Augur

Multi-Agent Ensemble Forecasting — 17 Specialists, One Answer

The Problem

Probabilistic forecasting is one of the most valuable and least commoditised capabilities in professional services. Hedge funds pay analysts hundreds of thousands of dollars per year to produce calibrated probability estimates for macro events, policy changes, and market outcomes. Strategy consultants build scenario models. Political risk firms charge per-report for geopolitical assessments.

The problem is not that AI can't forecast — it's that a single LLM produces a single perspective, with no framework for calibration, no domain specialisation, and no transparency into how it arrived at a number. Superforecasting research (Tetlock, IARPA) shows that the best forecasters are those who aggregate independent estimates from people with diverse analytical frameworks. No AI product has done this seriously.

Cindicator raised $15M in 2017 to crowdsource human predictions and blend them with ML. The crowd was too noisy and incentive-misaligned. The model never got good enough to justify the operational cost.

The Solution

Augur fans any probabilistic question out to a mesh of 17 domain AI specialists simultaneously. Each specialist applies a distinct analytical playbook — loaded from a TOML manifest at runtime — and returns a structured forecast:

{
  "probability": 0.72,
  "confidence": 0.85,
  "reasoning": "Fed futures are pricing ~65% probability of a cut...",
  "key_assumptions": ["Inflation stays below 3.5%", "No major financial shock"],
  "key_uncertainties": ["Q2 employment data not yet released"],
  "would_change_if": "A surprise CPI print above 3.8% would shift my estimate to ~0.35"
}

Responses aggregate via confidence-weighted average. A specialist with 0.9 confidence counts 9x more than one with 0.1. The result: a calibrated ensemble probability with transparent per-specialist reasoning.

The Specialist Mesh

Each specialist's playbook is defined in a TOML manifest. Update the manifest, update the specialist's behavior — no code changes.

API

POST /v1/forecast
{
  "question": "Will the Fed cut rates before September 2026?",
  "context": "Current fed funds rate: 4.25-4.5%...",
  "specialists": ["reasoner", "market_analyst", "intelligence_analyst"],
  "synthesize": true
}

→ {
    "ensemble_probability": 0.68,
    "ensemble_confidence": 0.74,
    "consensus": "lean_yes",
    "synthesis": "Specialists broadly agreed on a likely cut, with market_analyst most bullish (0.75) citing futures pricing, and intelligence_analyst most cautious (0.61) citing election-year political uncertainty. Key residual risk: Q2 employment print.",
    "specialists": [...],
    "successful": 3,
    "failed": 0,
    "latency_ms": 4200
  }

Performance

• >Parallel execution: All specialists run simultaneously via `asyncio.gather` — ensemble latency ≈ slowest single specialist, not sum
• >Graceful degradation: Timeout or API error per specialist → `status: "timeout"`, probability falls back to neutral 0.5, excluded from weighted average
• >TOML hot-reload: Manifest changes take effect on next request, no restart needed

Revenue Model

The addressable market for institutional forecasting tools is well-established: Bloomberg Terminal (~$6B/yr), Refinitiv (~$6B/yr), specialized political risk firms (Eurasia Group, Oxford Analytica). Augur is not competing with those databases — it's offering calibrated probabilistic synthesis that none of them provide.

Stack

• >Backend: FastAPI + Python, Anthropic Claude API (claude-sonnet-4-6)
• >Execution: asyncio parallel fan-out, per-specialist timeout, synthesis via claude-haiku
• >Config: TOML manifests (17 files, hot-loaded at request time)
• >Deploy: Production at eudaimonia.win

**Rising Sun** · [risingsun.name](https://risingsun.name) · April 2026

*"17 specialists. One calibrated answer."*