Skip to content

Architecture

The Tribunal is a separation-of-powers architecture for autonomous trading research. Its current doctrine is The Inversion: the trading layer is made of deterministic Python engines, while the LLM tribunal governs only when it proves measurable contribution over code.

This is not a retreat from the court metaphor. It is the court applying its own law. A seat that cannot prove contribution is retired or moved to shadow. The architecture obeys evidence before mythology.


Every architectural decision traces back to a small set of governing principles.

Trading and governance are separate authorities. Engines produce auditable target outputs. The King owns allocation and sizing intent. The Prophet supplies verified context. The Sentinel watches regime and capital danger. The executor owns final conversion into orders under risk constraints.

Leadership is earned, not given. No model or subsystem holds a permanent seat. A component starts as a candidate, proves itself under paired measurement, or remains outside the authority path.

Python truth precedes model authority. Computed facts cannot be contradicted by an LLM narrative. LLM counsel may be useful, but only measured contribution promotes it from commentary to authority.

The record is immortal. The model is mortal. Every verified result survives model replacement. The system compounds through accumulated testimony scored against market outcomes, not through loyalty to any one model generation.

No charlatans. A component must prove both performance and identity. It cannot win by secretly doing a different job than the seat it claims to occupy.


LayerComponentsFunction
Layer 1Reversal, Carry, Trend EnginesDeterministic trading seats; auditable target generation
Layer 1.5Candidate engines + shadow instrumentsNew systems run without authority until measured
Layer 2King, Prophet, SentinelGoverning court: allocation, context, regime conscience
Layer 3Succession InstitutionFormal process for replacing governors when challengers prove superior

The original design began with LLM signal seats voting on trades. That era served its purpose: it stress-tested the belief that specialized model agents could discover reliable directional signal by training cycle alone. The evidence did not justify keeping those seats on the critical path.

The current architecture therefore moves the power of three up one layer:

  • Three engines below — deterministic trading systems with auditable outputs.
  • Three governors above — King, Prophet, Sentinel, each probationary.
  • One rule between them — the model keeps authority only when it improves the measured result over a frozen Python stand-in.

The retired signal seats remain part of the origin story. They are the ordeal that clarified where authority belongs. The court did not fail; it judged its own premise and changed the seating chart.


The trading layer is code-first. Engines do not receive authority because they sound plausible. They are built as deterministic systems, evaluated against historical outcomes, measured under realistic costs, and frozen before they become candidates for governance.

The three intended seats are:

  • Reversal Engine — active daily target generation; market-neutral relative-value process, described publicly without proprietary recipe details.
  • Carry Engine — funding and basis research path; next deterministic engine to build and validate.
  • Trend Engine — longer-horizon continuation path; gated before construction so the system does not spend cycles on an unlearnable seat.

A trading engine is not an LLM. It is a truth-layer instrument. It produces a target, a record, and a result that the court can be judged against.


New systems do not enter the authority path directly. They run in shadow first: candidate engines, regime instruments, and diagnostic sources observe the same market history or live feed without changing deployment.

Promotion requires more than plausible behavior. A candidate must survive paired evaluation, realistic cost assumptions, and calendar durability. Rejected candidates remain recorded so the system does not rediscover the same false promise later.

This is the public commitment: audition before authority, measurement before mythology.


The King is the capital governor. It synthesizes engine outputs, portfolio state, risk envelope, and verified context. Its job is not to invent trades from noise. Its job is to decide how much authority each trading engine should receive right now.

The King is on probation by design. It earns standing only if its allocation, sizing, or stand-down decisions improve the result against a deterministic control on the same rows. If it does not, Python keeps the job.

The Prophet is a cross-layer intelligence officer. It computes and presents context before a decision: regime state, portfolio health, candidate observations, and claims that require verification.

The brief has two parts: a computed section and an advisory section. The computed section is canonical. The advisory section is counsel. It becomes more than counsel only after validated predictions over time.

The Sentinel is the conscience of the system. It watches whether the environment is safe for capital at all and reports directly to the King.

It does not originate trades. It issues posture mandates such as normal operation, caution, or stand-down pressure. It informs rather than commands, except for hard executor-level safety brakes reserved for catastrophic conditions.


The Scribe serves the Prophet by turning quantifiable observations into verification instruments. Python tests the claim against the recorded tablets. Numbers testify. If the claim passes, it can enter Canon. If not, the rejection is preserved.

The Canon is what survives after every model generation is retired. Verified strategic knowledge, confirmed regime behavior, operating law, and rejected hypotheses remain available to the next generation even if the weights are discarded.

A system that lets an LLM write its own Canon without verification worships its own hallucinations. The Tribunal closes that loop by design.


The King of Kings is not a model. It is the trial system that governs which governor holds authority.

CANDIDATEpoolTRIALmulti-regimehistoricalevaluationKING INWAITINGchallenger confirmedDUELledger decidesall gatesmust passwinsTHRONEincumbent reignsDEPOSEDreturns tocandidate poolThe incumbent does not vote on its own replacement — the ledger is the evidence.

A meta-model choosing which model to trust would be another black box. The institution — standardized replay, challenger trials, verified ledgers, and formal demotion — is more interpretable and more accountable.


The base model used for deliberation may itself be a Mixture-of-Experts: many subnetworks, a router, a small active set per token. Every forward pass already resembles a small tribunal.

MoE FORWARD PASSTRIBUNAL SESSIONall experts evaluatedactive subset selectedWEIGHTEDSYNTHESISmajoritystays silentSAMELOGICKINGSOVEREIGN GATEKPSall counsel heardweighted by evidenceCAPITAL CALL

The system uses that structure where it fits: deliberation, synthesis, context, and governance. It does not require the model to own every layer.


The system evolves toward measured usefulness, not a predetermined cast.

This applies across the stack:

  • Trading engines: candidates run in shadow before authority.
  • Governors: King, Prophet, and Sentinel keep standing only through measured contribution.
  • Adapters: candidates are tested against replay buffers and behavior gates.
  • Canon: 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.


Verified learning loop

Engines, governors, and candidates are separated until measured results justify authority.


A single model has no native mechanism for dissent, abstention, or accountability. A flat multi-agent panel often averages bias instead of resolving it. The court structure creates separation of powers:

  • Engines produce auditable target outputs.
  • Shadow candidates test new claims without authority.
  • The King governs allocation under risk.
  • The Prophet separates verified context from interpretation.
  • The Sentinel watches whether capital should be deployed at all.
  • The Canon records what survived and what failed.

The architecture is valuable because it can replace its own assumptions without losing its institutional memory.