding, LCP, Technicals, Time, Rebound, Pullback, Fear&Greed.
Weight Rebalancing—Redistributed weights to emphasize LEADING indicators (CVD velocity, TRAP) over LAGGING indicators (technicals, regime). Total still sums to 1.0.
SignalFusion CVD Velocity—cvd_velocity_per_min field added to SignalFusionResult dataclass. Calculated from CVD history over configurable time window (default 5 cycles).
Validation Manager Integration—evaluate_unified_short_gate() now receives cvd_velocity_per_min from both microstructure data and SignalFusion fallback.
Bear Trap Case Study—Trade #992 (Jan 14): Entry at 97425, MTF BULLISH 4/4, CVD was -2.5M but shifted to +25M in 5 min. Result: -$4.59 loss in 10.7 min. V7.1 would have detected the +5.5M/min velocity and blocked.
Configurable Thresholds—trap_cvd_velocity_extreme_pct: 5M, trap_absorption_threshold: 0.3, trap_momentum_exhaustion_threshold: 0.5. All in risk_presets.json.
January 12•v1.34.0
Unified Entry Gate V2 - Consolidated Architecture
Unified Entry Gate V2—Consolidated 20+ gate files into single unified_entry_gate.py. Principle: "tutti devono lavorare per l'obbiettivo comune" - everyone works toward the common goal with weights, no hard blocks.
Critical Gates (Hard Blockers)—Only 3 gates can hard-block: circuit_breaker (capital protection), edge_floor (positive EV required), spread (acceptable execution cost). Everything else is a weighted penalty.
Direction-Aware Scoring—Each factor produces 0.0-1.0 score based on trade direction. RSI 80 + LONG = low score (overbought), RSI 80 + SHORT = high score. Factors: rsi, cvd, orderbook, funding, trend, regime, signal_fusion, fear_greed, whale_flow, lcp.
Weighted Conviction Penalty—Final score = weighted average of all factors. conviction_penalty = (1 - score) * penalty_multiplier. Marginal setups trade with reduced conviction instead of being blocked.
Feature Flag Rollout—use_unified_entry_gate_v2 flag enables new system. Default: false for safe testing. When enabled, skips old unified_short/long gates and orchestrator.
Atomic Flip PnL Fix—Fixed atomic_flip to properly close old trade row with real PnL (including accumulated partial profits from Redis) and open new trade row for flipped position.
Design Document—Full architecture documented in docs/architecture/unified_gate_v2_design.md. Migration path: shadow mode → gradual rollout → full adoption.
January 9•v1.33.0
EV Optimization Suite - 10 Critical Bug Fixes
Atomic Flip Execution—Single 2x order to flip positions instead of CLOSE→WAIT→OPEN sequence. Saves fees, faster execution, no missed entries during volatility.
Signal-Adverse Exit Fix (P0)—LONG positions were exiting when fusion_score < 0.15, which included +0.10 (slightly bullish!). Now correctly uses <= -0.15 threshold.
Unified SHORT Gate Guard (P0)—Legacy orderbook_veto_short was running even when unified_short_gate was enabled, causing double-blocking. Now properly guarded.
ATR-Based Trailing Stop—Dynamic trailing distance based on Average True Range: 2x ATR multiplier with configurable min 0.3% / max 2.0% bounds.
Gate Conviction Applied (P1)—gate_result.total_conviction_adjustment was logged but NEVER applied to actual conviction. Now properly modifies decision conviction.
Neutral Dilution Prevention—Signal Fusion was dividing by total_weight including neutral signals, diluting the score. Now only counts directional signals (|direction| >= 0.1).
Cross-Tenant Isolation (P1)—FlipStrategy._flip_state was keyed by symbol only, affecting all tenants. Now uses tenant_id:symbol for proper multi-tenant isolation.
MRD Confidence Scaling—During regime transitions, detected_confidence was used without adjustment. Now scales with pending_count/min_persistence ratio.
`or 0.5` Parsing Fix—Signal Fusion used `x or 0.5` which returns 0.5 when x=0.0 (valid value). Changed to `x if x is not None else 0.5`.
