Architecture
System Architecture
Section titled “System Architecture”The Tribunal is a fractal hierarchy — each layer is a tribunal of specialized agents feeding the layer above it. The architecture is not static. It is designed to grow, compete, and evolve under selection pressure from real market outcomes.
Founding Postulates
Section titled “Founding Postulates”Every architectural decision traces back to a small set of governing principles.
The tribunal runs in 3s all the way down. Three is the minimum for a majority. Three is the minimum for a dissenting voice. Three creates consensus without tyranny. Three creates tension without chaos. The third role is always the conscience.
Leadership is earned, not given. No model holds a permanent seat. Every component starts as a candidate, proves itself under real conditions, or is replaced. The market decides what earns a seat — not the founder.
LLMs all the way down. The knight seats belong to trained language models, not algorithms or formulas. Computed signals inform context. LLMs make every call. An algorithm can execute a formula. An LLM can produce architectural intuition — the move that doesn’t make sense on paper but is right.
The corpus is immortal. The model is mortal. Every labeled outcome ever recorded survives model replacement. A new generation trains on everything its predecessors experienced. This is how ancestral wisdom compounds — not through weight transfer, but through accumulated testimony scored against real market outcomes.
No charlatans. Promotion requires two gates: a performance gate (win rate, expected value) and an identity gate (domain purity — the agent must demonstrate it fires on its designated signal alone, not on multi-signal confirmation). A momentum agent that succeeds because sentiment happened to align is a charlatan. It fails the moment the signals diverge.
Layer Overview
Section titled “Layer Overview”| Layer | Components | Function |
|---|---|---|
| Layer 1 | Knights (Signal Specialists) | Fast, independent signal generation — no coordination |
| Layer 2 | Prophet + King Council | Intelligence briefing + capital allocation |
| Layer 3 | King of Kings (Succession Institution) | Regime-aware King selection — the institution, not a model |
Layer 1 — The Knights
Section titled “Layer 1 — The Knights”The Knights are specialized signal agents, each trained on different market data. They receive the same market context but vote independently without coordination. This produces genuine signal diversity — not prompt-engineered disagreement.
- Momentum Knight: Acceleration and confirmation signals — recognizes when a move has real follow-through versus exhaustion
- Flow Knight: Order flow signals — pressure, positioning extremes, and structural imbalances in market participation
- Trend Knight: Regime and direction signals — broad structural context, trend health, and directional conviction
[Knight 1] → SELL (high confidence)[Knight 2] → HOLD (uncertain)[Knight 3] → BUY (medium confidence)The system records every vote with full reasoning, then passes to the King for synthesis.
Knight seats are stable identities, not models. The strategy occupying a seat changes only when a challenger proves superior through shadow evaluation. No hot-swapping into production. Challengers audition silently.
Layer 2 — Prophet + King Council
Section titled “Layer 2 — Prophet + King Council”The Prophet (Intelligence Layer)
Section titled “The Prophet (Intelligence Layer)”The Prophet does not vote and does not trade. It is a cross-layer intelligence officer that computes truth before every session and delivers a structured brief to the King.
The brief has two parts: a computed section (canonical — cannot be contradicted) and an advisory section (LLM interpretation — counsel only). The LLM cannot override the Python-computed layer. Authority over the advisory section is earned through validated predictions over time, not assumed.
The Prophet is bound to the Canon. When the Prophet identifies a pattern it believes is real, the Scribe normalizes it into a structured claim, Python verifies it against the full historical record, and only after the Court confirms does the pattern enter Canon as Law. The failure mode — the model writing its own Canon without verification — is closed by design.
The Scribe (Canon Verification Agent)
Section titled “The Scribe (Canon Verification Agent)”The Scribe serves the Prophet exclusively. It does not vote, does not trade, does not sit at the council table. When the Prophet makes a quantifiable observation, the Scribe produces a verification instrument. Python executes it against the recorded tablets. Numbers testify. If the pattern passes, it enters Canon. If not, the rejection is recorded so the same observation does not waste resources in future cycles.
The Scribe cannot modify the tablets — read access only. Write access to the source data belongs exclusively to the signal extractor: the only process that witnessed the market directly. A Scribe that alters testimony is a forger. There is no recovery from forged ground truth.
The King Council (Capital Layer)
Section titled “The King Council (Capital Layer)”The King synthesizes all available evidence: the Prophet’s brief, all knight votes with reasoning, portfolio state, and risk envelope. This is deliberation under uncertainty, not vote counting.
