DealsFlow: An Evidence-Gated Agentic Harness for Venture Capital Deal Triage (arXiv (Cornell University))

Summary

• DealsFlow is an agentic triage system for venture capital pitch decks built on the principle that the LLM reads and scores categories while deterministic code aggregates, caps, and only ever lowers scores.p.1
• The evaluation on a portfolio of 243 SEIS/EIS early-stage investments (£272.50m invested) focuses on 142 mature investments at least three years old to avoid time-censoring bias.p.5
• On 142 mature holdings, the full pipeline achieves a rank correlation with total-value multiples of Spearman rho = 0.451, an AUC0 of 0.709 for write-offs, and an AUC2 of 0.791 for deals returning over 2x.p.8
• Reviewers found that 84.2% of 240 cited claims were fully supported by their retrieved sources, although 70.0% of generated reports still required at least one correction.p.10

Key insights

1. External corroboration rates differ significantly across outcomes (63.6% for investments returning >2x versus 37.8% for write-offs), highlighting potential hindsight leakage when using undated retrieval.p.7
2. An asymmetric sceptic agent evaluates category scores and evidence, with code permitting adjustments only downwards by at most 20 points per category.p.4
3. The sceptic pass affected 95.1% of the 142 mature deals with a mean score reduction of 2.99 points, while the ceiling affected 48.6% with a mean reduction of 1.23 points.p.9
4. Verification and correction tasks accounted for 50.1% of assisted analyst time (9.45 out of 18.85 minutes) during deal triage.p.8