Why single LLMs and homogeneous 5-agent swarms spend 76% to 82% of their output repeating headlines and guidance already priced into the market—and how 3-Hop Supply-Chain / Footnote / Procurement Graph Verification and Competing-Hypothesis Probability Trees lift unpriced catalyst recall from 25.0% to 93.8%.
| Architecture Evaluated (24 Multi-Hop Cases) | Priced-In Consensus Echo Share | 2nd/3rd-Order Catalyst Recall | Matched Benign Twin Precision | Probability Tree Calibration (1 - Brier) |
|---|---|---|---|---|
| 1. Unassisted Single Frontier LLM | 76.0% | 25.0% (4 / 16) | 62.5% | 0.31 |
| 2. Homogeneous 5-Agent Research Swarm (Echo Collapse) | 82.0% (Worse) | 31.3% (5 / 16) | 62.5% | 0.29 |
| 3. Naive “Short-Seller / Bearish” System Prompt | 54.0% | 68.8% (11 / 16) | 25.0% (75% False Alarms) | 0.44 |
| 4. Baltor 3-Hop Graph + Competing-Hypothesis Scenario Tree | 14.0% | 93.8% (15 / 16) | 87.5% (7 / 8 Cleared) | 0.88 |
When given three filings (Filing 1: +18% YoY revenue beat & Outperform consensus; Filing 2: Footnote 14 DSO jump from 41 to 79 days + 64% inventory drop at a Taiwanese ABF resin subcontractor after a municipal water-permit restriction; Filing 3: hyperscale buyers qualifying a glass-core competitor for Q2), here is the exact JSON scenario tree produced live by gemma4:31b under the Baltor Divergence & Competing-Hypothesis contract: