Published Research
The evidence base
for our platform.
Every capability in our platform is grounded in peer-reviewed research and validated against historical market data. We publish our findings openly because we believe the field moves faster when the work is visible.
Performance data
Simulation accuracy over time
Coranor vs. traditional VaR (12-month rolling window)
Coranor Agent Simulation
Traditional VaR
Policy attestation pass rate
Monthly containment enforcement success over a 12-month period
Attestation pass rate (%)
Floor: 99.94% · Current: 100%
Historical crisis benchmark accuracy
Directional prediction accuracy across 10 major crisis events
% Accuracy
Mean: 93.8% · Range: 88% to 100%
Platform throughput scaling
Agent-steps per second as compute nodes scale
×10⁶ steps/sec
Near-linear scaling up to 96 nodes
Platform prototype
A preview of the Coranor Labs interface, currently in closed beta with Lighthouse partners.
Workspace
Risk Dashboard
LIVE ENGINEPortfolio: Global Macro Fund · As of Sept 28, 2026
VaR (99%)
-$4.2M
+0.3% vs benchmark
Simulated Drawdown
-12.8%
Severe shock profile
Agents Active
2.4M
Zero drift detected
Portfolio P&L Under Simulation (30-Day Window)
10⁶ agent iterationsWorkspace
New Scenario: Stagflation Shock
Rate shock
+350bps
Applied over 90 days
Inflation deviation
+4.2%
vs. 2% target
Liquidity stress index
78 / 100
Severe
Agent population
1,000,000
Heterogeneous
Published papers
The April 2025 Tariff Shock: A Multi-Agent Simulation
Pre-registered evaluation of multi-agent LLM forecasting under strict anti-lookahead protocols
We evaluated whether adversarial multi-agent debate improves corporate forecasting across 30 major U.S. corporations subjected to the April 2025 tariff shock. Across 26,000+ Monte Carlo simulations, multi-agent debate systematically degraded forecast accuracy and escalated overconfidence, performing worse than historical base rates and worse than an uninformative coin flip.
Key Finding: Multi-agent debate (A3, Brier 0.386) underperformed both a single-pass prompt (A1, 0.257) and naive base rates (A0, 0.235).
Ablation Ladder Results
| Level | Architecture | Brier Score ↓ | ECE Calibration ↓ |
|---|---|---|---|
| A0 | Naive Base Rates (No LLM) | 0.235 (Best) | 0.063 |
| A1 | Single-Pass LLM (Zero Debate) | 0.257 | 0.280 |
| A2 | FirmAgent + Monte Carlo | 0.362 | 0.406 |
| A3 | Full Multi-Agent Debate | 0.386 (Worst) | 0.441 |
*Brier score: mean squared error (0.00 = perfect, 0.25 = 50/50 baseline, lower is better).
Note: Pre-registered protocol with 180 dynamic trap questions quarantined against knowledge contamination.