LLM Swarm Terminals

// Agent Timeline

When each agent fires.

The Swarm operates on a 6-hour intelligence cycle. Four times per day, the full chain executes in sequence. Specialist agents run on their own schedules.

6-Hour Intelligence Cycle

05:50 BST

Shadow Matrix

Missed alpha scanner. Analyses every signal the system rejected in the past 24 hours, calculates how much profit was left on the table, and feeds the forensics directly to the Swarm for filter correction.

Repeats 11:50 . 17:50 . 23:50

06:00 BST

Sentinel Broadcaster

Full 9-point clinical intelligence broadcast via Claude Opus 4.7. Reads 97 indicators across 88 symbols, dissects every trade from the past 6 hours, analyses crime scenes with full physics snapshots, and produces a comprehensive market observation. 4 broadcasts per day build the intelligence chain the Swarm reads before evolving.

Repeats 12:00 . 18:00 . 00:00

06:05 BST

Watchdog

Account health guardian. Tracks losing streaks from peak, monitors 20-trade rolling win rate. Push alerts at 10% loss. Emergency swarm rebuild triggered at 15% to force immediate strategy evolution.

Repeats 12:05 . 18:05 . 00:05

06:10 BST

Performance Snapshot

Captures full system state. Win rates, PnL percentages, equity curve data points, open positions with unrealised returns, and engine breakdown across all 4 trading engines.

Repeats 12:10 . 18:10 . 00:10

Strategic Agents

Every 48h, 2h prior

Weekly Digest

Compresses the prior days of sentinel broadcasts, trade data, and swarm evolution into a single strategic summary. Identifies recurring failure patterns, regime cycles, and evolution priorities. Runs 2 hours before each swarm cycle so the full intelligence is fresh and ready.

Every 48h

AI Swarm

50-generation evolution cycle. Paradigm rotation is phase-based: gens 1-10 MOMENTUM_SQUEEZE, 11-20 STRUCTURE (Order Block / PDL-PDH Sweep / Break of Structure), 21-30 GEOMETRIC (Wyckoff Spring / Fibonacci OTE), 31-50 HONE_BEST which mutates the top-scoring candidate from phases 1-3. Gemini 3.1 (primary) with 2.5 Flash, Flash-Lite, and Claude Opus 4.7 fallback chain ingests sentinel broadcasts, the weekly digest, the shadow matrix of missed trades, persistent cross-cycle memory, and pure-numpy reference code for every structural primitive. Each paradigm ships a working numpy template the LLM mutates rather than reinventing from scratch. Every generation is backtested against 7 days of real data across 88 symbols with realistic fees, walk-forward validated across 3 rolling 2-day sub-windows. A peer-reviewer statically vets each candidate before evaluation runs; the post-mortem agent writes lessons to persistent memory; a meta-learner rewrites directive rules every 5 generations based on what is and isn't working. Scored by V9 Calmar + WR-50 Floor + TC-264 Ceiling + Drawdown Penalty + Per-Symbol Quality + Walk-Forward Floor. Only a candidate that strictly beats the live system on the same 7-day window takes the throne. If nothing beats it, the system holds. The decision not to evolve is itself intelligence.

// LLM Intelligence

50 generation evolution. 4 phases. Natural selection.

Every cycle runs 50 generations through four phases: 10 MOMENTUM_SQUEEZE, 10 STRUCTURE, 10 GEOMETRIC, then 20 HONE_BEST that mutates the top-scoring candidate from the first three phases. Powered by Gemini 3.1 with a 2.5 Flash, Flash-Lite, and Claude Opus 4.7 fallback chain. It ingests sentinel broadcasts, the strategic digest, the shadow matrix of missed trades, the full indicator physics from every recent win and loss, and persistent cross-cycle memory - past failures, paradigm performance stats, and meta-rules derived from what has gone wrong. Armed with 97 pre-computed indicator columns and a pure-numpy market-structure library (swing detection, BOS/CHoCH, order blocks, FVG consequent encroachment, volume profile, squeeze confluence). Candidates mutate a working paradigm template rather than reinventing from scratch. Successful candidates that compute synthetic indicators on-the-fly generate recommendations to add them to Streamline permanently. Then it evolves.

