AUGUR

Twelve specialised agents analyse prediction markets from deliberately different perspectives — probability estimation, news impact, momentum, correlation discovery, contrarian bias-hunting, cross-bookmaker odds comparison, sports form, sentiment, anomaly scanning. The agents never communicate directly. Each submits beliefs, news impacts, correlations and timing signals to a shared world model, which fuses them into one picture through Bayesian consensus — each belief weighted by confidence, by the agent’s earned trust, and by type-specific information decay.

The world model produces second-order intelligence no single agent could: when one market shifts, correlated markets update automatically through causal links, and Shannon entropy across agent beliefs marks the markets where disagreement — and therefore potential alpha — is highest.

Fixed timers are wasteful, so the system is event-driven: price spikes, breaking news and surprise belief shifts wake the relevant agents within seconds rather than minutes. An attention engine runs a multi-armed bandit over the swarm, directing agent cycles to the markets that need them — high entropy, recent surprise, approaching expiry — and speeding up agents that produce signal while slowing those that produce noise.

Information ages honestly. Every agent’s output carries an exponential half-life — a price snapshot is stale in minutes, breaking news in half an hour, structural analysis holds for hours — and the rates adapt to the market regime. Beliefs fade rather than expire. Above it all, a metalearning system tracks which agent combinations actually work, per market category, regime and time of day, and feeds that record back into the scoring — the swarm gets measurably smarter over time, not just statically clever.

Most trading systems try to find edges. One agent in AUGUR exists to destroy them. For every high-conviction trade the swarm wants to make, the adversarial agent runs a structured attack: a pre-mortem that imagines the trade has already lost and names the weakest link; a blind-spot attack that asks what the market knows that the agents do not; checks for positions that share a common failure mode; and a flag for stable-market logic applied in volatile conditions.

When all agents agree strongly against the market price, a sentinel refuses to treat that as high conviction — unanimous agreement is flagged as a possible blind spot, the very scenario that produces black-swan losses, and the consensus is sent back to be attacked. Plausible failure scenarios enter the consensus as counter-beliefs. The trade gate is hard: a dangerous rating cuts the composite score by 60%; an abort rating blocks execution entirely. The swarm must prove its consensus survives a structured attack before capital is committed.

Every trade generates a complete natural-language thesis: which agents contributed and what they believed, the specific evidence, the news context, the correlated positions, what the pre-mortem found, the kill conditions, and the full scoring breakdown. Every decision is auditable end to end.

Execution is risk-managed throughout — conviction is a seven-factor composite rather than raw edge size, positions are sized by the Kelly criterion behind drawdown protection and correlation guards, and the system tracks its own calibration curve: if its 80% confidences resolve correctly less often than 80% of the time, future confidence is scaled down automatically.

The swarm architecture, world model and intelligence layer are built and running. The system trades on paper by default; live execution, a real-time dashboard and a backtesting framework are roadmap. It is published as open-source software under the MIT licence.