FlipStrategy Logging Fix—Flip reason always said "Signal Fusion triggered" even when orderbook fallback was used. Now correctly reports ORDERBOOK_FALLBACK.
January 9•v1.32.0
Economic Calendar Intelligence - Macro Event Awareness
Macro Event Calendar—Daily economic event tracking (NFP, CPI, FOMC, Jobless Claims, Fed speeches). Events fetched at 06:00 CET via Finnhub API or manual input.
Morning Analysis (Gemini 3.0 Pro)—At 09:00 CET, Gemini 3.0 Pro Preview analyzes ALL events ONCE for the day. Creates bullish/bearish scenarios with key thresholds (e.g., "Jobless >220K = bullish BTC").
Pre-Event Caution Mode—30 minutes before high-impact events: position size reduced 50%, conviction requirement +10, edge requirement +0.1%. 5 minutes before: trading paused entirely.
Scenario-Based Reaction—Post-event, Flash compares actual results to Pro morning scenarios for immediate directional bias. No waiting for slow research - scenarios pre-computed.
Daily Risk Assessment—Morning analysis produces overall_risk (low/medium/high) for the day. High-risk days automatically reduce size 30% and boost conviction requirements.
Event Impact in Prompt—Alpha receives formatted calendar alerts with upcoming events, Pro scenarios, and key thresholds. Full context for informed decisions.
Dual-Model Architecture—Pro (expensive, smart) runs ONCE at 09:00 for deep analysis. Flash (fast, cheap) runs every cycle with Pro scenarios loaded. Cost-efficient macro awareness.
Today Example—Jobless Claims 14:30 CET: "If >220K = bullish BTC (weak labor = Fed cuts)", Fed Waller 16:00: "Watch for restrictive vs easing keywords".
January 9•v1.31.0
Flip Strategy & Adaptive Risk Controls
Flip Strategy V2—Continuous position flipping system - keeps exactly ONE position open at all times. When market reverses, closes current position and immediately opens opposite direction.
Flush Detection—Prevents flipping LONG→SHORT during extreme capitulation (imbalance < 20%). V-bounce detection avoids getting trapped shorting the bottom of flushes.
Anti-Whipsaw Protection—Minimum 3 cycles and 180 seconds between flips. Prevents rapid flip-flop from noise while allowing legitimate reversals.
Flip Cost Validation—Ensures edge > 2x flip costs before executing. Accounts for taker fees (0.035%), slippage (0.05%), and spread.
Loss Close Prevention Widened—RISK_004 threshold changed from -0.15% to -1.0% ROE. Allows flip strategy to exit small losses while protecting against panic closes on larger drawdowns.
Post-Restart Cooldown Reduced—STRAT_007 reduced from 3 cycles (6 min) to 1 cycle (2 min). Faster market response after service restarts while still allowing data initialization.
SHORT Parameter Wiring—Fixed wiring gaps: short_stop_loss_mult (0.8x), short_take_profit_mult (0.6x), short_max_leverage_mult (0.8x) now properly applied to SHORT trades.
Boss Size Modifier—Boss guidance affects SIZE only, not direction. Aligned: 125%, Neutral: 100%, Opposing: 75%. Direction decisions delegated to unified gates.
January 8•v1.30.0
Momentum Detection - Early Entry System
MomentumDetector Engine—Enterprise-grade LEADING indicator for early entry timing. Detects price velocity, volume spikes, CVD acceleration, and orderbook sweeps BEFORE lagging indicators confirm.
ROC (Rate of Change)—Price velocity tracking with configurable thresholds: WEAK (0.2%), MODERATE (0.4%), STRONG (0.6%), EXPLOSIVE (1.0%). 5-minute rolling window.
Volume Spike Detection—Unusual volume activity detection with ratios to average: 1.3x (weak) to 4x (explosive). Confirms momentum with participation.
CVD Acceleration—Tracks rate of change in Cumulative Volume Delta. Thresholds from $50K (weak) to $500K (explosive) in 3-min window.