PHASE 1 · MOMENTUM / SQUEEZE

F SQUEEZE FIRE CONFLUENCE

Squeeze release as trigger, quality-filtered on 6+ bar coil duration and compression ratio below 0.5, then confluence-gated by ADX > 22 and DI spread direction. Only high-quality coils fire - dud squeezes are suppressed.

PHASE 2 · STRUCTURE

B ORDER BLOCK RECLAIM

Pure numpy detection: last opposing candle before a 1%+ displacement swing. Enter on retest of the OB zone in the correct premium/discount half of the range, with volume expansion > 1.2× the 20-bar average.

PHASE 2 · STRUCTURE

C PDL/PDH SWEEP RECLAIM

Previous Day Low or High sweep-and-reclaim using pre-computed calendar pivots. Low-ADX ranging filter (ADX_14 < 25) prevents firing in trends where sweeps rarely reclaim. Ranging regime specialist.

PHASE 2 · STRUCTURE

E BREAK OF STRUCTURE

Close above the last 50-bar swing high (or below swing low) as the BOS trigger, confluence-gated by ADX > 22 and volume expansion > 1.3×. Classic trend-continuation entry on confirmed structural break.

PHASE 3 · GEOMETRIC

A WYCKOFF SPRING

Price wicks below a 200-bar range low then closes back above it - the classic Wyckoff accumulation spring. Confluence-gated by RSI_14 < 35 (oversold). Longs only; shorts suppressed in ranging regime.

PHASE 3 · GEOMETRIC

D FIBONACCI OTE

Optimal Trade Entry zone: 61.8-78.6 percent retracement of the dominant 100-bar swing. Confluence-gated by ADX > 20 and DI spread > 5 in the entry direction. Institutional re-accumulation window.

PHASE 4 · HONE BEST (gens 31-50)

★ MUTATE THE TOP SCORER

After the first 30 generations across MOMENTUM_SQUEEZE, STRUCTURE, and GEOMETRIC phases, the top-scoring candidate becomes the HONE parent. Gens 31-50 take the parent's full source and tighten one threshold, swap one filter, or add one confluence check per generation. Architecture preserved, edges sharpened. This is where the swarm stops exploring and starts converging.

THE EVALUATOR

7-day walk-forward evaluation across 88 perpetual futures symbols with realistic taker fees (0.06%) and slippage (0.02% per side). Every generation is scored on overall performance AND on each of 3 rolling 2-day sub-windows - a strategy that is profitable overall but breaks on any sub-window gets halved. Honest Physics enforced: T-120 lookups with a 120-second delay so lookahead bias is impossible. Hard 1.2 ATR stop floor. 4-hour maximum hold time. Target 50-264 trades per cycle; above the ceiling a multiplicative over-trading penalty applies. Absolute PnL floor - negative net equals score zero regardless of win rate. Drawdown penalty - above 8% max drawdown deducts $2,000, above 15% deducts $5,000. Per-symbol quality - if fewer than 30% of symbols are profitable, score is halved. A peer-reviewer statically vets every candidate before eval time is spent. Operator-set safety rails. Non-negotiable.

STRICT BEAT OR HOLD

A new candidate must strictly beat the reigning champion on the same fresh 7-day window or the throne stays. Champion and runner-up both saved to the Hall of Fame, so every surviving strategy remains available for analysis. If nothing beats the live config, the system holds. No change for the sake of change. The swarm evolves only when evolution produces something genuinely superior - and proves it holds up across every walk-forward sub-window.

When each agent fires.

50 generation evolution. 4 phases. Natural selection.

Direct feed from every agent.