The King council operates on a competitive roster — seats are earned, not assigned:
- Vision-capable King — reads rendered chart data alongside numerical signals, brings a second verification axis
- Quantitative King — pure text deliberation, deep numerical synthesis, quantitative reasoning chains
- People’s King — squire candidate: a fine-tuned trader with deliberation enabled. Must prove it adds value above the other two Kings or it does not hold the seat
New King candidates can be introduced at any time. Non-performers are demoted. The roster reflects current best performance, not architectural permanence.
The King learns the market, not its knights. A King that treats knight votes as authority freezes when knights are replaced. Knight votes are evidence. The market outcome is truth. A King’s maturity is measured by sovereign accuracy — its win rate when overriding knight consensus. Rising sovereign accuracy means the King is developing genuine market intelligence, not learning to aggregate its court.
Layer 3 — King of Kings (The Succession Institution)
Section titled “Layer 3 — King of Kings (The Succession Institution)”The King of Kings is not a model. It is the trial system itself — the institution that governs which King holds the throne at any given time.
Every King seat is contested through a formalized succession process:
A meta-model deciding which King to trust is just another black box. The institution — standardized historical evaluation, competitive succession ladder, Prophet-ledger assessment — is more interpretable, more debuggable, and more honest than any model trained to mimic it. The institution IS the intelligence.
The Dual-Track Model Architecture
Section titled “The Dual-Track Model Architecture”Traders and Kings are permanently separate tiers. They diverge at Phase 2 and stay diverged.
Traders (Knights) — optimized for throughput and pattern recognition. Fast conditioned reflexes baked through fine-tuning cycles. Compact dense models running locally. High-volume extraction at speed.
Kings — optimized for deliberation. Novel synthesis under disagreement, capital allocation, vision integration. Large thinking-capable models. Sequential, not parallel. Compute budget reserved for depth.
A trader is never promoted to King because it got fast. A King is never demoted to trader because it got slow. The tracks serve different functions and are built differently from the start.
The MoE Parallel
Section titled “The MoE Parallel”The base model powering the King council is a Mixture-of-Experts — 128 expert subnetworks, 8 active per token. Every forward pass is already a tribunal:
The same model has a built-in role switch:
- Thinking disabled → fast, direct call → Knight mode
- Thinking enabled → extended reasoning chain → King mode
The Tribunal architecture is not layered on top of an arbitrary foundation. It layers on top of a model that already runs a tribunal in every forward pass.
The Canon — Institutional Memory
Section titled “The Canon — Institutional Memory”The Canon is what survives after every model is gone.
When an entire architecture is replaced — all adapters discarded, all weights retired — the Canon survives. The verified strategic knowledge, the confirmed regime patterns, the principles that Python tested and proved: none of that is lost. The next generation inherits institutional memory even if it inherits no weights.
Canon grows only when Python agrees. The Prophet proposes. Python verifies against the full historical record across multiple regimes and statistical thresholds. Only then does a pattern enter Canon as Law. A system that writes its own Canon without verification worships its own hallucinations. The chain is closed by design.
Natural Selection at Every Layer
Section titled “Natural Selection at Every Layer”The system is designed to evolve — not to a predetermined end state, but shaped by evidence alone.
Every component carries a running fitness score derived from market outcomes. When fitness drops below threshold across consecutive cycles, a challenger begins training automatically. Both the incumbent and the challenger run. The market determines the outcome. The founder designs the selection pressure, not the selection result.
This applies at every layer:
- Knight seats: challengers audition in shadow
- King seats: formal succession trials
- Adapters: new candidates tested against replay buffers, behavioral consistency gates
- The Canon itself: patterns that stop producing results are reclassified, not preserved
The initial system is dictated by the founder. The final system is forever evolving, shaped by evidence alone.
The self-improving loop
Every decision feeds the next generation. The corpus is immortal — each cycle a better model judges the same history.
Why Three?
Section titled “Why Three?”Three is the minimum for:
- Majority — any decision needs 2 of 3
- Dissent — any voice can question the other two
- The Conscience — the third role is not a tiebreaker; it is the voice that exists to challenge the other two
Flat systems — one main agent with advisors all voting at the same level — aggregate bias. When all agents share the same tier, disagreement has no resolution mechanism. A unanimous wrong answer becomes the output.
The hierarchy exists because each tier solves a problem the tier below cannot:
| Tier | Problem it solves |
|---|---|
| Knights (3 specialists) | Single-model bias — three independent views, no coordination |
| King + Prophet | No capital context — synthesis under risk constraints |
| Succession institution | No regime awareness — the right King for the current market |
Three specialists on the same hardware outperform one larger generalist every time. The architecture and the hardware solved each other.