Orderbook Sweep Detection—Detects rapid consumption of price levels (3+ levels in 30 sec). Early signal for aggressive positioning.
Signal Fusion Integration—Momentum added as 7th signal with 15% base weight in MICRO timeframe. When STRONG/EXPLOSIVE, weight dynamically boosted +10-20%.
Dynamic Weight Override—Strong momentum reduces lagging indicators (Boss, Whale, Polymarket) proportionally. Prioritizes LEADING signals during fast moves.
Time Decay System—Momentum signals valid for 5 minutes with exponential decay. Fresh signals weighted higher, stale signals fade gracefully.
Entry Urgency Scoring—Returns "immediate", "normal", or "wait" based on momentum strength. Size multiplier suggestions for position sizing.
Solves Late Entry Problem—Previous issue: waiting for 3+ lagging signals = entering at 70-80% of move. Momentum detects moves EARLY for better R:R.
January 6•v1.29.0
Polymarket Integration - Prediction Market Macro Overlay
Polymarket Signal Fusion—New 6th signal in Signal Fusion Engine (10% weight). Integrates prediction market probabilities for BTC price targets, FOMC rates, and crypto catalysts.
Gamma API Integration—Real-time data from Polymarket's Gamma API. Fetches crypto-tagged markets: Bitcoin milestones ($100k, $150k, $75k), Ethereum targets, macro events.
Unified Signal Weights—Rebalanced weights: Boss (28%), Orderbook (25%), MRD (18%), Whale (14%), Polymarket (10%), Options (5%). Sum = 1.0 with normalization.
Risk Temperature Scoring—Computes macro_risk_temperature (0-1) from bullish/bearish probabilities. Low risk + high bullish = LONG friendly, high risk + bearish = SHORT friendly.
Event Volatility Alerts—Flags when 2+ high-volume catalysts approaching within 7 days. Reduces conviction during uncertain macro periods.
Direction Score Computation—Converts btc_bullish, btc_bearish, risk_temp into normalized direction (-1 to +1). Confidence capped at 0.8 (prediction markets not perfect).
Removed Standalone Adjustment—Previous direct conviction adjustment replaced with unified Signal Fusion integration. No more disconnected polymarket penalty/boost - all signals weighted equally.
5-Minute Cache—Prediction market data cached for 5 minutes to reduce API load while maintaining freshness. Fetched in gather_context() alongside other signals.
January 5•v1.28.0
Parlant Guidelines Engine & Signal Fusion Fix
Parlant Guidelines Engine—55 YAML-based trading guidelines replacing hardcoded rules. Categories: risk (14), signals (13), regimes (13), strategy (15). Full explainability traces.
GuidelineEngine Class—Core engine with condition parsing, priority evaluation, and conviction/edge adjustments. Supports operators: <, >, <=, >=, ==, !=, IN, NOT IN, BETWEEN.
Context-Aware Matching—Guidelines evaluated against 40+ context fields: HLSI, MRD regime, LCP risk, signals, position state, time, exchange state.
Preset Overrides—Each guideline can have different thresholds per risk preset (safe/balanced/growth). Zero hardcoding in Python code.
RISK_006 Bug Fix—Fixed OR condition parsing bug that was blocking ALL trades. Disabled guideline pending split into multiple AND-based rules.
Signal Fusion Logging Fix—Fixed structlog keyword conflict causing signal_fusion_failed errors. Weights normalization now logs correctly.
Nuclear Pause Reset—Added ability to reset cooling period after consecutive losses. Bot status properly updates to RUNNING.
Prompt Size Reduction—Context-aware guideline injection reduces LLM prompt size by 40-60% while maintaining decision quality.
January 4•v1.27.0
System Hardening & Observability Sprint
MRD Regime Persistence—Requires 3 consecutive cycles before confirming regime change. Prevents oscillation between ACCUMULATION/DISTRIBUTION/CHOPPING.
Signal Fusion Dynamic Weights—Weights auto-adjust based on regime: MRD +10% in DISTRIBUTION, Options +5% in high volatility, Whale +10% when flow >$100M.
