A public showcase

BTC, decided.

An experiment: Claude team agents + Bayesian inference, deciding when to trade BTC.

The question

Who's about to be forced to move and how do we trade it?

01

How a run goes.

READ

What's true in the market right now?

Five populations, each watching a different slice. Each agent executes code against the data and forms an independent view.

ARGUE

What does it mean as a whole?

Four agents over three rounds, each round answering a specific question. Then a debate, mechanically stopped at message twelve. No exceptions.

R1Is liquidation pressure quiet, elevated, or about to cascade?
R2Which population is constrained: fresh, trapped, crowded, or capitulating?
R3If a forced move happens, which direction does it push?
DECIDE

What does the Trader do?

One of six explicit decisions, with full reasoning chain.

OPENenter a new position
CLOSEexit the current one
CANCELkill a pending order
REPLACEswap a pending order
HOLDkeep what's there
STAND ASIDEflat, no opinion
02

The thesis.

The core problem is coordination under uncertainty: how do you combine five heterogeneous, asynchronous estimators — each seeing different data, operating on different timescales — without drowning in coordination cost or collapsing into groupthink?

We built a Bayesian inference system for BTC trading that addresses this directly. It observes 5 market populations — leveraged traders, options dealers, institutional flow, spot real-money, and basis arbitrageurs. Each population has hidden constraint states that determine what it must do next. The core edge is detecting forced moves before they happen: a crowded leveraged long MUST sell when margin calls hit, a short-gamma dealer MUST buy when price rises through their strike.

The question isn't what the market believes — it's who is constrained and what they're forced to do.

Five independent analysts rotate through different market domains each round, executing code against six specialized MCP servers — funding, options, on-chain flow, cross-venue arbitrage — to form independent views. The tools aren't menus; each agent writes its own queries. No analyst rates the same population twice — structural independence, not just instructed independence. A Bayesian engine tracks calibrated beliefs across rounds. Between rounds, analysts share brief structured estimates about each other's populations — what they think a peer's evidence implies, without revealing raw data. This preserves independence while allowing cross-domain inference to flow. The coordination cost is zero when analysts are confident, and targeted when they're not.

After three rounds, the analysts see the final posteriors — alongside measured anchors: the market's memory signature, recent liquidity sweep patterns, and the natural half-life of each population's influence — and collectively debate what the joint picture means. The trading thesis emerges from this contested debate: which population is about to be forced to move, in which direction, and how the other populations amplify or dampen that move. No single analyst holds the complete picture. The configuration — "leveraged rally without spot backing" or "gamma-amplified cascade into crowded longs" — arises from interaction between domain perspectives, not from rules or pattern matching. An independent canary population guards against groupthink: if it shifts without new evidence from its own domain, conformity is contaminating the signal.

The Trader then judges the debate, probes weak spots, and decides.

Code handles the math. Intelligence handles the meaning.
03

Recent decisions.

Built on
  • A Claude team of agents — 4 analysts + 1 Trader, rotating roles
  • Bayesian inference — 24-cell joint posterior, KL-tracked
  • Code-executing MCP servers — 6 servers, 100+ market-data tools