Technical Overview

The Tribunal Consensus Engine

The Tribunal is a self-improving orchestration framework for autonomous decision-making through hierarchical multi-agent consensus. It is formally called Ancestral Tribunals — each generation trains on verified outcomes, but only evidence that survives the truth layer earns authority. The weights carry accumulated wisdom. The institution compounds across generations.

read-only inference stream

Sanitized Consensus Trace

Phase 2

t+00s[INGEST]market-state packet sealed; feature names redacted

t+07s[KNIGHT-IMB]directional vote recorded; no capital authority

t+14s[KNIGHT-MOM]dissent preserved; no cross-agent coordination

t+21s[SQUIRE]shadow evidence scored; execution path untouched

t+28s[PROPHET]computed brief attached; advisory authority pending evidence

t+35s[KING]sovereign gate checked; capital intent withheld

Why Hierarchy — The Flat System Problem

Most multi-agent systems place agents at the same tier: a main agent, a set of advisors, all voting together. The problem is structural. When every voice occupies the same level, disagreement has no resolution mechanism. Unanimous wrong answers become the output. Bias aggregates instead of canceling.

The Tribunal’s hierarchy exists because each tier solves a problem the tier below it cannot:

Tier	The problem it solves
Independent signal seats	A single model has no check on its own bias. Specialist seats, zero coordination, produce measurable signal diversity.
Sentinel + shadow squires	The court needs two kinds of untrusted evidence: a regime watchman that flags dangerous deployment to the King, and new structural sources that audition without capital authority until durability is proven.
Prophet + King	Knights have no capital context. The King synthesizes under risk constraints. The Prophet brings verified intelligence the trading tier cannot see.
Succession institution	A single King has no accountability. The institution governs which King holds the throne based on demonstrated performance across all market regimes.

Each layer is a correction of the layer below’s structural blindspot — not a redundancy.

Architecture Philosophy

Three core principles underpin everything:

Multi-Agent Hierarchy — specialized agents with decoupled roles, each tier solving what the tier below cannot
Hardware-Accelerated Inference — optimized for enterprise-grade GPUs, dense models for volume, thinking models for deliberation
Self-Improving Lineage — every verified decision becomes training data for the next generation; the system compounds across cycles

1. The Multi-Agent Hierarchy

Layer 1 — The Knights

Independent signal specialists. Each is trained around a distinct evidence view — different market data, different signal focus. They receive market context and vote without coordinating.

The independence is structural, not prompt-engineered. Models trained on different data naturally diverge in their assessments. When the court agrees, the signal is strong. When it splits, the system expresses genuine uncertainty.

Knights are directional only. They do not abstain, size positions, or express portfolio intent. That separation is intentional: signal production and capital authority are different jobs.

Knight seats are stable identities. The role is permanent; the specific model occupying it earns its position through shadow evaluation. Challengers never hot-swap into production.

Layer 1.5 — The Sentinel + Shadow Squires

The Sentinel is the regime watchman, canonically replacing the Risk Warden working name. It is knight-lineage but not directional — a knight that zoomed out: it watches volatility, chop, crash pressure, and capital danger so the field knights are not asked to manage risks outside their job description. It issues regime mandates (OPEN / WATCH / LOCKDOWN) rather than trades and reports directly to the King as a second senior voice alongside the intelligence layer — it informs, it does not command, and it earns standing only by proving it improves outcomes.

Shadow squires are structural evidence candidates. They can monitor alternative data, produce independent diagnostics, and challenge the court — but they cannot touch execution until they survive paired validation, cost assumptions, and regime durability checks.

This is how the system absorbs new information without giving away its edge or poisoning the court with fragile patterns: audition first, promote only after evidence, reject explicitly when the edge does not hold.

Layer 2 — Prophet + King Council

The Prophet is a cross-layer intelligence officer. It computes regime state, portfolio health, and verified context before every session and delivers a structured brief to the King. Its computed output is canonical — the LLM cannot contradict Python-verified numbers. Its advisory interpretation is counsel. Authority over the advisory layer is earned through validated predictions, not assumed.

The King council synthesizes all evidence: the Prophet’s brief, all knight votes with full reasoning chains, portfolio state, and risk parameters. This is deliberation under uncertainty. The King owns HOLD/abstention and capital intent; the executor owns final dollar conversion. The King’s maturity is measured not by overall win rate but by sovereign accuracy — its win rate specifically when it overrides knight consensus. A King that never overrides its court is a vote counter. A King with rising sovereign accuracy is developing genuine market intelligence.