Whale Feed Bug Fix—Fixed misclassification of beacon staking as exchange inflow. Added DeFi protocol detection (Aave, Compound, Lido).
Execution Slippage Tracking—Every fill now logs slippage in basis points with maker/taker distinction. Stats persisted in Redis with 24h TTL.
Orderbook Quality Penalties—Moderate skew (50-200x): -5 conviction. Severe skew (200-500x): -10 conviction. Uses fresh skewed data instead of stale snapshots.
Maker Timeout Increase—Extended from 6s to 10s to reduce IOC fallback frequency and improve maker fill rates.
Regime Change Telegram Alerts—Automatic notification when MRD regime transitions (5-minute cooldown to prevent spam).
Whale Health Monitoring—Tracks consecutive fetch failures and parsing errors. Alerts when feed becomes unreliable.
January 3•v1.26.0
Unified Direction Gates V2 - Zero Hardcoding
Unified Gate Architecture V2—9-factor weighted scoring system replacing ALL scattered direction rules. RSI, CVD, orderbook, fear_greed, funding, whale, LCP, MRD, technicals - all in one gate.
Whale Flow Integration—Net whale flow (inflow/outflow) now scored: -$100M outflow = bullish for SHORT, +$100M inflow = bullish for LONG. Real institutional money flow.
LCP (Liquidation Cascade) Scoring—avoid_long/avoid_short signals converted to scores. CASCADE_RISK levels (LOW/MEDIUM/HIGH/CRITICAL) affect conviction penalty.
MRD (Micro Regime) Scoring—ACCUMULATION = great for LONG (1.0), DISTRIBUTION = great for SHORT (1.0), CHOPPING = neutral (0.5) with trading_bias consideration.
Technicals Integration—M15/H1/H4 EMA trends scored together. All BULLISH = 1.0, All BEARISH = 0.2. Intentionally low weight (5%) - technicals are lagging indicators.
Configurable Weights—unified_long_gate_weights and unified_short_gate_weights in dashboard. Default: orderbook 20%, cvd 15%, whale 15%, lcp 10%, mrd 10%, rsi 10%, funding 10%, fear_greed 5%, technicals 5%.
Prompt Cleanup—Removed hardcoded SHORT TRADING RULES, HLSI SHORT rules, and LCP avoid rules from AI prompts. Direction filtering now 100% programmatic.
Score-Based Penalties—penalty = (1 - score) × multiplier. Score 0.40+ passes, below 0.40 blocks. No more binary vetoes - gradual conviction adjustment.
January 2•v1.25.0
Signal Fusion Engine & Alpha Clarity
Signal Fusion Engine—Academic-grade weighted consensus system combining 5 signal sources: Boss (30%), Orderbook (25%), MRD (20%), Whale (15%), Options (10%).
Trading Signal Summary—Human-readable signal digest injected into Alpha prompts. No more raw JSON confusion—clear directional guidance with conflict warnings.
Conflict Detection—Automatic detection when signals disagree (e.g., Whale BULLISH vs MRD Distribution). Alpha now sees "HIGH CONFLICT" warnings before deciding.
Conviction Adjustments—+20 points when all 5 signals align, +15 for 4/5, -10 penalty for 2+ conflicting signals. Size multiplier: 1.1x consensus, 0.7x conflict.
Alpha Receives All Signals—Critical fix: Alpha now sees whale_alerts, alpha_intelligence, signal_fusion, boss_guidance—previously these were only passed to Boss agent.
Smart WAIT Decisions—Alpha correctly identifies signal conflicts and waits for clarity instead of forcing trades. "Technicals bullish but MRD Distribution + Bearish Orderbook = WAIT."
Direction Score—Normalized -1.0 to +1.0 directional score. >0.20 = LONG, <-0.20 = SHORT, between = NEUTRAL. Eliminates ambiguity.
MRD None Handler—Fixed confidence *= None crash in MRD parsing. Now uses `or 0.5` fallback for None values.
Missing a feature? Let us know