Skip to content

Architecture

The Tribunal is structured as a fractal hierarchy — each layer is a tribunal of specialized agents feeding the layer above it.


LayerComponentsFunction
Layer 1Knights (Signal Specialists)Fast, independent signal generation
Layer 2Prophet + KingIntelligence briefing + capital allocation
Layer 3King of KingsRegime detection + meta-strategy

The Knights are specialized signal agents, each trained on different market data:

  • Momentum Knight: RSI acceleration, OI confirmation, volatility filters
  • Flow Knight: Liquidation clusters, taker pressure, funding extremes
  • Trend Knight: SMA regime detection, crossover signals

All three Knights receive identical market context. Each votes independently without coordination. This produces genuine signal diversity — not prompt-engineered disagreement.

[Knight 1] → BUY (high confidence)
[Knight 2] → HOLD (neutral)
[Knight 3] → SELL (medium confidence)

The system records each vote, then passes to the King for consensus.


The Prophet (Silent Observation in Phase 2)

Section titled “The Prophet (Silent Observation in Phase 2)”

The Prophet does not vote or trade in Phase 2. It watches decisions and records observations:

  • Regime Detection: Bull / Bear / Sideways market classification
  • Knight Ledger: Historical win rates by regime
  • Macro Calendar: Scheduled events (Fed, CPI)
  • Pattern Matching: Historical analogs from past sessions

Its observations become training material for later King fine-tuning. Active advice only belongs in a later phase after the observations have a validated track record.

The King deliberates with full context:

  • Prophet’s brief
  • All Knight votes and reasoning
  • Portfolio state (drawdown, exposure, Kelly sizing)

The King makes the final call:

  • 3/3 Agree: Full position
  • 2/3 Agree: Half position
  • Disputed: Quarter position (weight to highest-EV agent)
  • All HOLD: No trade

The King of Kings is not a model — it is the institution itself:

  • Tracks regime transitions
  • Selects which King version is active
  • Manages adapter swapping based on regime

The system detects market regime and activates the appropriate specialist adapter.


State Data -> Signal Extraction -> Label (outcome)
| |
[Knights] -> [King Consensus] -> [Action / No Action]
| | |
[Prophet] watches | [Portfolio]
| |
[Decision Ledger] [SQLite DB]
| |
[Fine-tune -> Next Cycle]

Each cycle the model running signal extraction IS the previous cycle’s fine-tuned model:

  • Better model → Better signal quality
  • Better signal → Better training data
  • Better training data → Better next model

This recursive improvement is the core compounding mechanism.


Three is the minimum for:

  • Majority: Any decision needs 2 of 3
  • Dissent: Any voice can question the other two
  • Balance: No single agent dominates

Three specialists on one GPU beat one larger model every time. The architecture and hardware solved each other.