The council roster is competitive. Kings earn their seats. Non-performers are demoted. New candidates can be introduced at any time.

Layer 3 — King of Kings (The Succession Institution)

The King of Kings is not a model. It is the trial system — the institution that governs which King holds the throne.

A meta-model selecting the “right King” is another black box adding another failure point. The institution — standardized historical evaluation across all regimes, a competitive succession ladder, Prophet-ledger-based assessment — is more interpretable and more accountable than any model trained to mimic it. The institution is the intelligence.

2. The MoE Parallel

The base model powering the King tier is a Mixture-of-Experts architecture — a large pool of specialist subnetworks with only a small subset active on each forward pass. Every inference call is already a tribunal: the router evaluates all specialists, selects the most relevant voices, silences the rest, synthesizes.

The same model carries a built-in role switch:

Thinking disabled  →  fast, direct output   →  Knight mode
Thinking enabled   →  extended reasoning    →  King mode

The Tribunal does not sit on top of an arbitrary foundation. It layers on top of a model that already runs a tribunal at the weight level.

3. Inference Architecture

The system runs two permanently separated model tracks:

The Trading Track (Knights) — compact dense models, optimized for throughput and pattern recognition. Run locally on workstation hardware. High-volume corpus generation at speed. Fast conditioned reflexes, no deliberation overhead.

The Deliberation Track (Kings) — large thinking-capable models, optimized for depth. Sequential, not parallel. Compute reserved for synthesis: novel combinations under disagreement, capital allocation, vision integration. A trader is never promoted to King because it got fast. The tracks serve different purposes and are built differently from the start.

4. The Self-Improving Lineage

The Cycle

Market data → Directional extraction → Outcome labels (market is the judge)
                    ↓
              Dataset preparation → Fine-tune → Quantized inference model
                    ↓
              Next cycle runs with improved weights → richer signals → better data

Each cycle the model running signal extraction can inherit from the previous cycle, but inheritance is not trusted blindly. The current operating discipline is evidence-first: if raw signals plateau, the system auditions new structural evidence rather than stacking more hierarchy above a weak base.

The self-improving loop

Every verified outcome feeds the next generation. Raw emissions, training targets, and promoted evidence remain separate.

The Canon

The Canon is what survives after every model generation is retired.

When an architecture is replaced entirely — all adapters discarded, all weights gone — the Canon survives. Verified strategic patterns, confirmed regime behaviors, principles tested by Python across years of market history: none of this is lost. The next generation inherits institutional memory even if it inherits no weights.

Canon grows only when Python agrees. The Prophet proposes a pattern. Python verifies it — statistical threshold, multi-regime confirmation. Only then does it enter Canon as Law. A system that writes its own Canon without verification worships its own conclusions. The chain is closed by design:

Prophet observes → Scribe normalizes → Python verifies → Canon records

Adapter Succession

At the weight level, the same “earn your seat” principle governs adapter replacement. A new candidate adapter trains alongside a replay buffer — a stratified record of every regime the system has faced. The candidate must prove it improved on the current adapter without abandoning what the incumbent had mastered. A candidate that learns the new at the cost of behavioral regression fails the identity gate. No regressions are promoted to production.

5. Natural Selection

The system is designed to evolve — not to a predetermined end state, but shaped by evidence.

Every component carries a fitness score derived from market outcomes. When fitness falls below threshold across consecutive cycles, a challenger trains automatically. Both run. The market decides the outcome. The founder designs the selection pressure; evidence shapes the result.

This applies at every layer:

Knight seats — challengers audition in shadow, earn comparison data
King seats — formal succession trials across full regime history
Adapters — tested against replay buffers and behavioral consistency gates
The Canon itself — patterns that stop holding are reclassified, not preserved out of inertia

The initial system is dictated by the founder. The final system is forever evolving, shaped by evidence alone.

6. Infrastructure

Watchdog systems — automated monitors detect stalls and trigger recovery; pods self-terminate on completion
No lookahead — backtesting uses only data that existed at decision time; the integrity of labeled data depends on this completely
Consensus-driven execution — actions only execute when the deliberation chain completes and the sovereign gate permits exposure
Hardware-aware inference — dense models for volume extraction (local GPU), thinking models for deliberation (cloud accelerator when needed)
Moat-safe disclosure — architecture is visible, but thresholds, feature recipes, prompts, model weights, and labeling logic remain private